The AI ROI Mirage
Why Your AI Pilot Succeeded and Your Production Deployment Will Not
Five illusions that inflate AI value at the pilot stage and quietly erode it in production.
Abstract
Enterprise AI deployments share a common pattern: a pilot that exceeds expectations, a business case approved on that basis, and a production deployment that quietly underperforms. The gap between the two is rarely attributed to the AI itself. It accumulates instead across five structural illusions built into the standard evaluation framework.
The pilot was engineered to succeed on curated data with motivated users and a measurement window that captured the system at its best. The cost model excluded the governance, oversight, and maintenance layers that dominate the full cost stack. The productivity metric counted hours saved rather than value created, without establishing what the freed capacity would produce. The attribution framework credited the AI for every win and assigned losses to human error. The performance baseline set at launch was never monitored, leaving decay invisible.
McKinsey's 2025 State of AI report found that 78% of enterprises use AI, but only approximately 6% qualify as high performers who have achieved enterprise-wide impact. Goldman Sachs Research in 2024 asked the question directly: given the scale of investment, where is the output? Both findings are the aggregate consequence of the five illusions this paper examines.
The productivity dividend from AI is real. Brynjolfsson and colleagues at MIT documented genuine output gains for knowledge workers under controlled conditions. The gains are meaningful. They are also narrower, slower, and more expensive to sustain than the business cases that get AI projects approved. This paper identifies where the gap lives and what to do about it.
The number everyone cited
Every enterprise AI deployment in the last three years has been preceded by a number. McKinsey says a 40% productivity uplift for knowledge workers. Gartner says a 25% reduction in operational cost. Goldman Sachs says 300 million jobs will be transformed. The numbers circulate in board presentations, vendor proposals, and budget justifications. They are almost never auditable.
The construction of AI ROI studies has a standard architecture, and it is optimised for citation rather than validation. Participants are selected from the users most engaged with AI tools, not the median enterprise user. Tasks are selected because they are amenable to AI augmentation, not because they are representative of the workload. The before-and-after comparison uses self-reported productivity estimates, which are well-documented in the measurement literature to skew in the direction that confirms the researcher's prior. The time window is short enough to capture the novelty effect and short enough to avoid the maintenance cost and adoption friction that accumulate after the first quarter.
The result looks like evidence. It functions as a prior. The organisation that cites the McKinsey number in its AI business case has not validated that the McKinsey conditions (specific tasks, specific user populations, specific data environments) apply to its own context. It has used a published number to anchor an approval process that was going to happen anyway.
The five illusions below are structural features of how AI ROI is estimated, not edge cases. Each has a corrective, and each corrective applies to AI deployments the financial discipline that most organisations already apply to capital investments of equivalent size.
Illusion 1
The pilot paradox
The pilot succeeded because it was designed to. Most enterprise AI pilots share the same hidden architecture: they select the highest-signal use case the team can identify, curate the cleanest data available, assign the most technically capable users as test subjects, and measure success against criteria that were partially visible before the pilot began. The result looks like a proof of concept. It is closer to a demonstration.
Three structural features explain why pilots systematically overestimate production ROI.
Selection bias operates on every dimension. The use case chosen is the one most amenable to AI augmentation: well-structured data, clear success criteria, a task the team understands well enough to evaluate AI output. The users chosen are the most curious, most motivated, most technically literate subset of the eventual user population. The data is as clean as the organisation can make it for the duration of the experiment. None of these conditions survive the full user base.
The measurement window captures the system at its best. A pilot runs for four to twelve weeks and measures outcomes within that window. It does not capture how performance degrades as the system moves from curated to representative data. It does not capture the productivity cost of adoption, where users learn the system, develop workarounds, and restructure their workflow. It does not capture the maintenance cost that accumulates after launch. The ROI calculation freezes at the high-water mark.
The counterfactual is almost never constructed with rigour. The "before" state is often the user's self-report of how long the task previously took, not a measurement. Self-reported productivity estimates skew in the direction that confirms the team's prior expectations. The question of what the user would have produced differently, absent the AI, is rarely asked.
The corrective
Define success criteria before the pilot begins, against a pre-measured baseline, using production-representative data where possible. Treat the pilot as a hypothesis test with a falsifiable outcome, not a demonstration with a predetermined audience. If the pilot cannot be run on representative data, document the gap explicitly and factor it into the business case as a risk, not an assumption.
Illusion 2
The productivity displacement fallacy
The most common metric in enterprise AI ROI studies is hours saved. It is also the least informative. Hours saved is an input metric. Value created is an output metric. The relationship between them is not fixed, and in most enterprise deployments, it is far weaker than the business case assumed.
The logic failure is precise: an AI system that saves a knowledge worker four hours per week has not created four hours of value per week. It has created four hours of capacity. Whether that capacity translates into value depends entirely on what the worker does with it. If the freed time is absorbed by more of the same work (more emails, more reports, more meetings), the value is zero. If it is redirected to higher-value activities the worker previously had no time for, the value is real. If it is absorbed by the overhead of managing and correcting AI output, the productivity gain is negative.
The economic literature on automation has a name for the failure to distinguish capacity from value. In energy economics it is called the rebound effect, or the Jevons paradox: when a process becomes more efficient, the saved resource tends to be consumed rather than preserved. AI tools follow the same dynamic. The analyst who previously spent four hours on a report and now spends two hours on it rarely banks the other two hours as unstructured value-creation time. They produce two reports. Two reports instead of one is a productivity gain only if the second report was actually needed and actually read.
That same McKinsey data puts the displacement problem in relief: 78% of enterprises have deployed AI broadly, but fewer than one in sixteen has achieved enterprise-wide impact. The gap between adoption and performance is the rebound effect made visible. Organisations are acquiring capacity. They are not redesigning workflows to convert it into value.
The corrective
Before deployment, define what the freed capacity will produce. If the answer is "more of the same work," reassess whether the ROI case holds. Value measurement should track output quality and business outcomes, not input time. Hours saved is a leading indicator at best; it is not the result. The business case should specify the value-generating activity that AI-freed capacity will be redirected to, and that redirection should be part of the deployment plan, not an assumption.
Illusion 3
The invisible cost stack
The AI vendor's pricing page shows the cost of the model. It does not show the cost of deploying it, governing it, correcting it, maintaining it, and operationalising the human judgement required to make its output safe to act on. In most enterprise deployments, the model cost is the smallest line item in the full cost stack.
The full cost architecture has five layers, each of which is typically underestimated at deployment.
Integration and deployment. Connecting the model to the organisation's data, authentication infrastructure, and existing workflows is rarely trivial. For enterprise-grade deployments with compliance requirements, this phase routinely costs three to five times the first year's model spend.
Prompt engineering and maintenance. Prompts are not a one-time configuration. They require active maintenance as the model is updated, as the data changes, as the use case evolves, and as edge cases accumulate. Most organisations do not staff this work explicitly; it becomes invisible overhead distributed across the technical team.
Human oversight. For any AI system producing outputs that are acted upon, some fraction of outputs require human review. The review rate varies by domain and risk level, but it is rarely zero. In regulated domains it is often close to 100% for high-stakes decisions. The cost of this oversight is almost never included in the ROI calculation because it is borne by the existing team and does not appear as an AI line item.
Error remediation. AI systems produce errors. Some are detected immediately; some propagate into downstream decisions before they are caught. The cost of remediating propagated errors (correcting the client record, reversing the procurement decision, retrying the failed analysis) is real and material, but it is absorbed by the business without being attributed to the AI system that caused it.
Governance and compliance overhead. Under the EU AI Act, organisations deploying AI systems in regulated contexts are subject to documentation, transparency, and monitoring requirements. The cost of maintaining that compliance posture (risk assessments, conformity assessments, incident reporting, audit trails) is borne by the legal, compliance, and technical teams and is never traced back to the AI project's cost base.
The corrective
Build a full cost model before sign-off. The vendor pricing page is one input. The correct denominator includes integration, prompt maintenance, oversight labour, error remediation, and compliance overhead. For most enterprise deployments, this adds 150 to 300 percent to the visible cost. An AI business case built on model cost alone has not been costed. It has been started.
Illusion 4
The attribution gap
AI attribution is almost universally constructed to maximise credit and minimise accountability. The attribution frameworks applied retroactively to AI deployments share a common architecture: wins are credited to AI, losses are attributed to human error, and the baseline against which AI performance is measured is the worst available version of the pre-AI workflow.
The mechanism is the predictable output of measurement systems designed to justify a prior decision rather than evaluate it. No deliberate distortion is required.
The bias operates across four distinct paths.
Metric selection is the first lever. The metrics chosen to measure AI performance are the ones where the AI performs well. A legal review tool that processes contracts faster is measured on processing speed, not on the error rate in reviewed contracts. A sales-assist tool that generates more outreach volume is measured on volume, not on conversion rates downstream.
Baseline selection is the second. "Before AI" is measured at the worst moment in the process: peak manual workload, least experienced staff, outdated tooling. "After AI" is measured at the best moment: optimised workflow, trained users, curated data. The delta between these two states overstates the AI's contribution.
Authorship attribution is the third. A salesperson closes a deal using an AI-drafted proposal. The AI is credited with the revenue. The salesperson's relationship with the client, their understanding of the prospect's constraints, their judgement about timing and pricing are not separately attributed. AI contributed; the measurement framework awards it the whole outcome.
Losses go unaccounted. The contracts that AI-assisted legal review passed but should have flagged. The prospects that AI-generated outreach alienated. The analyses that contained subtle errors and propagated before detection. These costs are absorbed diffusely across the organisation and are never aggregated against the AI's ledger. The ledger only has a credit column.
The corrective
Design attribution at deployment, not after. Define what the AI system is responsible for, what success looks like, what failure looks like, and how the counterfactual will be constructed. The measurement framework should be agreed before deployment by the parties who will be evaluated against it, and it should be symmetric: the same rigour applied to crediting wins should be applied to surfacing losses.
Illusion 5
The decay curve
The ROI calculation that justified deployment is a point-in-time measurement applied to a system that changes continuously. The accuracy, quality, and business fit of an AI system on the day it goes live is not the accuracy, quality, and business fit of that system twelve months later. The business case does not account for this, and most deployments have no mechanism to detect when performance has degraded past the level that justified the original investment.
Four forces drive decay.
Data drift. Models are calibrated against data distributions that reflect the world at a point in time. As the business changes (new products, new customer segments, new market conditions), the distribution of inputs to the AI system diverges from the distribution on which it was calibrated. Performance degrades gradually, without any system change, as the world moves away from the training assumption. Data drift accumulates without triggering alerts, and it compounds.
Model updates. Foundation model providers update their models regularly. Most updates improve average performance across benchmarks while changing the distribution of specific behaviours. A prompt that produced reliable outputs under one model version may produce different outputs under the next. Most enterprise deployments have no regression testing infrastructure to detect these shifts. Model updates are applied silently, and the behaviour change surfaces when users report unexpected outputs.
Prompt obsolescence. Prompts are written to handle the use cases the team has seen. As the deployment encounters new use cases (which production always does), the prompt's coverage gaps become visible. Patching prompts to handle new cases without breaking existing ones is a skilled engineering task that accumulates complexity over time. Without active maintenance, prompt quality degrades as the gap between the prompt's intended coverage and the deployment's actual use cases widens.
Changing business context. The task the AI was deployed to perform evolves as the business evolves. New regulatory requirements change what outputs are acceptable. New competitive dynamics change what quality looks like. New organisational structures change who uses the system and for what. An AI system designed for last year's workflow is increasingly misaligned with this year's workflow, and the misalignment accumulates invisibly.
The corrective
Define a performance floor at deployment: the minimum acceptable quality level below which the system will be subject to mandatory review or suspension. Build monitoring that tracks performance against that floor in production, and budget the ongoing cost of maintaining performance above it (prompt maintenance, model evaluation, user-feedback processing). An AI system without a performance floor is a system whose ROI is assumed to be constant while the underlying conditions that justified it continuously shift.
The unified picture
The five illusions do not operate independently. They compound.
The pilot produces an inflated baseline. The cost model omits 60 to 70 percent of the real cost. The productivity measure counts hours saved rather than value created. The attribution framework credits the AI for every success and assigns errors to humans. The performance floor is never monitored, so decay is invisible. The result is a business case that was optimistic when approved, was never fully validated in production, and is never revisited unless the system produces a failure dramatic enough to reach the people who approved the original investment.
This describes the modal AI deployment in enterprise organisations today. It explains why the aggregate productivity data is so much weaker than the aggregate adoption data. Organisations are not getting the returns they were sold, and most are not building the measurement infrastructure to understand why.
The productivity dividend is real. The organisations that will generate genuine, sustained ROI from AI are the ones where measurement is honest, costs are complete, attribution is designed before deployment, and performance is monitored against a defined floor. The five illusions are the default. They are also correctable.
AI as a capital investment
The argument of this paper, compressed: AI deployments should be evaluated with the same financial rigour applied to capital investments of equivalent size and risk.
Capital investments are subject to NPV analysis, payback period calculation, sensitivity analysis against key assumptions, and post-implementation review against the original business case. These disciplines are routinely applied to equipment purchases, facility expansions, and technology infrastructure projects. They are not being applied to AI.
The same organisations that apply these disciplines to a €500,000 server purchase are approving €2 million AI deployments against a business case built on a pilot study, a vendor-supplied ROI estimate, and hours-saved projections that have never been subjected to counterfactual analysis. The gap is a failure of financial governance applied inconsistently across investment categories.
NPV analysis
Discount the full cost stack (not just model costs) against a realistic, sensitivity-tested value projection. The vendor's ROI estimate is an input, not the model.
Sensitivity analysis
What happens to the business case if productivity displacement is 40% of the projected gain? If error remediation costs double? If adoption takes 18 months instead of 6? The base case is one scenario, not the forecast.
Post-implementation review
12 months after deployment, measure actual ROI against projected ROI, using the same metrics defined in the original business case. If the review is not scheduled before deployment, it will not happen.
Performance floor
Define the minimum acceptable quality level and build monitoring to enforce it. An AI system below its performance floor is not delivering the ROI the business case projected, regardless of what it costs to run.
Attribution framework
Agree before deployment how AI contribution will be separated from human contribution in the measurement. The framework should be symmetric: the same rigour applied to crediting wins should be applied to surfacing losses.
None of these disciplines change the technology. They change the decision-making architecture around it. The organisations that apply them will make fewer AI investments, deploy more carefully, and generate more genuine value from each deployment. The measurement gap is a governance problem. Governance problems are solved by decisions, not discoveries.
Conclusion: the measurement gap
The AI industry has produced a large quantity of ROI claims and a small quantity of auditable measurement. This is the natural output of a market in which vendors are incentivised to quantify upside and buyers are incentivised to justify approved investments. The measurement infrastructure to distinguish genuine ROI from a well-constructed narrative has not yet been standardised, and until it is, the headline numbers will continue to overpromise.
The five illusions this paper describes are not properties of any particular vendor or tool. They are properties of the evaluation framework the industry has settled on by default: pilots run under conditions that cannot be replicated, cost models that exclude the layers that dominate in practice, productivity metrics that count activity rather than value, attribution designed to confirm rather than test, and performance baselines that are set once and never revisited.
What changes when organisations close these gaps is not the technology. The same models, the same tools, the same use cases produce materially different outcomes when the measurement is honest, the cost model is complete, and the performance floor is monitored. Brynjolfsson's research shows what that looks like: genuine, measurable improvements in output quality for specific tasks, achieved under conditions the organisation understood and maintained. That is the achievable version of the AI productivity dividend. It is smaller than the vendor-projected number and larger than the unmeasured default.
The work is in applying to AI the financial discipline that already exists for every other capital investment of comparable size. The tools are not new. The decision to use them is.
References
- McKinsey & Company (2025). "The State of AI: How Organizations Are Rewiring to Capture Value." McKinsey Global Institute.
- Brynjolfsson, E., Li, D., & Raymond, L. R. (2023). "Generative AI at Work." NBER Working Paper No. 31161.
- Gartner (2025). "Hype Cycle for Artificial Intelligence, 2025." Gartner Inc.
- Goldman Sachs (2024). "Gen AI: Too Much Spend, Too Little Benefit?" Goldman Sachs Research.
- IBM Security & Ponemon Institute (2025). "Cost of a Data Breach Report 2025."
- European Parliament (2024). "Regulation (EU) 2024/1689: Artificial Intelligence Act."
- KPMG (2025). "Enterprise AI Adoption Study: The Value Realisation Gap." KPMG International.
- Stanford University HAI (2026). "Artificial Intelligence Index Report 2026."