BLOG

The AI ROI Gap Is a Definition-of-Done Problem

Most enterprise AI spend returns almost nothing. The projects that pay off share one dull trait: everyone agrees on what finished looks like.

Enterprise AI adoption numbers read like a restaurant that keeps buying ovens and never serves dinner.

Spending is up everywhere, and every earnings call has its AI paragraph. Then you read the research. BCG surveyed 1,000 executives and found that only 4% of companies consistently get substantial value from AI, and only 26% have built the capabilities to get past the proof-of-concept stage at all. MIT's NANDA initiative looked at enterprise generative AI pilots and found that 95% of the organizations it studied had no measurable return to show for them.

My read on why is unglamorous. The models are mostly fine. The projects that fail never answered one question precisely: what does done look like? The projects that make money tend to be the ones where the answer is so obvious nobody thought to write it down.

The budget went to the hardest problems first

According to the NANDA report, more than half of corporate AI budgets go to sales and marketing tools. I understand the appeal. Revenue is measurable, and a CFO likes being able to draw a line from "AI" to "pipeline."

The returns point the other way. The report found the strongest returns in the back office: process automation, cutting outsourcing, operational plumbing. Nobody gives a keynote about invoice matching.

Through the definition-of-done lens this stops being surprising. "Generate more qualified leads" is a fuzzy target in an adversarial market. Your competitors bought the same tools, buyers have grown an immune response to AI-written email, and the parts that actually close deals are still human. Invoice matching has a correct answer: the invoice matches the purchase order or it doesn't. The incumbent you're replacing is a spreadsheet and a tired person on a Thursday afternoon.

My guess is that a lot of the struggling companies started with a chatbot, because a chatbot demos well, and invoice matching has never once demoed well.

(If you're trying to figure out where to start instead, this is what I do.)

"Data quality" is usually the wrong diagnosis

The figure that gets passed around is that more than 80% of AI projects fail. RAND repeated that estimate in a 2024 report, then did something more useful than repeat it. Its researchers interviewed 65 experienced data scientists and engineers about why projects fail. Bad data was on the list. But the most common root cause was that people misunderstood or miscommunicated what problem the project was supposed to solve.

That matches what I see. "Data quality" is often a polite name for "we never pinned down the task." You can't tell whether your data is good enough until you know what it's for, so the data initiative launches, runs its course, and leaves the real question exactly where it was.

ChatGPT works well for individuals because the individual quietly supplies the definition of done. They steer it, notice when it's wrong, and throw away the bad output without filing a ticket. An organization needs consistent, auditable output wired into processes that predate the transformer paper. There's no one in the loop doing that judgment for free anymore, so it has to be written down in advance, and writing it down is the step most "AI transformations" skip.

The half of the job nobody wrote down

Some agent failures have little to do with the model. The agent does the job it was given, and the trouble comes from the part of the job nobody wrote down.

In 2022, Air Canada's website chatbot told a customer booking a flight to his grandmother's funeral that he could buy a full-price ticket and claim the bereavement discount afterward. The airline's actual policy didn't allow that. When he asked for the money back, Air Canada argued that the chatbot was a separate legal entity responsible for its own actions. The tribunal rejected that and made the airline pay. The amount was about CA$800, but the precedent costs more. Whatever your agent tells a customer, you told them. The bot's job was to answer questions, and nothing stopped it from making promises the fare rules didn't back.

Lending has the same gap built into the law. In 2023 the CFPB reminded lenders that a complex model is no excuse: if you deny someone credit, the Equal Credit Opportunity Act still requires specific, accurate reasons, and ticking the closest box on a sample form doesn't count. A model can rank applicants well on every benchmark and still be unusable if nobody can say why it turned a particular person down. "Approve good loans" is half the objective. "Explain every denial" is the other half.

Both are specification bugs. The missing half of each objective was exactly the part a human employee would have known without being told, which is why nobody thought to tell the machine.

Humans in the loop wear out

The standard response to agents doing surprising things is a human in the loop: every action gets approved before it runs.

Anthropic has published numbers on how this goes in Claude Code, which asks before running commands. Users approve 93% of those prompts, and Anthropic describes the result as approval fatigue, where people stop paying close attention to what they're approving. Its fix was fewer prompts, not more careful reviewers. Sandboxing cut permission prompts by 84% in its internal use. The safety layer erodes without any attacker involved, purely through the predictable behavior of people who are interrupted all day.

An approval only means something if the reviewer can tell in a few seconds what they're approving. If they can't, you've built a slow rubber stamp and are paying a salary to operate it.

What working looks like

Fraud detection keeps coming up as a genuine success, and its shape is worth copying. One agent flags anomalies, another checks compliance, a third writes up the case, and a human makes the call. The machines do the part that means chewing through enormous volumes of transactions. The person does the judgment, on a handful of well-prepared cases rather than a firehose.

The task is narrow, the failure modes were known before launch, and you can tell afterward whether it worked, because the fraud was caught or it wasn't. Done was defined before anyone wrote a prompt.

That pattern won't fit every use case. It does explain the back-office numbers better than any argument about which model is smartest.


Sources: BCG, "Where's the Value in AI?" (October 2024) · MIT NANDA, "The GenAI Divide: State of AI in Business 2025" · RAND, "The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed" (2024) · International AI Safety Report 2026 · Moffatt v. Air Canada, 2024 BCCRT 149 · CFPB Circular 2023-03 (September 2023) · Anthropic, Beyond permission prompts and How we built Claude Code auto mode