You have read this article before.

Some version of it, anyway. AI changes everything. Agents are the future. Your competitors are already doing it. Here is a statistic about how many companies are failing. Here is a framework with four quadrants. Contact us to learn more.

I’ve read it too. Probably a hundred times this year. So I want to start somewhere else: with a question I’ve asked in about a dozen enterprise rooms over the past eighteen months, and the answer I get almost every time.

The question is this: can you show me the list of decisions your team actually makes?

Not the org chart. Not the process map. Not the dashboard inventory. The decisions: the moments where somebody has to choose, with money or risk or a patient or a customer on the other side of it. Who makes each one. How often it comes up. What changes depending on the answer.

Nobody has the list.

That’s not a criticism. In most organizations, no one has ever been asked to write it down. That missing list is the single best predictor I know of whether an AI program will make it into production or die in a steering committee.

The failure isn’t where everyone is looking

The prevailing story about enterprise AI is that it underdelivers because the technology isn’t ready, or the data is messy, or people resist change. Those are real. They’re also not what the evidence points at.

Look at what practitioners actually report. McKinsey’s Global Survey on the state of AI finds that 88% of organizations now use AI in at least one business function, up from 78% a year earlier, but only about a third have begun to scale, and just 39% report any EBIT impact at the enterprise level. Around 6% qualify as high performers, meaning they attribute more than 5% of EBIT to AI. What separates that 6%? McKinsey is unambiguous: they redesign workflows. They don’t bolt AI onto the existing ones.

Adoption is nearly universal. Impact is rare. And the gap has proved stubborn. McKinsey’s newest edition, published on August 25, 2026, is titled On the road to ROI. Three years in, the destination is still in the title.

Now look at where projects die. Gartner’s forecast is that more than 40% of agentic AI projects will be canceled by the end of 2027, and the three reasons given are escalating costs, unclear business value, and inadequate risk controls.

Read those three again. Not one of them is a model problem.

Cost, value, and control are all demand-side failures. They are what happens when you build capability before you have established what the capability is for. You cannot price a system whose output you never defined. You cannot prove value against a baseline you never took. You cannot control a process nobody ever mapped.

Every one of those failures traces back to the same absent artifact: the list nobody wrote down.

A short detour about a number

Before I go further, I want to do something slightly unusual for an article like this, because it makes the rest of the argument make sense.

You have almost certainly encountered the claim that 95% of enterprise generative AI pilots produce zero return. It is everywhere. It has been in board decks, keynotes, and, I would bet, at least one email you received this quarter.

It comes from a preliminary working paper circulated by MIT’s NANDA initiative in August 2025, built on roughly 52 interviews and distributed behind an email form. Within three days of the coverage breaking, the claim had already hardened in retelling into something the paper did not say.

I raise it because it is a perfect miniature of the problem: a number that sounds authoritative, gets repeated by serious people, and cannot survive being checked. It moved millions of dollars of attention. And when someone finally asked where it came from, there wasn’t much there.

That is exactly what happens to an AI answer inside your business, except the AI produces a thousand of them a day, and nobody is checking any of them either.

Which brings me to what I think the actual work is.

Three things have to be true

I’ve come to believe that AI becomes operational — as opposed to impressive — when three conditions hold. Miss any one and you have a demo.

01 · You know the questions.

Every decision the business makes, by role, with how often it recurs and what moves on the answer. This is the inventory almost nobody has. It sounds administrative. It is actually the whole strategy, because until it exists, every AI investment is a guess about what people will want to ask.

The order matters more than anything else in this article. Most enterprise AI starts with a tool and goes looking for questions to justify it. Starting with the questions and only then deciding what to build is a different discipline entirely, and the difference shows up eighteen months later as the gap between a renewed contract and a quiet cancellation.

02 · You answer inside the decision window.

Decisions are perishable. “Can I accept this contract change?” is worth something today and nothing next week: the negotiation moved on, the site went to another sponsor, the quarter closed. An answer that arrives after the window has shut is a wrong answer, however fast it was.

This is where the distinction between a chatbot and an operational capability actually bites. A chat window answers when someone remembers to go ask it. That’s fine for research. It is useless for a decision that has a clock on it. The answer has to arrive where the work already is, before the person gets to the moment of choosing.

03 · The answer survives contact with a review process.

The test I’d apply to any AI output in a regulated business: can it go into a board pack, an audit file, or a regulatory submission? Or does it stop at the chat window because nobody can vouch for it?

An answer that can’t be checked is a liability someone will eventually have to clean up.

Why “defensible” stopped being optional

If you’ve been treating auditability as a nice-to-have, the last two months should change your mind, though not in the direction the headlines suggested.

In July, the EU adopted the Digital Omnibus on AI, which moved the high-risk obligations of the AI Act from August 2, 2026 out to December 2, 2027, with product-embedded systems following in August 2028. A lot of compliance programs exhaled and re-baselined.

That was a misread. The delay changed the calendar; it changed nothing about the substance. The obligations still land: continuous risk management, technical documentation, automatic event logging, human oversight, conformity assessment. And the transparency obligations were not postponed at all. They took effect on August 2, 2026, on the original schedule.

Meanwhile FINRA, in its 2026 Regulatory Oversight Report, added a section on AI agents for the first time and named four risks with uncomfortable precision: agents acting without human validation; agents exceeding their intended scope and authority; multi-step reasoning that cannot be reconstructed after the fact; and agents mishandling sensitive data.

Notice that none of those are about whether the model is smart. They are all about whether you can say, afterward, what happened and why. Sixteen extra months is time to build that. It is not permission to skip it.

The part your CFO already suspects

There’s a second reason the question inventory matters, and it’s the one that tends to end the meeting fastest.

AI is not behaving like software. Traditional SaaS built once and served the next customer for almost nothing, which is what supported 80–90% gross margins. AI spends real inference on every single query. ICONIQ’s 2026 data puts average gross margin for AI-native companies at about 52%, improving from 41% two years ago but structurally well below the software economics everyone’s financial model assumes.

Inference is cost of goods sold.

And the model invoice is the smallest part of it. In one documented deployment, token cost accounted for just 22% of the fully loaded per-transaction AI cost; tool calls, vector queries, human review, and compliance made up the other 78%. Any program watching only the model bill will mis-cost its own product by a factor of four. Broader practice data says the fully loaded cost of a production workload typically runs two to four times the foundation-model invoice, and that 98% of FinOps practices now manage AI spend, up from 31% two years ago.

The question I’d put to any team proposing an AI investment: what is your cost per resolved question, by role?

If nobody can answer that, the instrumentation gap is the finding. And you cannot even begin to measure it until you’ve defined what a question is and what “resolved” means, which is, once again, the list.

What I’d actually do on Monday

I’m not going to pretend this is easy, but it is concrete, and it doesn’t require a platform decision to start.

Pick one function. Sit down with the people who actually decide things in it, not their managers and not the systems owners. Write down every recurring decision they make, who makes it, how often, and what’s at stake when it goes wrong. You will get somewhere between forty and two hundred. It takes a couple of weeks, not a couple of quarters.

Then do the uncomfortable part. Go through the list and mark each one: can we answer this today with data we actually have, partially, or not at all? For every “not at all,” write down the specific reason: the missing source, the undefined term, the system nobody has access to.

That ratio — answerable over total, calculated per role rather than in aggregate — is the most honest AI readiness score your organization will ever produce. Aggregate coverage flatters you. Per-role coverage tells you that your commercial team is at 70% and your regulatory team is at 12%, and that the second number is the one that will sink the program.

And the list of blocked questions, sorted by how many decisions each fix frees? That’s your build sequence, derived from what your people actually need to know rather than from a vendor’s roadmap.

The order is the whole difference

I said at the top that you’d read this article before. Here’s what I hope was different.

Most of the conversation about enterprise AI is a supply-side conversation: which model, which platform, which partner, which use case from the vendor’s list. It is a conversation about what can be built.

The organizations pulling ahead are having a different one. They started by writing down what their business actually needs to know, in the words the deciders use, and only then asked what to build. It’s slower to start and dramatically faster to finish, because nothing gets built that nobody asked for.

Sixteen months of regulatory runway. A cost structure your finance team hasn’t modeled yet. A cancellation rate north of 40% concentrated in projects that couldn’t prove their own value. Those are the reasons the demand side is the only side worth starting from.

The future belongs to the organizations that can say, precisely and on the record, what they need to know — and prove they knew it.

Most still can’t. That’s the opportunity.

Sources

  • McKinsey & Company, The State of AI: Global Survey. 88% regular AI use in at least one function (vs 78% a year earlier); ~one-third scaling; 39% reporting enterprise-level EBIT impact; ~6% high performers. mckinsey.com. Latest edition: The state of AI in 2026: On the road to ROI, August 25, 2026. QuantumBlack insights index
  • Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, press release, June 25, 2025; based in part on a January 2025 poll of 3,412 webinar attendees. Coverage via BigDATAwire
  • Regulation (EU) 2026/1744 of July 8, 2026 (Digital Omnibus on AI), OJ L 2026/1744, published July 24, 2026. Moves Annex III high-risk obligations to December 2, 2027 and Annex I to August 2, 2028. EUR-Lex
  • Council of the EU, Artificial intelligence: Council gives final green light to simplify and streamline rules, June 29, 2026. consilium.europa.eu
  • Al Jazeera, What came into force with the EU’s AI Act this week – and what didn’t, August 6, 2026. aljazeera.com
  • FINRA, 2026 Annual Regulatory Oversight Report, published December 9, 2025. First dedicated treatment of AI agents. finra.org
  • Debevoise & Plimpton, FINRA’s 2026 Regulatory Oversight Report: Continued Focus on Generative AI and Emerging Agent-Based Risks, December 2025. debevoise.com
  • ICONIQ Growth, State of AI 2026. AI-native gross margin ~52%, up from 41%; inference ≈23% of revenue at scaling-stage AI B2B companies. Via SaaS Mag
  • FinOps Foundation, State of FinOps 2026. 98% of practices now manage AI spend, up from 31% two years prior; fully loaded GenAI workload cost typically 2–4× the foundation-model invoice.
  • Per-transaction cost decomposition (token cost 22% of fully loaded per-loan AI cost). The DAILY BRIEF
  • Theodor Diaconu, The Number Nobody Checked, Cult of Coders. Documents the provenance and spread of the MIT NANDA 95% claim. cultofcoders.com

A note on that last source: the 95% figure is cited here as an example of an uncheckable claim, not as evidence of anything. That is the point of including it.