Read enough recent enterprise AI commentary and the recommendations begin to sound reassuringly consistent: move past pilots, redesign workflows instead of layering AI onto existing tasks, build the data foundation, assign accountable leadership and invest in people. Govern agents, measure value rather than usage, and move before competitors do.
The agreement is substantive. These are not frivolous ideas, and most of the leading consulting firms are more careful about their qualifications than the slogans extracted from their work. But once the recommendations travel into executive presentations, transformation roadmaps and public commentary, an association can start to sound like a cause, a forecast like an observation, and a sensible management principle like a necessary step toward success.
For a CEO or board, that is the uncomfortable point. The advice can be sound and still leave open the questions that determine whether to invest: What is the opportunity for this enterprise, what must change, and at what cost? What would count as success, and what would persuade us to stop?
We examined recent consulting and analyst publications alongside government statistics, firm-level economic studies, experiments, organizational research and disclosures from companies deploying AI. Our aim was neither to rank advisory firms nor to demand a randomized trial for every management judgment. It was to ask what each source establishes, what reasoning connects it to a prescription, and how that prescription should inform an enterprise decision.
The advice is converging, but the evidence is not all of the same kind
| Recurring prescription | Representative voices | What the evidence supports | What remains open |
|---|---|---|---|
| Redesign work, not just individual tasks | McKinsey, BCG, Deloitte | Better-reported outcomes accompany redesigned workflows; measured task gains can attenuate before delivery | Which redesign causes which outcome, at what cost? |
| Build data and operating foundations | IBM IBV/Oracle, BCG | Reliable access, permissions and integration matter for consequential work | Which architecture is economically best for a given enterprise? |
| Govern AI and agent autonomy | BCG, KPMG, EY | Policies, controls and accountability are widely reported; policy can diverge from practice | Which controls actually detect or prevent important failures? |
| Train people and redesign roles | BCG, Deloitte, PwC | Targeted practice and AI assistance can improve performance in tested tasks | Does participation become transferable, durable expertise? |
| Measure value and use stopping criteria | Gartner, KPMG | Organizations report major measurement gaps; stopping rules accompany better-reported results | What would count as attributable net benefit in this business? |
| Move quickly and keep adapting | McKinsey, BCG, Deloitte | Capability is changing and implementation creates organizational learning | When is commitment valuable, and when is reversibility worth paying for? |
This is a map of the advice, not a scorecard. Similar practices appear alongside reported success across several genuinely distinct survey samples. That convergence deserves attention, although reliance on executive self-report, differing definitions of “high performance” and respondent selection can create shared blind spots. The table cannot determine the right decision for a particular company.
A firm basis for action already exists
We should not lose the positive evidence while scrutinizing claims. Controlled studies show that AI can improve task performance; European firm-level research estimates a positive productivity effect under a stated econometric approach; and software-development evidence finds more completed releases as well as far more coding activity. Task-specific AI-assisted feedback has also improved measured workplace performance. These findings do not mean every enterprise transformation will pay off. They do establish a serious opportunity, and a reason to investigate where gains arise, what it takes to deliver them and how the enterprise might capture their value.
1. What is actually happening, and what are we counting?
Three headline figures illustrate the measurement problem. In McKinsey’s August 2026 survey, 80% of respondents report improvements in their individual productivity, but 37% attribute at least some positive earnings-before-interest-and-taxes impact to AI. KPMG’s Global AI Pulse reports that 89% of organizations see measurable value, although only 12% consistently assess value against cost. Gartner’s October CIO survey finds 13% reporting significant or transformational enterprise value.
These are not three estimates of one underlying ROI rate. The populations, questions and outcome thresholds differ, and a respondent’s attribution of financial impact is not an audited measure. The surveys help us understand what leaders report experiencing and which practices accompany those reports. Averaging their percentages would obscure rather than clarify the picture.
What does “AI value” actually mean?
Four published numbers. Four different outcomes. They should not share one ROI axis.
Report improved personal productivity
Respondents describing their own productivity, not audited enterprise returns.
Report “measurable value”
A broad senior-leader survey measure, not the same threshold as significant enterprise impact.
Report significant or transformational value
A higher outcome threshold in a CIO and technology-executive survey.
Businesses report using AI
A broad business survey, including firms beyond the most AI-active large enterprises.
Read the label, not just the percentage. Adoption, local productivity, reported value and significant enterprise outcomes are different constructs, with different respondents and thresholds.
The UK government’s October 2026 AI adoption report changes the population. Among 3,503 surveyed businesses, 32% report using AI. Among adopters, 50% report productivity improvements, generally described as modest, while 64% report no immediate operating-cost change. These are self-reports too, but across a more representative business population rather than a panel weighted toward leaders actively pursuing AI. They suggest a great deal of use is still incremental.
Independent operational studies provide another lens. In the September revision of Writing Code vs. Shipping Code, Demirer, Musolff and Yang draw on more than 500,000 GitHub developers. Their matched-event analysis estimates cumulative coding-activity gains of 30%, 180% and 240% across successive generations of tools. For the most autonomous generation, the 240% gain in coding activity translates into an estimated 30% increase in software releases. That is a substantial gain in delivered output, but much smaller than the increase in activity. The two deserve separate measures.
There is also broader evidence of productivity effects. A BIS/EIB working paper using matched firm data estimates a roughly 4% short-run labor-productivity gain under its instrumental-variable approach, with a positive alternative estimate using value added per employee. The study does not measure the 2026 frontier-agent economy. Its observations largely cover 2019 to 2024, the adoption variable combines big-data analytics and AI, and the primary productivity measure is turnover per employee. Those qualifications matter, without negating the positive result the researchers estimate.
Company reports illustrate a related issue. UBS’s own October 2026 results disclose more than 12,700 employee-run personal AI agents, alongside 608 live use cases. The figures describe a substantial reported deployment effort, but UBS does not define those agents’ degree of autonomy or isolate a corresponding net financial return. The absence of an isolated disclosure does not mean there was no benefit.
A decision worth taking now: keep agent counts, licences, prompts and course completions in the activity and adoption columns, not the value column, of an executive dashboard. Report separately what work improved, what reached customers or operations, what changed economically and what evidence links the change to AI.
2. Where will the enterprise create value, and who will capture it?
A decision to pursue AI should begin with the business outcome rather than the technology’s availability. Efficiency is one pathway, alongside better service quality, faster delivery, additional capacity, avoided losses, new products, wider market reach and differentiation. Each has distinct economics, measurement requirements and competing claims on capital.
The consulting work is broader than a cost-reduction narrative. BCG’s Applied AI Index argues for concentrating AI on the few opportunities that matter to an enterprise’s strategic position; its 2026 classification places almost half of surveyed companies in value-generating categories. McKinsey also emphasizes organizational change, while Wharton’s 2026 industry-partnered report finds that 35% of enterprises tracking at least one growth outcome report strongly positive ROI, compared with 13% tracking efficiency alone. That is a descriptive association, not evidence that merely adding a growth metric causes the return. Stronger firms may already possess better growth opportunities, leadership or capabilities.
The more useful economic distinction is between creating value and capturing it. A team may save time without reducing expenditure or using the released capacity. Faster product development matters most when there is demand, a route to market and an advantage worth exploiting. Quality improvements can preserve customers or prevent expensive errors even when immediate cost savings are small. And if a productivity improvement becomes widely available to competitors, some of its benefit may reach buyers through lower prices rather than the adopting firm’s profits.
Ranking growth above efficiency on this basis would simply replace one universal prescription with another. A manufacturer defending margins, a bank protecting the reliability of advice, and a growing software business competing on product speed can rationally invest in quite different AI capabilities. A focused, measurable efficiency gain may be a better investment than a speculative new offering; in other circumstances, a new capability may matter more than near-term cost savings.
A decision worth taking now: identify the value pathway for each significant initiative before selecting the application. Specify the beneficiary, how the benefit reaches an outcome, who captures it, and what alternative use of capital it must outperform. A business case framed around “employee productivity” alone is not enough.
Producing more is not the same as capturing more value
A useful outcome has to survive several different tests, not just the first performance benchmark.
Capability
The technology can perform or assist with a task.
and context
Task output
Work is faster, greater in volume or better in quality.
and integration
Delivered outcome
Approved work reaches users, customers or operations.
and economics
Captured value
The enterprise realizes a consequence worth the full cost.
The same revised study measures different outcomes for autonomous coding agents. The release gain is real; the gap is not a general AI-to-value conversion rate.
3. What turns AI potential into results?
On workflow redesign the consulting firms do have a meaningful convergence. McKinsey finds it much more prevalent among respondents classified as AI high performers. BCG advocates a limited number of consequential end-to-end workflows. Deloitte’s 2026 agentic research documents a gap between expectations of process transformation and reported readiness to do it.
It would be a mistake to dismiss these patterns merely because the surveys are observational. Separate samples reinforce a real proposition: organizations that report good outcomes often also report changing the work around the tools. What they do not establish is which element caused success or whether a broad reorganization beats a targeted intervention. Strong management, investment capacity and existing technology may help explain both the redesign and the outcome; early success may also provide the money to pursue further redesign.
The software evidence points to a practical question: where is the constraint? Generating more code can create code-review backlogs, rework and integration costs; some verification is necessary protection, not dispensable bureaucracy. A randomized workplace study across 66 firms and 7,137 workers found that some AI users spent less time on email, without detected broad changes in the composition of work. Narrow improvements can matter even before wholesale redesign. The study’s researchers had Microsoft affiliations, relevant when evaluating the tool-specific setting.
MIT Sloan’s September research adds a missing ingredient: the largely invisible work of experimenting, checking, coordinating and refining AI-enabled solutions until they become usable across an organization. That work takes people, ownership and time. It is a valuable explanation of why installation is not implementation, though it does not quantify the universal ROI of change management.
Data access and integration belong in this discussion, but they do not settle a platform strategy. IBM Institute for Business Value’s report, produced with Oracle, advocates a unified governed data foundation. Reliable context, permissions and interfaces are credible requirements for consequential systems. The available literature does not demonstrate that this particular architecture consistently beats a modular solution, an incremental integration or another design after accounting for implementation cost and switching risk.
A decision worth taking now: trace one important workflow from input to final business outcome. Locate where quality degrades, work accumulates, judgment is needed or value fails to be captured. Make the smallest change that can test the suspected constraint; redesign more broadly when the evidence from that workflow and the value at stake warrant it. Do not mistake every human check for waste—or every existing process for something that must be retained.
4. How much autonomy and which safeguards make sense?
The word agent is increasingly unhelpful without a description of authority. An assistant that drafts a response, a bounded agent that updates records, and a system allowed to execute transactions without approval present different opportunities and risks. Their counts cannot be compared as though they were the same operational unit.
“Agent” is not a unit of authority
The right safeguard depends on what a system can do, the consequences of error and whether mistakes can be detected or reversed.
BCG’s October governance analysis reports that companies scaling innovation alongside governance have around three times the reported ROI of low-innovation companies. That is a useful association, not evidence that installing a governance framework multiplies returns by three. EY’s survey of 202 senior AI decision-makers finds that 98% report formal policies while 47% acknowledge having bypassed governance processes for urgent deployments. A written policy and an operational control are not interchangeable.
Research on human–AI collaboration challenges the assumption that adding a human reviewer necessarily produces the best outcome. A 2024 meta-analysis in Nature Human Behaviour, examining 106 experimental studies, found that combined systems outperformed humans working alone on average, but performed significantly worse than the better of the human-only and AI-only alternatives. The effects varied by task and by the relative strengths of people and AI. Because the underlying studies were published through mid-2023, this is not a direct evaluation of today’s agent systems; it is evidence against assuming automatic human–AI synergy.
A more recent Anthropic experiment deliberately gave agents preferences between plausible analytical choices. Analysts using them often adopted those preferences. It demonstrates a mechanism by which nominal human review may inherit AI framing. The setting was deliberately manipulated and narrow, and the agent-assisted analysts also worked faster and made fewer serious errors than a small comparison group. We should learn from both observations rather than extrapolate a general enterprise failure rate from one experiment.
Verification is not simply a cost to eliminate: it can slow throughput while preventing expensive downstream errors. More checks, however, are not necessarily better checks. Unless reviewers have the evidence, competence and authority to detect specified failures, approval may become a ritual, while blanket human sign-off can make low-risk work unnecessarily expensive.
A decision worth taking now: decide autonomy at the workflow and failure-mode level. State what the system may do, what requires approval, what errors matter, what the safeguards must catch, and how their effectiveness will be tested. Include review, remediation and monitoring in the economics. If a control is legally or professionally mandatory, distinguish its compliance role from its contribution to factual correctness.
5. Are enterprises developing competence, or mostly recording participation?
Consulting firms rightly emphasize people and organizational change, but the scale of a training program says little about the skill available in a consequential workflow. Course completion measures participation; reliable performance, exception handling and judgment under uncertainty measure capability.
A pre-registered three-month experiment with 133 patent lawyers found AI-assisted drafting improved task quality. In a later unaided professional-judgment exercise, the mean improvement was concentrated among senior lawyers; juniors showed no average gain, with their scores shifting toward both better and worse performance. The paper discloses that Google conducted the experiment, funded its direct costs and employed several coauthors. Its single-profession design also limits generalization. PwC’s September workforce survey finds that 64% of nearly 50,000 respondents used AI at work during the previous year. It also identifies a core-workforce cohort comprising 56% of respondents, with comparatively less access to AI, learning and development opportunities. This is worker-reported experience, not proof of an intervention’s effectiveness. Together, these sources ask a serious question: what happens to the development of expertise when people can produce good work without practicing all the underlying reasoning themselves?
There is affirmative evidence about training too. In a financial-services call-centre study, AI-assisted feedback reduced handling time by around 10%. But using AI to deliver task-specific feedback is different from teaching employees to use AI in a generic awareness course. Neither study justifies treating all training interventions as equivalent. Meanwhile, the MIT Sloan research suggests that useful organizational capability is built through repeated practical adaptation, not only formal instruction.
The challenge for workforce strategy is to develop people who can use AI effectively today while retaining the capacity to recognize errors, solve unfamiliar problems and supervise more capable systems tomorrow. Uncritical assistance may weaken some learning pathways; designed practice, feedback and challenge may strengthen others. The evidence supports neither inevitable deskilling nor automatic upskilling.
A decision worth taking now: replace completion-only reporting with task-based assessments. Test employees on realistic cases, difficult exceptions and, where relevant, situations in which AI assistance is absent or wrong. Treat training as complete when capability is demonstrated, not when a module is finished.
6. How should leaders commit resources without pretending the future is settled?
Enterprise AI investment carries competing risks. Delay can forfeit opportunities and the practical learning needed to exploit future capabilities, while premature commitment can strand systems, create dependence and absorb management attention before the business case is credible. Neither “move immediately” nor “wait for proof” resolves that trade-off.
Here the advisory advice on measurement deserves to be taken seriously. KPMG reports that only 12% consistently measure value against cost. Gartner reports that 31% of CIOs have guidelines for keeping unsuccessful AI pilots out of production, and 17% have criteria for retiring low-value production initiatives. Organizations with such rules are more likely to be in Gartner’s stronger-performing category; that is an association, not the measured effect of installing a stopping policy. Nevertheless, an investment without a success measure or an exit condition is difficult to govern intelligently.
The complete cost includes more than model calls and software licences: data preparation, integration, exception handling, rework, human review, security, operational monitoring, training, changes to work and possible switching costs. The alternative might be a deterministic rules engine, conventional software, a process change or doing nothing. Comparing AI only with the unmodified current process can systematically flatter the AI business case.
Models are also improving. A 5 November METR assessment at Anthropic found evidence of accelerating progress on well-specified research tasks, but Anthropic redacted much of the evidence most informative to METR and attribution to AI remained uncertain. That is one reason to avoid treating any particular technical performance assumption as permanent—not a prediction of enterprise ROI or an argument to postpone all investment.
A sensible response is staged commitment with proportionate reversibility. Experimentation can begin precisely to build understanding, while costly architecture or automation choices should be justified against stronger evidence of fit. Major decisions benefit from retaining a short record of what was assumed, which alternatives were rejected, what result was expected and what would trigger reconsideration. That is ordinary decision discipline, not proof of a financial return from specialized reasoning infrastructure. Nor should review processes multiply without regard to their own costs.
A decision worth taking now: require a baseline, a plausible counterfactual, full operating costs and explicit scale/stop conditions before treating a pilot as a transformation program. For commitments that are hard to reverse, insist on a documented reason for the lock-in and what changing assumptions would require a new decision.
What should a board ask of an AI business case?
The wider research does not produce a universal AI transformation formula. It does produce a better set of questions for decisions that are already being made. These are not additional bureaucracy; they are the questions that make the recommendation testable.
Seven questions before an AI business case scales
Not a new framework to implement. A set of challenges to put to a consequential proposal.
What business outcome are we buying?
Identify the customer, operating, competitive or risk result, and how the enterprise captures it.
Compared with what?
Establish the baseline and a credible alternative, including conventional software or another use of capital.
Where will the gain be absorbed?
Trace verification, rework, integration and capacity that cannot be economically used.
What must be true here?
Test data, permissions, ownership, incentives, skills and workflow prerequisites.
What may the system do, and do safeguards work?
Specify permissions, failure modes, reviewer competence, escalation and shutdown; test them.
What constitutes demonstrated capability?
Assess performance and judgment in consequential work, not seats, credentials or output counts.
What would justify scaling, changing or stopping?
Include total lifecycle costs, decision gates, alternatives and reconsideration triggers.
Read the seven questions in the original article wording
- What business outcome are we buying? Specify the customer, operating, competitive or risk outcome rather than tool adoption or activity, and explain how the enterprise captures the benefit.
- Compared with what? Establish the baseline and a credible alternative: the existing process, conventional automation, a narrower workflow improvement or another use of capital.
- Where will the gain be absorbed? Trace the end-to-end work. Include verification, rework, integration, human attention and any capacity the organization cannot actually use.
- What needs to be true about this enterprise? Identify the data, permissions, leadership, incentives, skills and operational prerequisites. Decide which gaps are material and which can be addressed incrementally.
- What authority are we delegating, and how will we know safeguards work? Specify tasks, permissions, failure modes, reviewer competence, escalation and shutdown arrangements. Test the controls on cases that could actually go wrong.
- What constitutes demonstrated capability? Measure performance in consequential work, not simply seats, usage, credentials or AI-generated output.
- What are the costs, decision gates and reconsideration triggers? Include the full lifecycle and alternatives, and state what would justify scaling, redesigning, pausing or stopping after deployment.
A mature organization may already have much of this discipline. A smaller enterprise can apply it without creating a new governance apparatus. A regulated institution will legitimately require a deeper account of failure, authority and duty. The method scales with consequence; it is not a universal administrative burden.
The point is not to become suspicious of every claim. There is credible evidence that AI can improve important work, and enough possibility to warrant strategic attention and experimentation. There is also strong reason to distinguish what firms report doing from what they achieve, to test whether familiar management practices are producing results, and to connect investments to the distinctive priorities of each enterprise.
The consulting consensus offers useful directions, and wider evidence helps clarify the conditions and uncertainties that accompany them. The decisions remain with leadership.
There is enough evidence to take AI seriously. There is not enough to outsource enterprise judgment to a popular playbook.
Selected sources and reading notes
This is an evidence-informed opinion article, not a systematic review. Sources include surveys, government statistics, observational economics, experiments, organizational research and first-party vendor work. They are not equally independent or suitable for the same inference. Most contemporary prescriptions are from August–November 2026; earlier research is included for specific mechanisms.
Consulting, analyst and enterprise sources
- McKinsey, The state of AI in 2026: On the road to ROI, 25 August 2026. Executive self-report; the 80% productivity, 37% EBIT and 6% high-performer figures describe different thresholds of reported results, not audited returns.
- BCG, The Formula for Agentic AI Value: The Applied AI Index 2026, 30 September 2026. Survey-defined maturity and value categories; comparative performance associations and forecasts require careful interpretation.
- BCG, How to Create AI Value by Balancing Innovation and Risk, 26 October 2026. Applied AI Index (1,330 leaders), a separate governance-leader survey (142) and interviews; 3× comparison is against low-innovation firms.
- Deloitte, AI Agents are Only the Beginning: Deloitte Survey Examines the AI Readiness Gap and Reveals How Enterprises Can Prepare for Agentic Success, 12 August 2026. Agentic readiness and redesign reporting; survey selection differs from nationally representative adoption measures.
- KPMG, Global AI Pulse Q3 2026: 89% of organisations report value from AI, but only 12% consistently measure value against cost, October 2026 country release. Senior-leader survey; its “measurable value” wording is not an audited or comparable enterprise EBIT measure.
- EY US, EY survey finds that autonomous AI implementation outpaces oversight, yielding an AI governance gap, 15 September 2026. Self-report from 202 senior AI executives at large US listed companies.
- Gartner, Gartner Survey Finds Only 13% of CIOs Say Their Enterprise Has Realized Significant or Transformational Value from AI, 19 October 2026. CIO self-report; stop-rule comparisons are associations.
- Wharton / GBK Collective, 2026 AI Adoption Report, October 2026. Industry-partnered survey; growth-measure and ROI association is not a causal estimate.
- IBM Institute for Business Value / Oracle, Agentic architecture research, 2026. Co-produced vendor-partnered architecture prescription; no independent like-for-like comparison of platform returns in this corpus.
- PwC, Global Workforce Hopes and Fears Survey 2026, 29 September 2026. Survey of 49,364 workers; reported AI use and uneven development opportunities, not an intervention experiment.
- UBS, Third-quarter 2026 results, 28 October 2026. Issuer disclosure of personal agents and use-case counts, not independent financial attribution.
Institutional, experimental and other evidence
- UK Department for Science, Innovation and Technology, AI Adoption Survey 2026: Executive Summary, 29 October 2026. Broad business survey; productivity and cost effects are self-reported.
- Aldasoro et al., AI adoption, productivity and employment: evidence from European firms, BIS Working Paper 1325, January 2026. Firm-data instrumental-variable study; mostly 2019–2024 observations; bundled AI/big-data adoption measure, turnover per employee as primary productivity outcome, alternate value-added result.
- Demirer, Musolff and Yang, Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools, NBER Working Paper 35275, September 2026 revision. Matched-event study rather than randomized trial; 500,000-plus developer figure and 30/180/240% coding-activity estimates are from the revised abstract.
- Dillon, Jaffe, Immorlica and Stanton, Shifting Work Patterns with Generative AI, NBER Working Paper 33795, May 2025; November 2025 revision. Randomized study of 7,137 workers across 66 firms; Microsoft-related affiliations are disclosed.
- MIT Sloan Office of Communications, Why Some Organizations Turn AI Experiments into Business Value While Others Quietly Fail, 9 September 2026. Press account of field research by Katherine Kellogg, Batia Wiesenfeld and Arvind Karunakaran on the often-unrecognized work of implementing AI solutions.
- Vaccaro, Almaatouq and Malone, When combinations of humans and AI are useful: A systematic review and meta-analysis, Nature Human Behaviour, 28 October 2024. Meta-analysis of 106 experimental studies published from 2020 through mid-2023; mean performance of human–AI teams was worse than the better standalone condition despite improvement over humans alone.
- Autor et al., Does AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent Drafting, NBER Working Paper 35720, September 2026 issue; October 2026 revision. Preregistered trial with 133 patent lawyers. The paper discloses that Google conducted the study, funded its direct costs and employed several coauthors.
- Fouarge et al., How AI-Augmented Training Improves Worker Productivity, IZA Discussion Paper 18224, October 2025. AI-supported task-specific feedback, not generic AI-literacy training.
- Anthropic, How AI Agents Shape Experts’ Judgment Calls, October 2026. First-party experiment deliberately assigning preferences; a mechanistic warning, not an estimate of enterprise oversight failure prevalence.
- METR, Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, 5 November 2026. High-access external assessment; major redactions and uncertain attribution limit public scrutiny. Context for changing capability, not enterprise ROI evidence.