Where AI Actually Pays in an Agency. The skeptic was half right.
The most quoted AI statistic says 95 percent of pilots returned nothing. The findings underneath it are more useful: what the paying minority bought, how they measured it, and the risk most agencies are already carrying without having bought anything at all.
The invoice arrives every month now: the seats, a couple of specialist tools, an enterprise plan someone signed up for in the spring. The team says the tools save time, and you believe them. But nobody can point to where the saving shows up in the numbers, and at some point in a partners' meeting someone asks it out loud: what are we actually getting for this?
If you have been the one asking, this piece is written for you. And the most-quoted number in the whole debate is on your side.
In 2025, MIT's NANDA initiative reviewed more than 300 public enterprise AI deployments, interviewed 52 organizations and surveyed 153 leaders, and reported that 95 percent of generative AI pilots produced no measurable impact on profit and loss, against $30 to $40 billion of spending.
That headline deserves the same scrutiny you would give any vendor's claim. The report is industry research, not peer-reviewed; its full dataset has not been published; and "no measurable impact" is not the same finding as "failed." Critics have argued that much of what it records is spending with no baseline to measure against. Read that way, the report is less an obituary for AI than a description of how most of it was bought.
The more useful findings sit under the headline. Tools bought from specialized outside vendors succeeded about twice as often as systems built in-house. Budgets leaned heavily toward sales and marketing tools, while the report found better returns in narrower back-office work. The report's own explanation for most of the gap was organizational rather than technical: tools that did not fit or learn the workflow they were dropped into. The practical corollary is to measure a task's cost before the tool arrives, because without that no return can be proven either way.
Each of those is a capital-allocation decision, the kind a finance lead already knows how to make. None of them is a bet on the technology.
Companies with an official AI subscription, against companies whose workers reported using personal AI tools for work every day. The gap is not an adoption statistic. In an agency, it means client briefs and strategy are already going through tools nobody approved, logged or reviewed.
Your people adopted AI a year ago, whether or not the firm did. The open question for most agencies is not whether to use it but whether its use is governed. For a business that handles other companies' plans, launch dates and sometimes customer data, ungoverned use is an unpriced liability, and it stays unpriced until one client asks where their material went.
The clients are also moving. In Gartner's 2025 CMO Spend Survey of 402 marketing leaders, 39 percent said they planned to cut agency budgets, and 22 percent said generative AI had already reduced their reliance on external agencies for creativity and strategy. The work AI makes cheapest is the work clients can increasingly do without you.
| Question | The answer that tends to pay | Why |
|---|---|---|
| Bought or built? | Bought from a specialist | About twice the success rate of in-house builds in the MIT NANDA sample |
| Specialized or general? | Specialized, pointed at one workflow | The returns clustered in narrow, defined tasks |
| Measured or unmeasured? | Measured against a before-the-tool baseline | Without one, no return can be shown in either direction |
| Governed or ungoverned? | A written rule for client material | The exposure exists already, whether or not the firm bought anything |
A decision aid built from the MIT NANDA findings, not a scoring model. Each row is a yes-or-no call a finance lead can make this quarter.
Apply the four questions to how most agencies use AI and a pattern shows up: the spending sits in drafting, variations and first-pass audits, the layer whose price is falling fastest. The companion piece on this site, on why AI multiplies the process a firm already has, covers the other half of the problem.
The MIT NANDA figures are not peer-reviewed, and the dataset is unpublished. The initiative also studies agent infrastructure, so it has a point of view about what the answer looks like. Treat the 95 percent as direction, not measurement.
None of the samples are agencies specifically. The enterprise findings describe large companies; how exactly they carry to a thirty-person shop is a judgment.
Gartner's figures are from spring 2025 and describe intentions, not completed cuts.
"Bought beats built" is not a rule for every firm. A specialist vendor is still a dependency. A firm with a genuinely unusual process may have to build, and should do it knowing the base rate.
- Nobody owns the AI budget, and nobody measured the work before the tool — no owner and no baseline means no one can show it paid.
- All of it sits in production — drafting, variations, audits: the layer clients are learning to do themselves.
- People use personal AI on client work and there is no written rule — the exposure exists whether or not you bought anything.
- Time saved goes back to clients as fewer billable hours — the efficiency is being given away rather than kept as margin.
- You are about to build something a specialist already sells — check the base rate before you commit a developer's quarter to it.
Write a one-page rule for what client material may go into which AI tools, and send it to the whole firm this month.
List the tools the firm approves, what may be put into each (public information, internal drafts, client confidential material), who can approve an exception, and where outputs that use client material are stored. Ask the team, anonymously if needed, what they actually use today, and write the rule around reality rather than around the tools you wish they used.
- What it costs
- An afternoon to draft, an hour with whoever handles your contracts, and possibly a paid business subscription to replace the personal accounts people are using now. It also means saying no to some habits the team likes.
- How you'll know
- Within the month, every person has acknowledged the rule and the tool list matches what they report using. The real test is the first client who asks how you handle their material with AI: you answer in one sentence and send the page.
The skeptic was mostly right about the headline and mostly wrong about the conclusion. AI does not reliably pay. The firms it pays for bought narrowly, measured first, and knew where their clients' material was going.
Related from Sound Decisions: AI Doesn’t Give You Leverage. It Multiplies the Process You Already Have. · Creative Is Now a Line Item · You Don’t Need More Clients. You Need Longer Ones.
One conversation. One AI line in your budget.
Not a vendor recommendation. A read on whether that spend is bought or built, measured or not, and pointed at work your clients will keep paying for.
Get a free positioning review Or read the frameworks →