A2A Research · Agencies · July 2026

Where AI Actually Pays in an Agency

Most enterprise AI returns nothing — and the finance chief who refused to chase the hype protected the firm. But the same evidence shows exactly where the payback lives, and it's a capital-allocation question, not a leap of faith. A read for the skeptic who signs the checks.

Reader · The agency CFO Evidence · MIT · Gartner · HBR · Deloitte Read · 9 minutes + a 5-minute diagnostic

01 · The skeptic was right

Most of what you've heard about AI is true — if you heard the skeptical version

Start with the number a finance chief asks for first.

In 2025, MIT's NANDA initiative studied more than 300 enterprise AI deployments and found that 95% of them produced no measurable impact on profit and loss. Not modest returns — no measurable return at all, against thirty to forty billion dollars spent. If you've been the person in your agency asking where the payback is, quietly declining to bet the firm on the technology everyone else was celebrating, the data is on your side. Not on the side of the enthusiasts.

That deserves saying plainly, because it rarely gets said: the skeptic was right. Every tech wave the agency has lived through arrived with the same promise and mostly delivered the same thing — a little more speed, rarely more growth. Treating this one with the same suspicion wasn't timidity. It was arithmetic, and it protected the firm.

But "AI doesn't pay" and "our AI didn't pay" are different sentences. The 5% who got a real return weren't luckier or braver. They did three specific things — and every one is a capital-allocation decision a CFO is already equipped to make.

02 · The 5% left a map

The failures weren't random — they read like a risk checklist

Most of that spend produced nothing. Three things separated the returns from the write-offs.

95%
Of enterprise AI deployments produced no measurable impact on profit and loss
MIT NANDA · 2025
Tools bought from specialized outside vendors succeeded ~67% of the time — about twice the rate of systems built in-house
MIT NANDA · 2025
>50%
Of AI budgets went to sales-and-marketing tools — among the lowest-returning uses; the payback sat in narrow, defined workflows
MIT NANDA · 2025
1
Thing the winners all did: measured the task cost before the tool, so they could prove it paid
MIT NANDA · 2025

That first finding deserves a hard look from any agency that has said "we could just build this ourselves." The instinct is understandable — it feels cheaper and more controllable. The data says it's the more likely way to join the 95%. Building your own isn't the safe choice; it's the one that fails twice as often.

So the CFO's real question isn't should we use AI. It's narrower and answerable: is our AI bought or built, specialized or general, measured or unmeasured? Three yes/no questions, each a capital-allocation call.

03 · Follow the dollars

Agencies are spending in the wrong layer

Apply those questions to how agencies actually use AI today, and a pattern appears that should concern whoever watches the money.

A Q1 2026 survey of 250 independent agencies found about 41% had at least one AI agent in production — concentrated almost entirely in execution: roughly two-thirds in brief and content generation, half in SEO audits. The bottom quartile wasn't covering its token bill, and the single most-cited blocker was the inability to prove the output beat the manual baseline.

Read that against your own economics. Agencies have poured their AI into the production layer — the drafts, the variations, the first-pass audits. But that's the layer whose price is falling: the efficiency paradox from our first Houston report, where billing for effort collapses precisely as AI makes the effort cheap.

You are automating the work you can no longer charge much for — and calling it progress.

04 · The unpriced liability

The risk is already inside the building

Before we get to where the money should go, one liability a finance chief will want named — because it's already on the books, unpriced.

Only about 40% of firms had official AI subscriptions — while roughly 90% of workers were using personal AI tools for work every day.

MIT NANDA · State of AI in Business 2025

That gap isn't an adoption statistic. It's a governance exposure. On an agency's work it means client material — briefs, strategy, sometimes confidential data — is being run through consumer chatbots that no one approved, logged, or reviewed, at the very moment clients are growing more anxious about exactly that.

The question was never adopt or don't. Your people adopted a year ago. The only open question is governed or ungoverned — and for most agencies, the honest answer is ungoverned. That's not a technology decision. It's an unpriced risk on the balance sheet, waiting for one client to ask the wrong question.

05 · Where the dollar pays

Where AI actually defends revenue

So where does an AI dollar earn its return in an agency? Not where the industry is spending it.

The clients aren't waiting, and Gartner's CMO data says so in three directions at once. Meanwhile the broader research — HBR, Deloitte — keeps finding the same thing: AI reliably lifts productivity, but rarely growth. It makes the cheap work cheaper. What it doesn't do is the scarce thing: judgment, interpretation, knowing which client to pursue and which relationship is quietly at risk.

39%
Of marketing chiefs plan to cut agency budgets
Gartner CMO Spend Survey
22%
Say generative AI has already reduced their reliance on agencies for creativity and strategy
Gartner CMO Spend Survey
50%
Of agencies’ own proprietary AI platforms are expected to be obsolete by 2029
Gartner · forecast
41%
Of independent agencies run at least one AI agent in production — almost all of it in the execution layer
Digital Applied · Q1 2026 · n=250
The production layer is where AI cuts cost — a race your clients can run without you. The decision layer — which clients you win, which you keep, which review you see coming — is where revenue is defended, and it's nearly empty of AI today. One kept client, at agency margins, dwarfs a year of shaved production minutes.

The CFO's conclusion writes itself: move the AI dollar from the layer that erodes your price to the layer that defends your revenue.

06 · The five signs — a five-minute mirror

Is your AI spend in the right layer? Score it honestly.

A capital-allocation check on your firm's AI dollars. Nothing you tap here is stored or sent anywhere — it runs entirely on your screen.

S1
No single person owns your AI spend, and no one measured what the work cost before the tool.No owner, no baseline — so no one can prove it pays.
S2
Your AI is all in production — drafting, variations, audits — the layer whose price is falling.Every dollar in the depreciating layer.
S3
Your team uses personal AI on client work, and there's no written rule for what may go into which tool.Shadow AI, unpriced and ungoverned.
S4
AI's time savings go to clients as fewer billable hours — not captured as margin.You're giving the efficiency away.
S5
No AI touches pipeline, reviews, or relationship health — the layer that actually keeps clients.Nothing where revenue is decided.
/ 5
Answer all five to see your read

Five honest answers. That's all the mirror needs.


07 · What the 5% do

Four moves, startable Monday — none a big bet

Name an owner and a baseline

One person accountable for AI spend, and a before-the-tool cost for every workflow — so return is provable, not asserted.

Move one dollar to the decision layer

Take a single AI budget line out of production and put it where clients are won or kept.

Write the one-page shadow-AI policy

What may and may not go into which tools on client work. Price the risk before a client does.

Measure before you scale

No AI workflow expands until it beats its manual baseline in a blind check.

That's the full prescription, and it's yours regardless of what you do next. Act on those four and never contact us — the report did its job.

The honest problem with that advice

Every move points at the same scarce resource — and it isn't the tool. It's senior judgment: someone who knows which decisions matter enough to defend, and has the standing to govern the rest. AI doesn't supply that. It only pays off when it's aimed by it. The bottleneck was never the technology. It's the small number of people who can tell a decision that matters from one that doesn't.

Why we built A2A

A2A — Aligned to Act — is built to the three conditions the 5% share: it's bought, not built (you don't maintain it); specialized, not general (it does one thing — win and keep profitable clients — not everything); and measurable (its job is decisions with a visible return, from the first room). It's AI aimed at the decision layer, where the return is structural — instead of the production layer, where the price is falling.

The refusal, plainly: A2A won't make a bad AI bet good, won't fix work that's behind or priced wrong, and won't replace the judgment it depends on. It exists for one layer — the decisions that win and keep clients — and it's honest about the rest.

One useful next step — and it isn't a sales call

Where does your AI dollar actually pay?

If the five signs landed, bring one real allocation question and let's talk about moving the dollar to the layer that defends your revenue.

Request a conversation