> ## Content Index
> Fetch the complete content index at: https://www.the-economics-of-enterprise-ai.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# AI Economics Series Article 5 | From pilots to production: why AI ROI is harder than it looks
- URL: https://www.the-economics-of-enterprise-ai.com/ai-economics-series-article-5-from-pilots-to-production-why-ai-roi-is-harder-than-it-looks/
- Published: 2026-08-12T22:27:23.000Z
- Updated: 2026-08-21T11:45:24.000Z
- Author: Tony Kehoe

Enterprise AI has moved into a less forgiving phase.

For the first wave of adoption, the question was whether the technology could do something useful. Summarise a board pack, draft a client email, write code, search policy documents, help a service agent respond faster. Most organisations now have enough evidence to answer yes. The harder question is what happens next.

Can AI change the economics of the business: reduce cost, increase capacity, improve risk control, accelerate revenue, compress cycle time, create a new operating model? And can it do those things after the pilot team has left, when the system is exposed to ordinary users, messy data, audit requirements, security constraints and the politics of real process change?

That is where the argument becomes uncomfortable. A good AI pilot is rarely a smaller version of a good production business case. It is a different creature altogether.

Recent surveys show the gap clearly. AI use is now widespread, with **88%** of respondents saying their organisations use AI in at least one business function. Yet nearly two-thirds have not begun scaling AI across the enterprise, and only **39%** report enterprise-level EBIT impact. [(McKinsey)](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai?ref=the-economics-of-enterprise-ai.com) A separate review reached a similar conclusion. Looking back in January 2026, it judged that at least half of generative AI projects had been abandoned after proof of concept by the end of 2025, citing “poor data quality, inadequate risk controls, escalating costs or unclear business value”. [(Gartner)](https://www.gartner.com/en/articles/genai-project-failure?ref=the-economics-of-enterprise-ai.com) That is a marked revision upward from the 30% forecast for the same date in 2024\. [(Gartner)](https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025?ref=the-economics-of-enterprise-ai.com)

None of which means enterprise AI is failing. It means many organisations are still confusing technical promise with economic conversion.

## **The pilot is the easy room**

A pilot is a controlled environment. The problem is narrowed and the user group is sympathetic. The data has often been cleaned, sampled or hand-selected, and the workflow is simplified with edge cases avoided. Senior sponsors pay attention because the initiative is visible, and expert teams sit close to the business, smoothing away friction that would otherwise become obvious.

Production removes those protections. The system meets inconsistent data, impatient users, legacy platforms, access controls, audit trails, operational handoffs, model drift, exception handling, vendor constraints and change resistance. The model may still be capable and the demo may still be impressive. But the business case now depends on everything surrounding the model.

That is why AI ROI is so easily overstated. The pilot proves that the tool can produce a useful output. Production has to prove that the organisation can absorb that output into work. A proof of concept can show that an AI assistant drafts a client note in seconds. It cannot show that relationship managers will trust it, that supervisors will approve it, that the source data is complete, that conduct risk is acceptable, that review time falls, that the CRM is updated, or that the organisation can turn the saved time into measurable value.

Pilots measure possibility. Production measures change.

## **The productivity trap**

Many early enterprise use cases sit in familiar territory: summarisation, research, meeting notes, document search, email drafting, software development and customer support. These are useful. They make people faster, reduce blank-page effort and remove some of the friction that sits around knowledge work. But productivity is not profit.

If a thousand employees each save twenty minutes a day, the benefit looks large in a spreadsheet. The calculation is seductive: minutes saved, multiplied by people, multiplied by salary. It feels precise. It is usually anything but.

The saving reaches the P&L only if something changes: capacity released, higher volumes handled without additional headcount, service levels improved, cycle times shortened, revenue activity increased, vendor spend cut, controls retired, backlogs brought down. Otherwise the benefit remains real for the individual but invisible to the enterprise.

A global survey of enterprise leaders captures the tension. Two-thirds of organisations report productivity and efficiency gains from enterprise AI. Only **40%** report cost reductions, and only **20%** report revenue increases. Revenue growth remains more ambition than reality: **74%** hope to grow revenue through AI in the future, compared with **20%** already doing so. [(Deloitte)](https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html?ref=the-economics-of-enterprise-ai.com)

That is the productivity trap. AI can make work feel faster before it makes the business perform better. What matters is what the organisation did with the time.

## **The hidden cost of cheap intelligence**

AI also has a more complicated cost structure than most business cases allow for. A single model call may be cheap. A production system is rarely a single model call. It may include retrieval, embeddings, re-ranking, orchestration, guardrails, tool calls, memory, evaluation, logging, monitoring, red-team testing, audit evidence and human review. Agentic systems add another layer, because they plan, retry, search, call tools, maintain context and loop through intermediate steps before producing a final result.

The invoice shows token consumption. The true operating cost is wider than that, and it scales with behaviour: usage, context length, model choice, latency requirements, automation depth, exception rates, human review and governance burden. In conventional software, more adoption usually improves the economics. In AI, more adoption can also mean more consumption.

None of this makes AI uneconomic. It makes economic design essential. A high-value reasoning task may justify a frontier model. A routine classification task probably does not. A sensitive workflow may require sovereign hosting, private deployment or an open-weight model, because the economics include control, data boundary and regulatory costs, not only inference price. A customer service agent that resolves cases end-to-end may be worth a higher consumption cost than a chatbot that drafts suggested replies.

This is where AI financial discipline becomes central. In practice it is an extension of the FinOps discipline enterprises already apply to cloud, now being adapted to token and inference economics. [(FinOps Foundation)](https://www.finops.org/wg/token-economics-saas/?ref=the-economics-of-enterprise-ai.com) Its toolkit is by now reasonably well established: model routing, usage budgets, prompt and semantic caching, consumption telemetry, chargeback models, exception monitoring and clear ownership of unit cost by use case. Harder, and much less settled, is managing to a cost per useful outcome rather than a cost per call.

## **Integration is where the business case breaks**

The model is rarely the hardest part. Value usually sits inside workflows, and workflows are untidy. They run across systems of record, document stores, approvals, controls, customer channels, data entitlements and human judgement. An AI tool sitting beside the workflow may help an individual without changing the workflow itself.

Many pilots operate as overlays. A user copies material into a tool, receives an answer and then manually transfers the output back into the process. That can be useful, but it leaves the economics largely untouched. The user is faster. The process is not redesigned.

Production ROI requires AI to be connected to the next business action. That means identity and entitlement controls, APIs into systems of record, audit trails and data lineage, exception handling, operational ownership and clear decision rights. It also means knowing when the AI can act, when it can recommend and when a human must decide.

A sales assistant that drafts a call note is helpful. A coverage intelligence system that identifies next-best actions, prepares meeting materials, checks profitability, updates CRM, flags constraints and routes follow-up tasks starts to change sales productivity. A document summariser is useful. A regulatory obligation engine that maps rules to controls, tests evidence, escalates gaps and maintains an audit trail changes the economics of compliance. The difference is operating design, not better AI.

## **Controls are not a drag on ROI**

In regulated industries, AI business cases often look stronger before the control model is added. That is a warning sign.

A proof of concept can rely on informal judgement, manual checking and a small group of trusted users. Production cannot. Once AI touches client communication, legal interpretation, credit, conduct, financial reporting, employee data, operational risk or regulated advice, the controls become part of the product. Human validation, access control, prompt and response logging, model evaluation, bias testing, resilience, vendor oversight, auditability and accountability are not optional extras. Together they are the conditions that allow AI to operate at scale.

High-performing AI organisations are more likely to define when model outputs require human validation. [(McKinsey)](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai?ref=the-economics-of-enterprise-ai.com) A survey of 3,235 business and technology leaders across 24 countries found that only **21%** had mature governance in place for agentic AI, even as **74%** expected to be using AI agents at least moderately by 2027\. [(Deloitte)](https://www.deloitte.com/us/en/insights/topics/emerging-technologies/ai-agents-scaling-faster.html?ref=the-economics-of-enterprise-ai.com)

That creates an uncomfortable trade-off. The more autonomous the AI, the greater the potential economic value, and the greater the need for governance. If the business case ignores that governance cost, the ROI is inflated before the system is even built. The question is whether the organisation has designed controls that let AI move safely, rather than whether controls slow AI down.

## **The portfolio problem**

Many organisations now have too many AI pilots and too few AI value systems. Each pilot can be defended on its own terms. One improves research, another supports HR, a third drafts documents or helps developers or summarises customer interactions. Individually, they all look plausible. Collectively, they may not move a single enterprise metric.

This is the portfolio problem. AI investment gets scattered across activity instead of concentrated around value.

Leading companies take the opposite approach. They focus on fewer use cases: **3.5 on average, compared with 6.1** for other companies. They also expect materially higher ROI, and allocate more than **80%** of AI investment to reshaping key functions and inventing new offerings, rather than spreading spend across smaller productivity initiatives. [(BCG)](https://www.bcg.com/publications/2025/closing-the-ai-impact-gap?ref=the-economics-of-enterprise-ai.com)

That is the lesson many enterprises need to absorb. AI value compounds when use cases cluster around an economically meaningful workflow. A bank does not need fifty isolated experiments. It needs a small number of AI-enabled value chains: client coverage, onboarding, credit, financial crime, regulatory reporting, software delivery, service operations. Within those, AI can remove handoffs, compress cycle time, improve decision quality and reduce rework.

The unit of transformation is not the prompt. It is the workflow.

## **A better ROI test**

The standard AI business case is often too narrow. It starts with time saved and turns that into a benefit. Sometimes that is a reasonable first estimate. It should not be the whole case.

Before an AI initiative is funded for production, the organisation should be able to answer a more demanding set of questions. What is the baseline today, in cost, cycle time, error rate, backlog, revenue leakage, control failure, customer outcome or operational risk? What specific mechanism converts the AI output into value, and what adoption level does the case require? What is the full production unit cost, including model usage, orchestration, integration, support, controls and human review? What operating model changes are needed, and how will the benefits actually be harvested?

That last question is often the one that exposes weak cases. A process becomes faster, but no capacity is released. Employees save time while workloads stay unchanged. A team can handle more volume that never materialises, or a control becomes easier to test while the old manual control remains in place. The AI is useful. The economics are not captured.

The distinction between value created and value harvested is an old one. It sits at the centre of benefits realisation management, and it mirrors the longer-standing strategic distinction between creating value and capturing it. What is new is only how rarely AI business cases bother to apply it. Production ROI requires both.

## **What this means for boards**

Boards should be wary of AI portfolios that report activity instead of outcomes. A count of pilots is not a strategy, licences are not transformation, and demo quality is not ROI.

The more useful questions are sharper. Which workflows are being redesigned, which cost pools changed, which revenue processes improved, and which controls strengthened or retired? Which AI use cases have positive unit economics at expected volume? Where human review is required, has it been priced? And where is usage rising without measurable benefit?

This is becoming a CEO-level issue. A 2026 survey found that corporations expect to roughly double AI spending in 2026, from **0.8% to about 1.7%** of revenue, and that half of CEOs believe their job is on the line if AI does not pay off. [(BCG )](https://www.bcg.com/publications/2026/as-ai-investments-surge-ceos-take-the-lead?ref=the-economics-of-enterprise-ai.com) That level of expectation cannot be met through a collection of pilots. It requires industrialisation.

## **Production is the real test**

AI ROI is harder than it looks because access to a capable model does not, by itself, create business value. The real test is in production, whether AI can be embedded into everyday workflows and deliver measurable improvements in speed, quality, cost, revenue or risk. That depends as much on process redesign, trusted data, controls, adoption and management discipline as it does on the technology itself. Therefore AI should be treated not simply as another software purchase, but as an operating-model change whose returns emerge only when technology, people and processes improve together. The operating-model change must be the primary goal and focus, achieved using available technology tools, including AI.

**The pilot proves that AI can work. Production proves whether the organisation can.**

## 

## **Sources**

[McKinsey — The State of AI in 2025](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai?ref=the-economics-of-enterprise-ai.com)

[Gartner — Why Half of GenAI Projects Fail](https://www.gartner.com/en/articles/genai-project-failure?ref=the-economics-of-enterprise-ai.com) (January 2026)

[Deloitte — State of AI in the Enterprise 2026](https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html?ref=the-economics-of-enterprise-ai.com) (January 2026)

[Deloitte — Agentic AI is scaling faster than guardrails](https://www.deloitte.com/us/en/insights/topics/emerging-technologies/ai-agents-scaling-faster.html?ref=the-economics-of-enterprise-ai.com) (January 2026)

[BCG — From Potential to Profit: Closing the AI Impact Gap](https://www.bcg.com/publications/2025/closing-the-ai-impact-gap?ref=the-economics-of-enterprise-ai.com) (AI Radar, January 2025)

[BCG — As AI Investments Surge, CEOs Take the Lead](https://www.bcg.com/publications/2026/as-ai-investments-surge-ceos-take-the-lead?ref=the-economics-of-enterprise-ai.com) (AI Radar, January 2026)

[Gartner — Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept by End of 2025](https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025?ref=the-economics-of-enterprise-ai.com) (July 2024 — the earlier forecast, superseded above)

[FinOps Foundation — Token Economics and SaaS working group](https://www.finops.org/wg/token-economics-saas/?ref=the-economics-of-enterprise-ai.com)