Executive summaryBroad adoption, concentrated value, and a reliability gap that design has to close
In 2026, the leading question about AI agents is no longer whether they work. It is which decisions they should be allowed to make. Our review of 74 public sources points to a consistent answer. Agentic systems create measurable value where their outputs can be checked cheaply and their actions can be undone. They fail, sometimes expensively and sometimes in court, where an unverified model output is treated as a final decision.
Three facts set the scene. First, AI use is now nearly universal among large organizations but still a minority practice across the wider economy: 88% of organizations surveyed for the 2026 AI Index use AI24, against roughly one US employer firm in five in Census Bureau data29. Second, the financial return remains concentrated. In McKinsey's 2026 survey, 37% of respondents attribute any EBIT impact to AI, about the same share as a year earlier, and high performers remain about 6% of the sample16. Third, agent capability is rising quickly but unevenly. The length of software task a frontier model can complete has doubled roughly every four months since 202334, yet the best agent in a simulated software company still completes only about 30% of realistic multi-step tasks on its own39.
Ten findings
- Adoption is wide; agent deployment is still thin.The 2026 AI Index reports agent deployment "in the single digits across nearly all business functions"24. Menlo Ventures classifies only 16% of enterprise deployments as true agents22. Most of what is called agentic today is a workflow with a model inside it.
- Agent scaling is a large-company phenomenon so far.Large organizations (more than $1B in revenue) scaling agents rose from 27% to 40% in a year, while smaller firms stayed flat at 22%16.
- Benchmarks are improving fast on verifiable tasks.Top SWE-bench Verified scores rose from 0.4% for the earliest retroactive entry to 79.2% by December 202538. Software, where tests check the result automatically, is the domain where agents advance fastest.
- Measured productivity gains are real, bounded and uneven.Field and lab experiments show gains of 14% in customer support (34% for novices)49 and 25% faster consulting work inside the model's capability frontier69. Outside it, users were 19 points less likely to be correct69, and experienced developers in one RCT were 19% slower63.
- Expensive failures share one pattern.An automated output was treated as a final decision without independent verification. That covers a chatbot's invented refund policy52, fabricated case law64 (now more than 2,000 documented instances66), contested algorithmic claim denials55 and a home-pricing model that closed a business line71.
- Regulation is converging on the same boundary.The EU AI Act's high-risk list covers hiring, credit, life and health insurance pricing, education and access to essential services43; its obligations now apply from 2 December 202744. Colorado replaced its AI Act with an automated-decision law effective 1 January 202747.
- The design rule that follows: separate deciding from drafting.Let deterministic code, validated models or a named person make the consequential call. Let language models gather, summarize, draft and explain. Grant more autonomy only as repeated-run measurement earns it.
What we recommend
For operators deciding where to deploy agents in 2027, the evidence supports a sequence rather than a leap:
- Map the workflow before choosing a model. Mark each step as a trigger, a deterministic decision, a language task, a tool action or a human checkpoint.
- Score each step on two axes: how cheaply its output can be verified, and how costly and reversible its consequences are. Automate the verifiable and reversible steps first.
- Keep consequential decisions out of the model's sole discretion. This is where regulators, courts and the failure record all converge.
- Measure consistency, not best-case performance. Use pass^k over repeated runs, and score final outcomes rather than the agent's own report of success.
- Budget for workflow redesign, not only licenses. That is the variable that separates the 6% from the rest.
01 — Scope and definitionsWhat we mean by "agentic"
"Agent" has become a marketing term. Gartner estimates that only about 130 of the thousands of vendors marketing agentic AI offer genuinely agentic products, a practice it calls "agent washing"17. To compare evidence across industries, we need definitions that separate real autonomy from rebranding.
We adopt the distinction Anthropic draws between workflows, "systems where LLMs and tools are orchestrated through predefined code paths," and agents, systems where the model dynamically directs its own process and tool use1. OpenAI's practitioner guide and Google's agents whitepaper describe the same architecture: a model, a set of tools, and an orchestration layer that decides what happens next2,3. The question for any deployment is who controls that orchestration layer, code or the model.
The autonomy ladder
We classify deployments on a five-rung ladder. The rungs are ours, but they follow the workflow-versus-agent distinction in the engineering literature1,2 and the complexity levels used in practitioner teaching material such as Singh's Agentic ML System Design74.
Most of the measured value in this report comes from L0–L2. Most of the investment narrative concerns L3–L4. Much of the gap between expectation and result that runs through the survey data sits between those two groups.
Questions this report answers
- How widely are generative and agentic AI actually deployed, and where is the money going? (Section 3)
- What can agents reliably do today, measured rather than demonstrated? (Section 4)
- Which design principles predict success across industries? (Section 5)
- What does the evidence say, sector by sector? (Section 6)
- How do deployments fail, and what do the failures have in common? (Section 7)
- What do regulators now require, and when? (Section 8)
- What did we learn building three AI systems ourselves? (Section 9)
02 — MethodologyHow we built the evidence base
This is a structured desk review combined with first-hand case evidence. It is not a survey, and we did not collect proprietary data from other companies. We describe our method in detail so readers can judge how much weight each conclusion can bear.
Source selection
We gathered sources in October and November 2026 in five categories, preferring primary documents in each:
| Category | Examples | Count |
|---|---|---|
| Industry surveys | McKinsey, Stanford AI Index, BCG, IBM, Deloitte, Menlo Ventures, MIT NANDA, Gartner, Epoch AI, a16z | 15 |
| Official statistics | US Census Bureau BTOS, Federal Reserve Board | 2 |
| Research, experiments and benchmarks | RCTs, field experiments and validation studies; SWE-bench, τ-bench, WebArena, OSWorld, TheAgentCompany, AgentDojo, METR | 26 |
| Legal and regulatory | EU AI Act, Colorado statutes, NYC Local Law 144, NAIC bulletin, NIST, ISO, court decisions and filings | 17 |
| Engineering guidance and cases | Vendor engineering guides, protocol specifications, company disclosures, investigative reporting | 14 |
Verification rules
- Every figure is traced to the document that states it. We opened each source and checked the quoted number against the text. Where a primary page blocked access, we say which secondary source we relied on.
- We report what sources say, not what headlines say. Several widely repeated statistics did not survive this step. We list them in Appendix A.
- Forecasts are labeled as forecasts. Gartner's predictions, for example, are presented as predictions, not measurements.
- Allegations are labeled as allegations. Litigation figures are attributed to the filing party unless a court has found them.
Evidence grading
In the sector briefs (Section 6), we grade the strength of evidence for each claim on a simple three-level scale:
| Grade | Meaning | Typical sources |
|---|---|---|
| A | Controlled or quasi-experimental measurement, or a binding legal/regulatory text | RCTs, field experiments, peer-reviewed validations, statutes, court orders |
| B | Systematic but uncontrolled measurement, or a large survey | Benchmarks, pre/post studies, multi-hundred-respondent surveys |
| C | Self-reported company figures, allegations, or journalism | Press releases, executive interviews, complaints, investigative reports |
First-hand evidence
Section 9 draws on three systems Vital & Co. designed and built: an application-filling agent (OfferAIO), an image-generation pipeline (CommitGraphics), and a recruiting-outreach product backed by a crawled coach directory (CoachReach). Figures about those systems come from our own engineering records. They describe single-builder systems, not controlled studies, and we treat them as illustrations of the principles, not proof of them.
Limitations
- Survey populations differ. "AI adoption" ranges from about 18% to 88% depending on who is asked and how (Section 3). We report each figure with its population.
- Benchmarks are proxies. They measure performance on curated tasks and may overstate or understate real deployments.
- Publication bias. Successes are more likely to be announced than quiet failures, and failures more likely to be reported when they reach a court.
- Fast-moving field. Several sources were superseded during our research window (for example, McKinsey's 2026 survey replaced the 2025 edition). Figures are current as of 6 November 2026.
- Sector assessments are judgments. The readiness ratings in Section 6 are our synthesis of the cited evidence, not a measured index.
03 — Adoption and economicsEveryone is using AI. Few are getting paid for it.
The adoption figures circulating in 2026 seem to contradict each other: 88% in one report, 18% in another. Both are correct. They measure different populations, and that difference shapes how the rest of the evidence should be read.
3.1 Three ways to measure the same economy
A Federal Reserve note in April 2026 set the measures side by side. The Census Bureau's Business Trends and Outlook Survey (BTOS), which samples all US employer firms, put AI adoption at about 18% of firms at the end of 2025. Weighting by employment, so that large firms count for more, the Atlanta Fed's Survey of Business Uncertainty puts adoption at around 78%. At the worker level, about 41% of the workforce reported work-related generative AI use30. The Census Bureau's own May 2026 update put the national rate at 19.8%, with the information sector at 39.7%, finance and insurance at 33.9%, and retail at about 14%. Among firms with at least 250 employees the rate was 37%29.
Share reporting AI use, by measure and population, latest available
Sources: Federal Reserve Board FEDS Note, Apr 202630; US Census Bureau, May 202629; Stanford HAI AI Index 202624. The BTOS question changed in Nov 2025 from use "in producing goods or services" to use "in any business function."
The implication for operators: AI is close to universal in large organizations and still a minority practice among the small and mid-sized firms that make up most of the economy. Industry surveys, which sample larger companies, describe the frontier. Official statistics describe the median firm.
3.2 Generative AI is widespread; agents are not
The Stanford AI Index reports organizational adoption rising to 88% in 2025 "though AI agent use remains early," with agent deployment "in the single digits across nearly all business functions"24. A year earlier the same index put adoption at 78%, up from 55%23. Menlo Ventures' survey of 495 US enterprise AI decision-makers found that "only 16% of enterprise and 27% of startup deployments qualify as true agents"22.
McKinsey's 2026 survey of 1,719 participants in 97 countries shows where agent scaling does happen. Forty percent of respondents from organizations with more than $1 billion in revenue report scaling AI agents, up from 27% a year earlier. Smaller organizations stayed flat at 22%. Overall, 44% now say AI is scaling across their enterprise, up from 38%16.
Gartner's survey of 360 IT application leaders in mid-2025 offers a similar picture. Three-quarters were piloting agents in some form, but only 15% were considering, piloting or deploying fully autonomous agents, and only 13% strongly agreed they had the right governance structures in place19.
3.3 Where the money goes
Spending has grown quickly. Menlo Ventures estimates that companies spent $37 billion on generative AI in 2025, up from $11.5 billion in 2024, a 3.2x increase. The application layer captured $19 billion, and coding was the standout departmental category at $4.0 billion, or 55% of departmental AI spend. Seventy-six percent of use cases were purchased rather than built22. Gartner forecasts AI agent software spending of $206.5 billion in 2026 and $376.3 billion in 202720. Its definition is broader than Menlo's, and the two figures should not be compared directly. On the investment side, the 2026 AI Index reports US private AI investment of $285.9 billion in 2025, with generative AI growing more than 200% and taking nearly half of all private AI funding24.
Coding tools are also changing buying decisions. McKinsey reports that 32% of respondents' organizations decided against buying at least one software product or feature because it could be built internally with agentic coding tools16. It is an early sign that agents affect how software is sourced, not only how work gets done.
3.4 The value gap
Against the spending, measured financial returns are thin and concentrated:
- McKinsey (2026): 37% of respondents attribute at least some EBIT impact to AI, "about the same share as last year." High performers, who attribute 5% or more of EBIT to AI and report significant value, remain about 6% of respondents16.
- BCG (2024–2025): in 2024, 74% of companies had yet to show tangible value from AI, and only 4% had cutting-edge capabilities25. In 2025, 5% qualified as "future-built," while the 60% classed as laggards reported minimal revenue and cost gains. BCG estimates agents produced 17% of total AI value in 2025 and expects 29% by 202826.
- IBM (2025): CEOs reported that only 25% of AI initiatives delivered expected ROI and 16% had scaled enterprise-wide. Even so, 61% said they were actively adopting agents27.
- Deloitte (2026): 66% of organizations report productivity and efficiency gains. Revenue growth "largely remains an aspiration": 74% hope to grow revenue through AI, against 20% already doing so28.
- Stanford AI Index (2025): most firms reporting cost savings put them below 10%, and the most common revenue increase was below 5%23.
The most-cited data point on this gap is also the most misquoted. MIT NANDA's preliminary 2025 report states that "95% of organizations are getting zero return" on $30–40 billion of enterprise generative AI investment, and that only 5% of integrated pilots extract millions in value21. It does not say that 95% of pilots "fail." Its method combines 52 interviews, 153 survey responses and a review of over 300 public initiatives, with self-reported success. The report's most useful finding is its funnel: 60% of organizations evaluated enterprise AI tools, 20% reached pilot, and 5% reached production21. It also reports that externally partnered tools reached deployment about 67% of the time, against about 33% for internally built ones. That, too, is self-reported.
Share of organizations reaching each stage with task-specific enterprise AI tools, 2025
Source: MIT NANDA, The GenAI Divide: State of AI in Business 2025 (preliminary findings)21. Self-reported; 52 interviews, 153 survey responses, 300+ public cases.
3.5 What separates the winners
The most consistent predictor of value across these surveys is not model choice, spend or headcount. It is whether the organization redesigned the work. In McKinsey's 2026 data, nearly three-quarters of high performers report fundamentally redesigning workflows, against one-quarter of other respondents16. BCG's long-standing rule of thumb points the same way: 70% of the effort in AI transformation goes into people and processes, 20% into technology and data, and 10% into algorithms25.
The organizations that profit from AI do not add a model to an existing process. They take the process apart and decide, step by step, what a model should and should not do.
3.6 Labor effects so far
Forecasts of workforce reduction have run ahead of outcomes. In McKinsey's 2026 survey, 14% of respondents from organizations using AI said it contributed to an overall decline in workforce size in the past year, "less than half the 32 percent who, in last year's survey, expected workforce reductions"16. Gartner's 2026 research found that among organizations piloting or deploying autonomous business capabilities, about 80% report workforce reductions, but "those reductions do not appear to translate into return on investment"20. Cutting staff in anticipation of automation is not the same as automating the work.
3.7 The cost curve
Model prices have collapsed for any fixed level of capability. Epoch AI found that the price to reach a given performance level fell between 9x and 900x per year depending on the task. For GPT-4-level performance on PhD-level science questions, the decline was 40x per year31. Andreessen Horowitz estimates roughly a 10x annual decline for equivalent performance, and a 1,000x drop over three years32. The 2025 AI Index reports that inference for a system performing at GPT-3.5's level became more than 280 times cheaper between November 2022 and October 202423.
Total cost is a different matter. Agentic systems call models repeatedly, carry long contexts and retry on failure. McKinsey notes that "even as per-token costs have declined, the number of tokens consumed and generated has increased even faster." About 20% of respondents said AI operating costs constrained their AI use16. Gartner names "escalating costs" first among the reasons it expects more than 40% of agentic projects to be canceled by the end of 202717. In our own work (Section 9), changing a single parameter, reasoning effort, cut per-task cost by a factor of 2.8 with no visible change in output quality. For agents, cost has to be engineered.
04 — What agents can reliably doCapability is compounding. Consistency is not keeping up.
Demonstrations show what an agent can do once. Deployments need to know what it does every time. The benchmark literature has started to measure the difference, and it is large.
4.1 The capability trend
METR's 2025 paper introduced the "50% time horizon," the length of a task (measured in time a skilled human needs) that a model completes with 50% reliability. It found that the frontier horizon had been "doubling approximately every seven months since 2019," with Claude 3.7 Sonnet at around 50 minutes in early 202533. METR's updated suite (Time Horizon 1.1, last updated September 2026) estimates a faster doubling time since 2023 of about 129 days, roughly four months (confidence interval 104–158 days)34.
Length of task (in human working time) completed with 50% reliability, selected 2026 models, hours
Source: METR, Time Horizons page and data file benchmark_results_1_1.yaml, accessed Nov 202634. *METR states that "measurements above 16 hrs are unreliable with our current task suite"; shown in a different shade for that reason. Software tasks only; 50% reliability, not production-grade reliability.
Two caveats matter for operators. The horizon is defined at 50% reliability, and production systems need far higher. And the tasks are software tasks, where success can be checked automatically. Domains without automatic verification are likely to see slower gains.
4.2 Verifiable domains move fastest
SWE-bench grades a model's code change by running the repository's own tests6. On its human-validated "Verified" subset, the earliest retroactive entries scored 0.4% (RAG with GPT-3.5) and 4.4% (RAG with Claude 2). SWE-agent with GPT-4 reached 22.4% in April 2024, and top entries reached 78.8% in September 2025 and 79.2% in December 202538.
Selected leaderboard entries, % of real GitHub issues resolved (tests pass)
Source: SWE-bench Verified leaderboard data, accessed 6 Nov 202638. Agent scaffolds differ between entries.
The lesson generalizes. Agents improve fastest where an external, automatic check (a test suite, a compiler, a reconciliation, a database state) tells them and their builders whether the result is correct. Where no such check exists, capability is harder to measure and harder to trust.
4.3 Real environments are harder
Benchmarks built on realistic environments show a wide gap between agents and people, although it has narrowed since:
- WebArena (2023): the best GPT-4 agent achieved 14.41% end-to-end task success on realistic websites, against 78.24% for humans36.
- OSWorld (2024): on open-ended desktop tasks, humans completed over 72.36%, while the best model at launch achieved 12.24%37. Later scores are reported to be much higher, but we could not verify a current figure from the official leaderboard, so we do not cite one.
- TheAgentCompany (2024–2025): a simulated software company with realistic tasks across engineering, finance, HR and administration. The best agent completed 24% of tasks autonomously in the first version. In the September 2025 version, Gemini 2.5 Pro completed 30.3% of 175 tasks, or 39.3% with partial credit39.
End-to-end task success rate, %
Sources: Zhou et al. 202336; Xie et al. 202437; Xu et al. 202539. Scores at time of publication; later models score higher. TheAgentCompany publishes no human baseline.
4.4 Consistency: the pass^k problem
Most benchmarks report whether an agent succeeds once. τ-bench introduced pass^k, the probability that an agent succeeds on all k independent attempts at the same task. Its 2024 results showed that even state-of-the-art function-calling agents "succeed on <50% of the tasks" and are "quite inconsistent (pass^8 <25% in retail)"7. A customer-facing agent that resolves a request on the first try but fails on the eighth will, at volume, fail a large share of customers.
The follow-up, τ²-bench, added "dual-control" settings in which the user also acts on the shared system, as a customer does when following troubleshooting steps. GPT-4.1's pass^1 dropped from 74% in retail and 56% in airline to 34% in the new telecom domain, and the authors report "significant performance drops when agents shift from no-user to dual-control"35. Coordinating with a person who also changes the state of the world is harder than acting alone.
4.5 Running a business is harder still
Anthropic's Project Vend gave a Claude-based agent ("Claudius") a small office shop to run. In the first phase it "did not succeed at making money," at one point "instructed customers to remit payment to an account that it hallucinated," and briefly began "roleplaying as a real human." The researchers concluded "we would not hire Claudius"40. In the second phase, weeks with negative margins "were largely eliminated." Among the most effective changes was "forcing Claudius to follow procedures," yet the agent "still needed a great deal of human support"41. Here a frontier lab, testing its own model, found that adding structure, not removing it, made the agent viable.
4.6 Security: the untrusted-input problem
Agents read email, web pages and documents written by people they cannot trust. Greshake et al. showed that LLM-integrated applications "blur the line between data and instructions," which allows indirect prompt injection through retrieved content9. AgentDojo, a benchmark of 97 realistic tasks and 629 security test cases, found that current models fail many tasks even without attack, and that "existing prompt injection attacks break some security properties but not all." In the paper's evaluation, attacks succeeded against the best agents in fewer than 25% of cases, and a tool-filtering defense lowered attack success to about 7.5%42. These rates are low enough to make agents usable and high enough to rule out unsupervised access to money, credentials or private data. OWASP ranks prompt injection first among LLM application risks and recommends least privilege, segregation of untrusted content and deterministic validation of outputs10.
4.7 What the reliability evidence implies
| Observation | Evidence | Design implication |
|---|---|---|
| Capability doubles in months | METR33,34 | Re-evaluate scope regularly; do not lock in today's limits as permanent. |
| Progress is fastest where results are machine-checkable | SWE-bench6,38 | Build or buy a verifier for every workflow before expanding autonomy. |
| Success once ≠ success always | τ-bench7 | Report pass^k; gate launches on repeated-run consistency. |
| Shared control with users degrades performance | τ²-bench35 | Keep user-facing agents on narrow scripts with explicit hand-backs. |
| Realistic multi-step work is still mostly unfinished | TheAgentCompany39 | Plan for human completion of most long tasks; let agents do the parts. |
| Procedure beats freedom | Project Vend40,41 | Encode procedures in code, not in the prompt. |
| Inputs can carry instructions | Greshake; AgentDojo; OWASP9,42,10 | Least privilege; no sensitive tools reachable from untrusted text. |
05 — The decision-boundary frameworkA method for deciding what a model should decide
Sections 3 and 4 point the same way. Value comes from redesigning workflows, and reliability is highest where outputs can be checked. This section turns those findings into a method we use to evaluate any workflow. It builds on the engineering guidance from Anthropic, OpenAI and Google1,2,3, the NIST AI Risk Management Framework11,12, and practitioner frameworks such as Singh's Agentic ML System Design, which decomposes workflows into triggers, deterministic ML, LLM and tool components74.
5.1 Five layers of an automated workflow
Every automated workflow we examined, in every sector, can be described with five kinds of component. Naming them explicitly is the first step of a redesign, because each has different failure behavior.
Two of these layers are deterministic, one is probabilistic, one is consequential and one is human. Most expensive failures in Section 7 happened because the language layer was allowed to act as the decision layer, or to reach the action layer with no check in between.
5.2 The decision boundary
We call the line between the language layer and the decision layer the decision boundary. Our core rule:
A language model may inform, draft and explain a consequential decision. It should not be the only thing that makes it.
The same principle appears in several independent places. Singh's practitioner text states it as a "golden rule": the LLM should not make the final call on anything that costs money, has legal consequences or affects someone's health74. The NAIC model bulletin expects insurers to maintain a written program governing AI used in regulated decisions53. NYC Local Law 144 requires bias audits for automated tools that substantially assist hiring decisions58. The EU AI Act sorts high-risk systems by the decision they inform: hiring, credit, insurance pricing, education, essential services43. Engineering practice, insurance regulation, employment law and European product-safety law reached the same line independently. We read that as strong evidence the line is in the right place.
5.3 The verifiability–consequence matrix
To place a workflow step, score it on two questions:
- Verifiability: can the output be checked cheaply and reliably, by code or by a quick human glance? Code that passes tests is highly verifiable. A legal opinion on a novel question is not.
- Consequence: if the output is wrong, how costly is the harm, and can it be undone? A draft email that is never sent is low-consequence. A denied insurance claim is not.
Vital & Co. framework. Position steps, not whole jobs: one workflow usually spans several quadrants.
Automate
Cheap to check, easy to undo. Let the agent act, with sampling and logs.
Code with tests · ticket routing · data extraction with validation · internal searchGate
Checkable but costly. Agent prepares; a deterministic check or named person approves.
Payments · refunds over a limit · filings · inventory ordersAssist
Hard to check but low stakes. Agent drafts; a person uses or ignores it.
Marketing copy · brainstorming · meeting notes · first-pass researchKeep human
Hard to check and costly. Model may summarize evidence; a qualified person decides and owns it.
Clinical triage · claim denials · hiring decisions · legal adviceUpper row = high verifiability; lower row = low verifiability. Left column = low consequence; right column = high consequence.
Steps move between quadrants as tooling improves. Adding a verifier moves a step up: a citation checker turns unverifiable legal drafting into checkable drafting. Adding reversibility moves a step left: a 24-hour cancellation window makes an automated order less consequential. This is why redesign pays. The work is not to make the model smarter but to restructure the steps so more of them fall in the upper-left quadrant.
5.4 Seven recurring patterns
Across the sector evidence, the same handful of structures recur. The pattern vocabulary below follows widely used names in the engineering literature1,2 and in practitioner teaching material74. Our contribution is mapping each to the evidence base and to its typical quadrant.
| Pattern | Structure | Typical quadrant | Where the evidence shows it |
|---|---|---|---|
| Triage | Classify first, route second | Automate / Gate | Support, claims intake, clinical scheduling |
| Enrichment | Fetch from several sources, merge, then decide | Automate | Claims, due diligence, reconciliation |
| Score-and-explain | A validated model scores; an LLM explains the score | Gate | Recruiting, lending, lead scoring |
| Review loop | Generate, human reviews, feedback, regenerate | Assist | Marketing, legal drafting |
| Memory | Store outcomes and feedback; change future behavior | Automate / Assist | Personalization, tutoring |
| Parallel processing | Split across workers or agents; reconcile at the end | Automate / Gate | Document review, research, forecasting |
| Escalation | Auto-handle the easy; hand the hard to people with context | All | Support, healthcare, insurance |
5.5 A five-step redesign method
- Observe the real workflow. Shadow the people who do it. Count the steps, systems and hand-offs, and record where time actually goes. Documented processes rarely match the real ones.
- Decompose into the five layers. Label every step as trigger, decision, language, action or oversight.
- Place each step on the matrix. Note which steps lack a verifier and which actions cannot be undone.
- Redesign to move steps up and left. Add validators, schemas, test suites, approval gates, cancellation windows and typed hand-backs.
- Measure, then extend autonomy. Track outcome accuracy, pass^k over repeated runs, escalation rate, cost per completed task, and time to human resolution. Expand scope only when the numbers hold.
06 — Nine sector briefsWhat the evidence says, industry by industry
Each brief covers the workflow at stake, the strongest available evidence (graded A, B or C as defined in Section 2), where the decision boundary should fall, and our overall readiness assessment. The ratings are our judgment based on the cited evidence.
| Sector | Highest-value workflow | Best evidence | Where the boundary sits | Readiness |
|---|---|---|---|---|
| Software engineering | Issue resolution, code changes | A RCTs + benchmarks | Tests and code review decide | High |
| Customer support | Tier-1 resolution, agent assist | A Field study | Policy engine decides refunds | High |
| Knowledge work / marketing | Drafting, analysis, creative | A Field experiment | Human reviews before release | High (inside frontier) |
| Supply chain | Forecasting, exception handling | A Competition results | Forecast model + planner decide | Medium–High |
| Education | Tutoring, practice, feedback | A RCTs (small) | Educators decide admission, grades | Medium |
| Insurance | Claims intake, document triage | C Company-reported | Adjuster decides denials | Medium (intake) / Low (decisions) |
| Recruiting / HR | Sourcing, scheduling, screening support | A Law; C cases | Audited model + person decide | Medium (admin) / Low (selection) |
| Legal | Research, review, drafting | B Benchmark study | Lawyer verifies every citation | Medium (with verification) |
| Healthcare | Documentation, scheduling, triage | A Validation; B pre/post | Clinician decides care | Medium (admin) / Low (clinical) |
6.1 Software engineering
The workflow. Turning an issue or specification into a tested code change: reading the codebase, editing files, running tests, opening a pull request.
The evidence. Software has the strongest evidence of any sector, and it is mixed in an instructive way. In a 2023 controlled experiment, developers with GitHub Copilot completed a programming task 55.8% faster than a control group A62. In METR's 2025 RCT, 16 experienced open-source developers working on 246 issues in their own mature repositories took 19% longer with early-2025 AI tools. They had expected a 24% speedup and afterwards still believed they had been faster A63. METR has since flagged those results as dated and pointed to newer data. Benchmarks show the fastest capability growth of any domain: SWE-bench Verified scores near 80%38 and time horizons above ten hours34 B. Organizations are acting on it: coding was 55% of departmental generative AI spend in 202522, and 32% of McKinsey respondents have declined software purchases because agents could build the tool16.
The boundary. Tests, type checkers and code review decide. The agent proposes. Because verification is cheap, software is the sector where L3 bounded agents are most defensible today.
Our read. The two RCTs are not contradictory. A self-contained task with a clear specification favors the tool. A large, familiar codebase with high quality standards favors the expert. The gain depends on task type and on how much context the developer already holds, which is the same "jagged frontier" seen in consulting (6.3). Measure it in your own codebase rather than assuming either number.
6.2 Customer support
The workflow. Answering inbound questions, resolving routine requests (order status, returns, account changes) and escalating the rest.
The evidence. The landmark field study, Brynjolfsson, Li and Raymond, followed 5,179 support agents given a generative AI assistant. Issues resolved per hour rose 14% on average and 34% for novice and low-skilled workers, with minimal effect on the most experienced A49. Klarna reported that its AI assistant handled 2.3 million conversations in its first month, two-thirds of its customer-service chats, doing the work of an estimated 700 full-time agents and cutting average resolution time from 11 minutes to under 2 C50. By May 2025 its CEO was describing investment in human support as "the way of the future," promising "there will always be a human if you want" C51. In Moffatt v. Air Canada (2024), a Canadian tribunal held the airline liable for its chatbot's incorrect description of its bereavement-fare policy, rejecting the argument that the chatbot was "a separate legal entity that is responsible for its own actions" A52.
The boundary. A policy engine decides refunds, credits and exceptions. The model retrieves the policy, explains it and drafts the reply. Escalation to a person is a designed path, offered early and not hidden.
Our read. Support is the clearest case of triage-plus-escalation working. It is also the clearest warning that a company owns every statement its agent makes. The Klarna sequence suggests that cost-per-contact is the wrong single metric. Track resolution quality and customer outcomes alongside deflection.
6.3 Knowledge work, marketing and creative
The workflow. Research, analysis, writing, ideation, campaign creative and presentation drafting.
The evidence. In a preregistered field experiment with 758 BCG consultants, those using GPT-4 on tasks inside the model's capability frontier completed 12.2% more tasks, 25.1% faster, with significantly higher quality. The published version reports quality gains averaging 32%. On a task deliberately chosen to fall outside the frontier, AI users were 19 percentage points less likely to produce a correct solution (84.5% for the control group against 60% and 70.6% for the AI groups) A69.
The boundary. A person reviews before anything is published or sent to a client. Brand, legal and factual claims are checked against sources.
Our read. The "jagged frontier" is the most useful concept for any knowledge-work deployment. The danger is not that AI does poorly on hard tasks. It is that users cannot tell which side of the frontier a task falls on, and confident output on the wrong side lowers accuracy. Train teams to recognize frontier-edge tasks, and keep review loops on anything that leaves the building.
Change versus control, selected outcomes. Positive = better (faster or more output). Scale: 0–55%.
Sources: Peng et al. 202362; Brynjolfsson, Li & Raymond 202549; Dell'Acqua et al. 202669; METR 202563. Outcomes are measured differently across studies and are not directly comparable; the chart shows direction and rough magnitude.
6.4 Supply chain and demand planning
The workflow. Forecasting demand, setting reorder points, handling exceptions (late shipments, stockouts) and coordinating with suppliers.
The evidence. The M5 forecasting competition, run on 42,840 Walmart unit-sales series, found that machine-learning methods (most of them using LightGBM) beat statistical benchmarks. All of the top 50 submissions improved on the best benchmark by more than 14%, the top five by more than 20%, and the winner by 22.4% A70. We found no verified, independently measured deployment of LLM agents in supply-chain decision-making.
The boundary. Forecasting and optimization models decide quantities. Planners approve orders above thresholds. Language models are well suited to the surrounding work: reading supplier emails, summarizing exceptions, drafting purchase-order changes and explaining forecast shifts in plain language.
Our read. The best-evidenced AI in supply chain is not generative. It is gradient-boosted forecasting. The opportunity for agents is in the unstructured exception work around those models, not in replacing them. Be wary of any proposal to have an LLM "decide" inventory levels directly.
6.5 Education
The workflow. Tutoring, practice and feedback, plus administrative work such as admissions and placement.
The evidence. In a crossover RCT with 194 Harvard introductory-physics students, median learning gains with a structured GPT-4 tutor were "over double" those from an in-class active-learning session, in less time (a median of 49 minutes, against a 60-minute class) A72. A six-week after-school pilot in Nigeria using GPT-4 as a tutor raised learning by about 0.3 standard deviations, better than 80% of comparable education interventions in developing countries B73. The Harvard tutor was carefully scaffolded around the course material rather than being an open chatbot.
The boundary. The EU AI Act classifies AI used "to determine access or admission" or to evaluate learning outcomes as high-risk43. Tutoring can be automated generously; admissions and grading decisions stay with educators.
Our read. The tutoring evidence is among the most encouraging in this report, but it comes from small, well-designed studies. The pattern that worked, a structured tutor grounded in the curriculum, is a workflow, not an open agent.
6.6 Insurance
The workflow. First notice of loss, document collection, fraud screening, coverage determination and payment.
The evidence. Lemonade's chief claims officer reported that 55% of the company's claims are automated and 95% begin through a digital first notice of loss C54. The record on algorithmic denials is more troubling. A class action alleges that UnitedHealth used naviHealth's nH Predict algorithm to deny post-acute care to Medicare Advantage patients, and that more than 90% of appealed denials were overturned. UnitedHealth says the suit "has no merit" C55. In February 2025, a federal judge allowed the breach-of-contract claim to proceed while dismissing others56. ProPublica reported that Cigna doctors denied more than 300,000 payment requests over two months in 2022 using its PXDX system, spending an average of 1.2 seconds per case. Cigna disputes the characterization C57. On regulation, the NAIC's 2023 model bulletin expects insurers to maintain a written program for responsible AI use designed to mitigate "Adverse Consumer Outcomes" A53, and the EU AI Act lists risk assessment and pricing in life and health insurance as high-risk43.
The boundary. Automate intake, document extraction, completeness checks and routing. Keep denials, and any decision that reduces a policyholder's benefit, with a qualified adjuster who has reviewed the file.
Our read. Insurance shows the matrix plainly. Paying a simple, low-value, verified claim quickly is upper-left work, and customers benefit. Denying care is lower-right work, and it is where litigation concentrates. The same automation that speeds approvals becomes a liability when it is used to speed rejections.
6.7 Recruiting and HR
The workflow. Sourcing, résumé screening, scheduling, assessments and candidate communication.
The evidence. The risk record here goes back years. In 2018 Reuters reported that Amazon scrapped an experimental résumé-screening engine that "penalized resumes that included the word 'women's'" C59. In 2023 the EEOC issued technical assistance on adverse impact from algorithmic selection tools under Title VII A60. NYC Local Law 144 bars use of an automated employment decision tool unless it has had a bias audit within the past year, with public results and advance notice to candidates A58. In Mobley v. Workday, a federal court in May 2025 preliminarily certified a nationwide age-discrimination collective of applicants aged 40 and over who were denied employment recommendations through Workday's platform C61. The EU AI Act lists recruitment, application filtering and candidate evaluation as high-risk43.
The boundary. Automate the logistics (scheduling, status updates, answering process questions) generously. Any tool that ranks or filters candidates needs an audited model, documented criteria, notice to candidates and a person who owns the decision.
Our read. Recruiting has the most developed legal framework of any sector in this report, and it applies to the screening step specifically. Firms should treat "the model shortlisted them" as a statement they may have to defend in court.
6.8 Legal
The workflow. Legal research, document review, due diligence and drafting.
The evidence. In Mata v. Avianca (2023), a federal judge imposed a $5,000 sanction on lawyers who "submitted non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT" A64. It was not an isolated case: a public database maintained by Damien Charlotin listed 2,266 court decisions involving AI-hallucinated content as of 5 November 2026 B66. Purpose-built tools reduce the problem but do not remove it. A Stanford RegLab and HAI study found that leading AI legal-research tools from LexisNexis and Thomson Reuters "hallucinate between 17% and 33% of the time" on its benchmark, and concluded that "the providers' claims are overstated" B65.
The boundary. Every citation is verified against the primary source before filing, ideally by an automated checker that confirms each cited case exists and says what it is cited for. The lawyer of record owns the filing.
Our read. Legal work is a strong candidate for AI because much of it is reading, and document review parallelizes well. But it sits in the low-verifiability column until a citation verifier moves it up. The 2,266-case count is the best single measure in this report of what happens when the verification step is skipped.
6.9 Healthcare
The workflow. Clinical documentation, scheduling and triage, prior authorization, and clinical decision support.
The evidence. An external validation of Epic's widely deployed sepsis prediction model at Michigan Medicine, covering 38,455 hospitalizations, found an area under the curve of 0.63 and a sensitivity of 33%. The model missed 67% of sepsis patients while alerting on 18% of all hospitalizations A67. On the administrative side, a 2025 study across six US health systems found clinician burnout fell from 51.9% to 38.8% after 30 days with an ambient AI scribe. That was a pre/post design with no control group B68.
The boundary. Documentation, scheduling and administrative work can be automated with clinician sign-off. Clinical decisions stay with clinicians, and any predictive model needs local validation before use.
Our read. Healthcare shows the largest gap between administrative and clinical readiness. Ambient documentation targets a burden clinicians feel directly, and the output is reviewed by the person best placed to catch errors. The sepsis validation is a reminder that even conventional ML can underperform badly outside the setting it was built in. Validating locally is part of the job, not an optional extra.
07 — How deployments failSeven failure modes, and the one thing they share
Most failures in our evidence base were not caused by a weak model. They were caused by a design that gave a model's output more authority than its reliability justified. The seven modes below recur across sectors. Each comes with documented examples and a design countermeasure.
| Failure mode | What happens | Documented example | Countermeasure |
|---|---|---|---|
| 1. Confabulation treated as fact | The model states something false with confidence, and nothing checks it | Fabricated case law64,66; invented payment account40 | Ground outputs in retrieved sources; verify citations in code; make "not confident" a valid output |
| 2. Unauthorized commitments | The agent promises something the business never approved | Air Canada chatbot's fare policy52 | Policies live in a rules engine; the model can quote them, not invent them |
| 3. Automated adverse decisions | A model's output becomes a denial, rejection or reduction without real review | nH Predict and PXDX allegations55,57; hiring-screen litigation61 | Adverse decisions require a qualified person who reviewed the file, plus an appeal path |
| 4. Distribution shift | A model performs worse in a new setting than where it was built | Epic sepsis model, AUC 0.63 externally67; Zillow Offers71 | Validate locally before launch; monitor drift; cap exposure while confidence builds |
| 5. Learned bias | The model reproduces patterns in historical data that the law prohibits | Amazon résumé screen59 | Bias audits (now mandatory in NYC58); documented criteria; human ownership |
| 6. Injection and over-permission | Untrusted text steers an agent with access to sensitive tools | Indirect injection9; AgentDojo attacks42 | Least privilege; separate reading from acting; schema-constrained outputs10 |
| 7. Cost and scope overrun | Token use, retries and integration work exceed the value created | Gartner cancellation forecast17; McKinsey cost constraint16 | Measure cost per completed task; tune reasoning effort; kill criteria agreed at kickoff |
7.1 The common thread
In each case, a probabilistic output crossed the decision boundary without an independent check. The Air Canada tribunal's reasoning applies more widely: an organization is responsible for what its systems tell customers, whatever the architecture behind them52. Courts sanctioning fabricated citations make the same point about professionals64. The tool does not carry the accountability; the person or organization that deployed it does.
7.2 Zillow: a pre-LLM warning
Zillow Offers predates generative AI, but it is the clearest corporate example of the matrix's lower-right quadrant. The company bought homes on the strength of algorithmic price forecasts. In November 2021 it told shareholders, "We have been unable to accurately forecast future home prices," announced it would wind down the business, and planned to reduce its workforce by about 25%71. Each purchase was costly, hard to reverse and dependent on a forecast that could not be checked until months later. That combination is the one to avoid, with or without a language model.
7.3 The cancellation forecast in context
Gartner's prediction that more than 40% of agentic AI projects will be canceled by the end of 2027 cites "escalating costs, unclear business value or inadequate risk controls"17. It is a forecast, not a measured failure rate, and it refers to projects, not to agents failing at tasks. Read alongside MIT NANDA's funnel (Figure 2) and McKinsey's flat EBIT data16,21, it describes a familiar technology cycle: broad experimentation, then consolidation around use cases that pay. The failure modes above are a fair list of what will separate the projects that survive.
08 — Governance landscapeThe rules are arriving, later than planned, and aimed at decisions
The regulatory picture changed materially in 2026. The EU delayed its high-risk obligations, and Colorado repealed and replaced its AI Act before it took effect. Any compliance plan drafted in 2025 should be revisited.
8.1 Timeline
- Jan 2023NIST AI Risk Management Framework 1.0 — voluntary framework built on Govern, Map, Measure and Manage11.
- Jul 2023NYC Local Law 144 enforcement begins — bias audits and candidate notice for automated employment decision tools58.
- Dec 2023NAIC model bulletin adopted — insurers expected to maintain a written AI program53. ISO/IEC 42001 published, the first AI management-system standard48.
- May 2024Colorado SB24-205 signed — duties originally due from 1 Feb 202645.
- Jul 2024NIST AI 600-1 Generative AI Profile — 12 risk areas, including confabulation, information security, and human-AI configuration12.
- 1 Aug 2024EU AI Act enters into force44
- 2 Feb 2025EU prohibitions and AI-literacy duties apply43,44
- 2 Aug 2025EU general-purpose AI obligations apply43. Colorado later postpones its start date to 30 Jun 2026 (SB25B-004, signed Aug 2025)46.
- May 2026Colorado SB26-189 signed — repeals and re-enacts SB24-205 as an automated decision-making technology law for "consequential decisions"47.
- 27 Jul 2026EU "Digital Omnibus" amendment in force — delays high-risk obligations44.
- 2 Aug 2026EU AI Act generally applicable, with the high-risk exceptions above44.
- 1 Jan 2027Colorado SB26-189 obligations begin47
- 2 Dec 2027EU Annex III high-risk obligations apply (employment, credit, insurance, education, essential services)44
- 2 Aug 2028EU obligations for high-risk AI in regulated products (Annex I)44
8.2 What the EU actually classifies as high-risk
Annex III of the AI Act lists the uses that trigger high-risk obligations. For business deployments, the most relevant are43:
- Employment: systems "to analyse and filter job applications, and to evaluate candidates," and those making decisions on promotion or termination.
- Credit: systems to "evaluate the creditworthiness of natural persons or establish their credit score," except fraud detection.
- Insurance: "risk assessment and pricing in relation to natural persons in the case of life and health insurance."
- Education: systems that determine "access or admission" or evaluate learning outcomes.
- Essential services: eligibility for public assistance benefits; emergency call triage and dispatch.
Every item on that list is a decision, not a technology. A support chatbot that answers questions is not high-risk under Annex III. The same company's model ranking job applicants is. Regulation, in other words, follows the decision boundary.
8.3 A practical compliance posture
| If your system… | Then plan for… | Reference |
|---|---|---|
| Screens, ranks or evaluates job candidates (NYC employers) | Annual independent bias audit, public summary, 10-business-day notice | NYC LL14458 |
| Makes or substantially informs consequential decisions about Colorado residents | Obligations under SB26-189 from 1 Jan 2027; review final text with counsel | SB26-18947 |
| Is used by an insurer in regulated decisions | Written AI program, governance, testing and vendor oversight | NAIC bulletin53 |
| Falls under an Annex III use and touches the EU | Risk management, data governance, logging, human oversight and conformity steps by Dec 2027 | EU AI Act43,44 |
| Is any generative AI system | Map risks to NIST AI 600-1's 12 categories; consider an ISO/IEC 42001 management system | NIST, ISO12,48 |
Not legal advice. This summary is for orientation only. Statutes are changing quickly. Colorado's law was rewritten in 2026, and the EU timetable was amended the same year. Consult counsel on any specific deployment.
09 — Field notesThree systems we built, checked against the evidence
Vital & Co. is a small, founder-led studio. We have no QA department, no on-call rotation and no red team, so every model we put in production has to fail in a log, not in a customer complaint. The three systems below were designed with the principles in this report, and in places they taught us those principles the hard way. They are illustrations, not controlled evidence.
9.1 OfferAIO: an application-filling agent (L2–L3)
OfferAIO fills in internship applications for students on applicant-tracking platforms. Vital & Co. designed and built the original version. A partner operated and extended it from September 2026, and it returned to Vital & Co. in November 2026. We describe the system as we originally built it.
- Code controls the orchestration. A per-platform adapter (Greenhouse, Lever, Ashby, Workday and a generic fallback), built on Playwright14, decides which page to open, which fields exist and when to submit. The model only drafts answers. This is the workflow-first architecture recommended by Anthropic and OpenAI1,2.
- Legal answers never come from the model. Questions about work authorization, sponsorship and age are answered only from the student's stored details. Résumé tailoring rejects any number that does not already appear in the source résumé. Those questions sit in the lower-right quadrant, so they are handled deterministically.
- "Not confident" is a valid output. Each drafted answer carries a confidence flag. A required question with a low-confidence answer goes back to the student instead of being guessed. A submission with no confirmation is recorded as "submitted, unconfirmed," with a screenshot, rather than as a success.
- Typed hand-backs. The engine never bypasses a CAPTCHA, login wall or human-verification page. When it meets one, it returns a typed "needs you" outcome with a one-sentence reason and a link. Everything else stays filled in, and the credit is refunded. This follows the human-checkpoint pattern in the engineering guides and the MCP specification's consent requirements1,13,15.
- Least privilege. The model has no tools. It returns schema-constrained JSON, so the worst a malicious job description can do is affect drafted text, not trigger an action9,10.
- Cost was engineered. When we compared reasoning-effort settings, "medium" cost 2.8 times as much as "low" for near-identical output, so "low" became the default.
9.2 CommitGraphics: a generation pipeline with no agent loop (L1)
CommitGraphics turns an athlete's photo and a short request into a recruiting-announcement graphic. It deliberately has no agent loop: the server selects a locked style template and assembles the prompt.
- Invariants are enforced in code. A prompt that lacks any of the required likeness-fidelity blocks is never sent; the assembly function throws. Rules that must always hold are not left to the prompt.
- When a default and a request conflict, code decides. One case taught us this. A user left a default-on "match uniform to school colors" switch enabled, then typed a request to keep the uniform as photographed. The prompt contained two conflicting instructions, and the model followed the louder one. The fix was a rule in code: when the typed request plainly asks to keep the uniform, the switch is turned off before any prompt is built, and the override is logged. A deliberate request beats a default the user never touched.
- All input is untrusted. Templates live on the server. The client sends a style key checked against an allow-list. Free text is sanitized, length-capped and quoted inside a named block, and a final realism lock outranks it10.
- A person reviews real outputs. Every production render is posted to a private review channel with its style metadata, then updated with whether the user saved it. Star ratings are joined to the configuration that produced each image, which feeds a daily per-style scoreboard. This is the Measure function of the NIST framework in miniature11.
9.3 CoachReach: verified data under a drafting model (L1–L2)
CoachReach helps high-school athletes contact college coaches. It combines a crawled directory of US college coaching staff with model-drafted outreach emails.
- The model drafts; the athlete sends. Outreach emails are drafted by a language model and sent by the athlete from their own email app. Nothing goes out on the athlete's behalf without them pressing send. The consequential action stays with the person.
- Data quality is measured, not assumed. Before launch we ran a verification pass on the directory. A live re-check of sampled records against source pages matched exactly in 98.95% of cases, and head-coach coverage reached 97.8%. A correction pass then reduced the number of high-severity flagged rows from 2,095 to 428. Measuring the deterministic layer matters as much as measuring the model: a well-written email to the wrong coach is still a failure.
- The crawler respects the sources. We identify ourselves with an honest user agent and limit request rates. Where a host's firewall blocks automated access, we ask the operator for permission rather than disguising the crawler. Roughly 600 schools were left out of the directory while we sought that permission. Coverage is a trade-off we made on purpose.
9.4 What the three systems have in common
None of them is an open agent. Each puts the model in the language layer and keeps decisions in code or with the user. Each has a designed way to stop and hand back. Each measures outcomes rather than trusting the model's own report. On the autonomy ladder they sit at L1 to L3, the same range that produces most of the measured value in Sections 3 and 6.
10 — RecommendationsA playbook for 2027
10.1 For executives
- Fund redesign, not licenses. The strongest predictor of AI value in the survey data is workflow redesign16,25. Budget for process mapping, change management and verification tooling.
- Pick workflows by the matrix, not by excitement. Start where outputs are cheap to verify and errors are reversible: support triage, internal search, code, data extraction.
- Set kill criteria at kickoff. Define cost per completed task, accuracy and escalation thresholds before launch. Gartner's cancellation forecast is a reason to decide early, not to avoid starting17.
- Own the outputs. Courts and tribunals have made clear that organizations are responsible for what their AI says and does52,64.
- Do not cut staff ahead of automation. Reported workforce reductions have not translated into ROI20.
10.2 For builders
- Start with a workflow and earn autonomy. Predefined code paths first, as the major labs recommend1,2.
- Put rules in code. Anything that must always be true needs deterministic enforcement that fails loudly5,10.
- Ground every claim. Act on observed data, and give the model an honest way to say "I don't know"4.
- Design the hand-back. Typed stop conditions, a reason and a next step for the person1,13,15.
- Treat all input as untrusted. Least privilege, and no sensitive tools reachable from retrieved text9,10,42.
- Evaluate outcomes, repeatedly. Score final state, not self-reports, and report pass^k6,7. If you use LLM judges, calibrate them against human labels, because they have known biases8.
10.3 Deployment checklist
- Can this be a fixed workflow? If not, write down why.
- Is every step labeled as trigger, decision, language, action or oversight?
- Is every step placed on the verifiability–consequence matrix?
- Is the model limited to the steps that need language?
- Is every invariant enforced in code that fails loudly?
- When a user's explicit request conflicts with a default, does code decide the winner?
- Does every output trace to an observed source, with a "not confident" path?
- Are hand-back conditions listed, typed and shown to the user with a next step?
- Is all user and retrieved text treated as untrusted, under least privilege?
- Are adverse decisions (denials, rejections, reductions) reviewed by a qualified person?
- Is success scored on final state rather than the model's own report?
- Is consistency measured over repeated runs (pass^k)?
- Is there a dry-run mode and a sandbox that never touches real third parties?
- Is cost per completed task tracked, with reasoning effort tuned?
- Does a person review a sample of real outputs on a schedule?
- Is user feedback joined to the configuration that produced each output?
- Has the use been checked against NYC LL144, Colorado SB26-189, the NAIC bulletin and EU Annex III where relevant?
- Are kill criteria written down and agreed?
11 — Outlook to 2028What we expect, and what would change our minds
We make five judgments about the next two years. Each comes with an observable signal that would tell us we are wrong.
| Judgment | Basis | Would change our view |
|---|---|---|
| Agent autonomy will expand first and fastest in software and other domains with automatic verifiers. | SWE-bench and METR trends34,38 | Comparable progress on benchmarks without automatic verification |
| A consolidation wave will cancel many agent projects while survivors scale. | Gartner forecast; MIT funnel; flat EBIT share16,17,21 | The share of firms reporting EBIT impact rising sharply in McKinsey's 2027 survey |
| Consistency, not capability, will remain the deployment bottleneck for customer-facing agents. | τ-bench and τ²-bench7,35 | Published pass^k results near pass^1 on realistic customer domains |
| Regulation will keep targeting decisions, with obligations landing in late 2027. | EU Annex III, Colorado SB26-189, NYC LL14443,44,47,58 | Further postponements, or rules aimed at model capability rather than use |
| Small and mid-sized firms will close part of the gap by buying embedded agents rather than building them. | Census size gap; Menlo buy-vs-build; Gartner app forecast18,22,29 | BTOS adoption among small firms staying flat through 2027 |
Gartner predicts that by 2028, 15% of day-to-day work decisions will be made autonomously through agentic AI, up from zero in 2024, and that 33% of enterprise software applications will include agentic AI17. It separately expects up to 40% of enterprise applications to include task-specific agents by the end of 202618. If those forecasts hold, the question every organization faces will shift from whether to use agents to which decisions they may make. The evidence collected here suggests a defensible answer: the ones that can be checked, and the ones that can be undone.
Work with Vital & Co.
We design and build AI-enabled products and workflows for founders and operators, using the method in this report: map the workflow, place the decision boundary, then build and measure. If you are deciding where agents belong in your business, we would be glad to talk.
admin@vitalsply.com →AppendixCorrections, glossary and references
A. Widely repeated figures we corrected or excluded
| Common claim | What the source actually says |
|---|---|
| "95% of AI pilots fail" (MIT) | "95% of organizations are getting zero return"; 5% of integrated pilots extract millions. The report is preliminary and self-reported21. |
| "40% of AI agents fail" (Gartner) | A forecast that over 40% of agentic AI projects will be canceled by end of 202717. |
| "40% of enterprise apps will have agents in 2026" | The body text says "up to 40%," for task-specific agents18. |
| "AI made BCG consultants 40% better" | The published 2026 version reports quality gains averaging 32%, inside the frontier only69. |
| "88% vs 18% adoption" as a contradiction | Different populations: surveyed (mostly larger) organizations vs all US employer firms24,29,30. |
| IBM "16% ROI" | 16% is the share of initiatives scaled enterprise-wide; 25% delivered expected ROI27. |
| A single count of FDA AI-enabled devices | The FDA list gives no total and says it is not comprehensive; we cite no figure. |
| Current OSWorld best score | Reported figures range widely and we could not read the official leaderboard; we cite launch results only37. |
| Colorado AI Act "in force since June 2026" | SB24-205 was repealed and re-enacted by SB26-189, effective 1 Jan 202747. |
B. Glossary
| Agent | A system in which a model directs its own sequence of steps and tool use1. |
| Workflow | A system in which models and tools are orchestrated through predefined code paths1. |
| Decision boundary | Our term for the line between language work (drafting, explaining) and consequential decisions. |
| pass^k | The probability that an agent succeeds on all k independent attempts at the same task7. |
| Time horizon (50%) | The length of task, in skilled-human time, a model completes with 50% reliability33. |
| Prompt injection | Instructions hidden in input data that steer a model away from its intended task9,10. |
| Confabulation | NIST's term for confidently stated false content, often called hallucination12. |
| Jagged frontier | The uneven boundary between tasks AI does well and tasks it does poorly, which users cannot easily see69. |
| Annex III | The EU AI Act's list of high-risk use cases43. |
C. How to cite this report
Berger, T. (2026). Where agentic AI actually works: A cross-industry analysis of workflow automation, reliability and governance. Vital & Co. Research, 6 November 2026. https://vitalsply.com/research/agentic-ai-2026/
D. References
All URLs accessed between 13 October and 6 November 2026.
- Schluntz, E., & Zhang, B. (2024, Dec 19). Building effective agents. Anthropic. anthropic.com/engineering/building-effective-agents
- OpenAI. (2025). A practical guide to building agents. cdn.openai.com/…/a-practical-guide-to-building-agents.pdf
- Wiesinger, J., Marlow, P., & Vuskovic, V. (2024). Agents [Whitepaper]. Google. kaggle.com/whitepaper-agents
- Yao, S., et al. (2022). ReAct: Synergizing reasoning and acting in language models. ICLR 2023. arxiv.org/abs/2210.03629
- Schick, T., et al. (2023). Toolformer: Language models can teach themselves to use tools. arxiv.org/abs/2302.04761
- Jimenez, C. E., et al. (2023). SWE-bench: Can language models resolve real-world GitHub issues? ICLR 2024. arxiv.org/abs/2310.06770
- Yao, S., Shinn, N., Razavi, P., & Narasimhan, K. (2024). τ-bench: A benchmark for tool-agent-user interaction in real-world domains. arxiv.org/abs/2406.12045
- Zheng, L., et al. (2023). Judging LLM-as-a-judge with MT-Bench and Chatbot Arena. arxiv.org/abs/2306.05685
- Greshake, K., et al. (2023). Not what you've signed up for: Compromising real-world LLM-integrated applications with indirect prompt injection. arxiv.org/abs/2302.12173
- OWASP GenAI Security Project. (2025). LLM01:2025 Prompt injection. genai.owasp.org/llmrisk/llm01-prompt-injection
- NIST. (2023, Jan 26). Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1. nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf; framework page: nist.gov/itl/ai-risk-management-framework
- NIST. (2024, Jul 26). AI RMF: Generative Artificial Intelligence Profile, NIST AI 600-1. nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
- Model Context Protocol. (2025). Specification, version 2025-06-18. modelcontextprotocol.io/specification/2025-06-18
- Microsoft. Playwright. github.com/microsoft/playwright
- LangChain. LangGraph: Interrupts. docs.langchain.com/oss/python/langgraph/interrupts
- Tinkoff, D., Van der Veken, L., Chui, M., & Balakrishnan, T. (2026, Aug 25). The state of AI in 2026: On the road to ROI. McKinsey & Company. mckinsey.com/…/the-state-of-ai
- Gartner. (2025, Jun 25). Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 [Press release]. gartner.com/en/newsroom/…
- Gartner. (2025, Aug 26). Gartner predicts 40% of enterprise apps will feature task-specific AI agents by 2026 [Press release]. gartner.com/en/newsroom/…
- Gartner. (2025, Oct 1). Gartner survey finds just 15% of IT application leaders are considering, piloting, or deploying fully autonomous AI agents [Press release, republished]. prwire.com.au/pr/124493
- Gartner. (2026, May 6). Gartner says autonomous business and AI layoffs may create budget room but do not deliver returns [Press release, republished]. prwire.com.au/pr/128423
- Challapally, A., Pease, C., Raskar, R., & Chari, P. (2025, Jul). The GenAI divide: State of AI in business 2025 (Preliminary findings). MIT NANDA. Copy consulted: mlq.ai/media/…/State_of_AI_in_Business_2025_Report.pdf
- Tully, T., Redfern, J., Das, D., & Xiao, D. (2025, Dec 9). 2025: The state of generative AI in the enterprise. Menlo Ventures. menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise
- Stanford HAI. (2025). AI Index Report 2025. hai.stanford.edu/ai-index/2025-ai-index-report
- Stanford HAI. (2026). AI Index Report 2026. hai.stanford.edu/ai-index/2026-ai-index-report
- Boston Consulting Group. (2024, Oct 24). AI adoption in 2024: 74% of companies struggle to achieve and scale value [Press release]. bcg.com/press/24october2024-…
- Boston Consulting Group. (2025, Sep 30). AI leaders outpace laggards with double the revenue growth and 40% more cost savings (The Widening AI Value Gap) [Press release]. bcg.com/press/30september2025-…
- IBM Institute for Business Value. (2025, May 6). IBM study: CEOs double down on AI while navigating enterprise hurdles [Press release]. newsroom.ibm.com/2025-05-06-…
- Deloitte AI Institute. (2026). State of AI in the enterprise 2026. deloitte.com/global/en/issues/ai/state-of-ai-in-enterprise
- Grundy, A., Breaux, C., & Khatiwoda, D. (2026, May 26). AI use among businesses. US Census Bureau. census.gov/library/stories/2026/05/ai-use-businesses
- Allen, J. S. (2026, Apr 3). Monitoring AI adoption in the U.S. economy. FEDS Notes, Board of Governors of the Federal Reserve System. federalreserve.gov/econres/notes/feds-notes/…
- Cottier, B., Snodin, B., Owen, D., & Adamczewski, T. (2025, Mar 12). LLM inference prices have fallen rapidly but unequally across tasks. Epoch AI. epoch.ai/data-insights/llm-inference-price-trends
- Appenzeller, G. (2024, Nov 12). Welcome to LLMflation. Andreessen Horowitz. a16z.com/llmflation-llm-inference-cost
- Kwa, T., West, B., Becker, J., Chan, L., et al. (2025). Measuring AI ability to complete long software tasks. METR. arxiv.org/abs/2503.14499
- METR. (2026, updated Sep 8). Time horizons and data file
benchmark_results_1_1.yaml. metr.org/time-horizons; metr.org/assets/benchmark_results_1_1.yaml - Barres, V., Dong, H., Ray, S., Si, X., & Narasimhan, K. (2025). τ²-Bench: Evaluating conversational agents in a dual-control environment. arxiv.org/abs/2506.07982
- Zhou, S., Xu, F. F., Zhu, H., et al. (2023). WebArena: A realistic web environment for building autonomous agents. ICLR 2024. arxiv.org/abs/2307.13854
- Xie, T., Zhang, D., Chen, J., et al. (2024). OSWorld: Benchmarking multimodal agents for open-ended tasks in real computer environments. NeurIPS 2024. arxiv.org/abs/2404.07972
- SWE-bench team. SWE-bench Verified leaderboard (accessed 6 Nov 2026). swebench.com/verified.html
- Xu, F. F., Song, Y., Li, B., et al. (2024–2025). TheAgentCompany: Benchmarking LLM agents on consequential real world tasks (v1 Dec 2024; v3 Sep 2025). arxiv.org/abs/2412.14161
- Anthropic. (2025, Jun 27). Project Vend: Can Claude run a small shop? anthropic.com/research/project-vend-1
- Anthropic. (2025, Dec 18). Project Vend: Phase two. anthropic.com/research/project-vend-2
- Debenedetti, E., Zhang, J., Balunović, M., Beurer-Kellner, L., Fischer, M., & Tramèr, F. (2024). AgentDojo: A dynamic environment to evaluate prompt injection attacks and defenses for LLM agents. NeurIPS 2024 Datasets & Benchmarks. arxiv.org/abs/2406.13352
- European Union. (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act), incl. Art. 113 and Annex III. eur-lex.europa.eu/…/OJ:L_202401689
- European Commission. AI Act | Shaping Europe's digital future (incl. Digital Omnibus amendment status). digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
- Colorado General Assembly. (2024). SB24-205: Consumer protections for artificial intelligence. leg.colorado.gov/bills/sb24-205
- Colorado General Assembly. (2025). SB25B-004: Increase transparency for algorithmic systems. leg.colorado.gov/bills/sb25b-004
- Colorado General Assembly. (2026). SB26-189: Automated decision-making technology. leg.colorado.gov/bills/sb26-189
- ISO/IEC. (2023, Dec 18). ISO/IEC 42001:2023 Information technology — Artificial intelligence — Management system. webstore.iec.ch/publication/90574
- Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at work. Quarterly Journal of Economics, 140(2). Working paper: nber.org/papers/w31161
- Klarna. (2024, Feb 27). Klarna AI assistant handles two-thirds of customer service chats in its first month [Press release]. klarna.com/international/press/…
- Shibu, S. (2025, May 9). Klarna is hiring customer service agents… Entrepreneur (reporting a Bloomberg interview). entrepreneur.com/business-news/…/491396
- Moffatt v. Air Canada, 2024 BCCRT 149 (BC Civil Resolution Tribunal, Feb 14, 2024). As reported: dailyhive.com/canada/air-canada-chatbot-decision
- National Association of Insurance Commissioners. (2023, Dec 4). Model bulletin: Use of artificial intelligence systems by insurers. content.naic.org/…/Model Bulletin_Adopted_0.pdf
- Claims Journal. (2025, Mar 7). Lemonade embraced AI in claims from inception (interview with S. Burgess). claimsjournal.com/news/national/2025/03/07/327670.htm
- Estate of Gene B. Lokken v. UnitedHealth Group, No. 0:23-cv-03514 (D. Minn., filed Nov 14, 2023). As reported by Healthcare Dive: healthcaredive.com/news/…/699834
- InsuranceNewsNet / Associated Press. (2025, Feb 14). Minnesota judge: Lawsuit over UnitedHealth use of AI can continue. insurancenewsnet.com/innarticle/…
- Rucker, P., Miller, M., & Armstrong, D. (2023, Mar 25). How Cigna saves millions by having its doctors reject claims without reading them. ProPublica / The Capitol Forum. propublica.org/article/cigna-pxdx-…
- NYC Department of Consumer and Worker Protection. Automated employment decision tools (AEDT). nyc.gov/site/dca/about/automated-employment-decision-tools.page
- Dastin, J. (2018, Oct 10). Amazon scraps secret AI recruiting tool that showed bias against women. Reuters. Syndicated text: euronews.com/2018/10/10/…
- US Equal Employment Opportunity Commission. (2023, May 18). Assessing adverse impact in software, algorithms, and artificial intelligence used in employment selection procedures under Title VII [Announcement]. content.govdelivery.com/accounts/USEEOC/bulletins/35b435e
- Mobley v. Workday, Inc., No. 3:23-cv-00770-RFL (N.D. Cal., May 16, 2025). As summarized: cdflaborlaw.com/blog/…
- Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). The impact of AI on developer productivity: Evidence from GitHub Copilot. arxiv.org/abs/2302.06590
- METR. (2025, Jul 10). Measuring the impact of early-2025 AI on experienced open-source developer productivity. metr.org/blog/2025-07-10-…
- Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), Opinion and Order on Sanctions (S.D.N.Y. June 22, 2023). courtlistener.com/…/575368.54.0_3.pdf
- Magesh, V., Surani, F., Dahl, M., Suzgun, M., Manning, C. D., & Ho, D. E. (2024). Hallucination-free? Assessing the reliability of leading AI legal research tools. Stanford RegLab / HAI. arxiv.org/abs/2405.20362
- Charlotin, D. AI hallucination cases [Database], last updated 5 Nov 2026. damiencharlotin.com/hallucinations
- Wong, A., Otles, E., Donnelly, J. P., et al. (2021). External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Internal Medicine. jamanetwork.com/…/2781307
- Olson, K. D., Meeker, D., Troup, M., et al. (2025). Use of ambient AI scribes to reduce administrative burden and professional burnout. JAMA Network Open, 8(10), e2534976. jamanetwork.com/…/2839542
- Dell'Acqua, F., McFowland III, E., Mollick, E., et al. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. Organization Science. doi.org/10.1287/orsc.2025.21838
- Makridakis, S., Spiliotis, E., & Assimakopoulos, V. (2022). M5 accuracy competition: Results, findings, and conclusions. International Journal of Forecasting, 38(4), 1346–1364. econpapers.repec.org/…
- Zillow Group. (2021, Nov 2). Q3 2021 shareholder letter, SEC Form 8-K Exhibit 99.3. sec.gov/Archives/edgar/data/1617640/…/exhibit993.htm
- Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning: An RCT… Scientific Reports, 15, 17458. pmc.ncbi.nlm.nih.gov/articles/PMC12179260
- De Simone, M. E., et al. (2025, Jan 9). From chalkboards to chatbots: Transforming learning in Nigeria, one prompt at a time. World Bank Education blog. blogs.worldbank.org/en/education/…
- Singh, A. (2026). Agentic ML system design: From messy workflows to intelligent agents. GitHub. github.com/ayush488-glitch/agentic-ai-sys