WRITER's 2026 AI Adoption in the Enterprise survey, run with Workplace Intelligence across 2,400 respondents, contains two numbers that should be read side by side. 97% of executives deployed AI agents in the past year. Only 23% report significant ROI from them.
That gap is the entire commercial opportunity in enterprise AI ROI right now, and almost nobody is pricing against it correctly.
The rest of the picture is worse than the headline
- 79% of organisations report AI adoption challenges — up year over year, not down.
- 54% of C-suite executives say AI adoption is "tearing their company apart".
- 75% admit their AI strategy is "more for show" than actual guidance.
- 48% describe AI adoption as a "massive disappointment".
- 59% of companies invest over $1 million a year in AI; only 29% see significant ROI from generative AI.
- 36% have no formal plan for supervising AI agents at all.
Read that as a buyer, not as a commentator. Over half of large companies are spending seven figures a year on something they cannot demonstrate a return on, and they know it. They are not looking for someone to tell them AI is important. They are looking for someone to make the number go green.
Why most offers miss the enterprise AI ROI problem
The typical pitch is capability-shaped: we build agents, we fine-tune models, we integrate LLMs. Every one of those describes an input. The survey says buyers are drowning in inputs and starving for outcomes. A proposal that opens with an architecture diagram is answering a question the buyer stopped asking in 2025.
The offers that convert now tend to share three traits: they name a specific process, they carry a measurable before-and-after, and they take responsibility for the measurement itself. "We will cut ticket handling time by 30% in 90 days, measured this way, and here is what happens if we miss" is a fundamentally different sale from "we build AI agents".
The supervision gap is a product in itself
Two figures deserve their own line item. 67% of executives believe their company has suffered a breach from unapproved AI tools. 36% have no formal plan for supervising AI agents. Meanwhile 69% of companies are planning AI-related layoffs and 60% plan to cut non-adopters.
So: agents are being deployed at scale, into companies with no supervision plan, while headcount that understood the underlying process is being removed. That is not a stable configuration. Governance, evaluation and monitoring of agent behaviour is going to be a mandatory purchase within a year, and right now almost nobody is selling it as a standalone service with a price on it.
How to position against the gap
- Lead with the failure statistic, not the capability. Buyers already believe AI works; they doubt that it works for them.
- Sell a measured outcome on one process. Narrow scope, hard number, short cycle.
- Price on value, not hours. A workflow that saves $40,000 a quarter supports a retainer that has nothing to do with your time.
- Build the evaluation layer. Whoever owns the measurement owns the renewal.
If you are choosing a partner rather than becoming one, browse vetted providers in our AI agency directory, compare specialists by discipline in our categories, or get matched with an agency that has delivered this before.
The survey in one table
| Measure | Figure |
|---|---|
| Executives who deployed AI agents in the past year | 97% |
| Reporting significant ROI from AI agents | 23% |
| Reporting significant ROI from generative AI | 29% |
| Investing over $1M annually in AI | 59% |
| Facing adoption challenges | 79% |
| Calling adoption a "massive disappointment" | 48% |
| With no formal plan for supervising agents | 36% |
| Believing they suffered a breach via unapproved AI tools | 67% |
Survey data deserves a caveat: respondents self-report, "significant ROI" is undefined, and executives who bought something expensive have mixed incentives when asked whether it worked. But the internal consistency is what makes this credible. A dataset where deployment is high and satisfaction is inflated would be suspicious. One where deployment is high and satisfaction is candidly low is people telling on themselves, which is usually the truthful direction.
Three offers that are working right now
The single-process guarantee. One named workflow — invoice matching, ticket triage, lead qualification, contract review. A baseline measured before anything is built. A target with a date. A fee that steps down if the target is missed. This converts because it is falsifiable, and after eighteen months of unfalsifiable AI promises, falsifiable is a differentiator.
The evaluation retainer. The client already built agents; 36% of them have no supervision plan. You define the test set, run it on a schedule, catch regressions and report on behaviour drift. Low delivery cost, high renewal rate, and it makes you the party who decides whether the thing works — a structurally excellent position to occupy.
The consolidation audit. With 67% suspecting breaches from unapproved tools, most enterprises have a shadow AI estate nobody has mapped. Inventory it, classify by risk and data exposure, recommend a sanctioned stack. It sells on security budget rather than innovation budget, which matters because security budget survives a downturn and innovation budget does not.
How to price against an ROI gap
The pricing mistake is anchoring on your cost rather than on enterprise AI ROI. If a workflow saves a client $40,000 a quarter, the question is not what it costs you to run — it is what share of the saving is defensible. A third is usually accepted without argument when the measurement is credible and shared. That is roughly $4,400 a month for something that might cost a few hundred dollars in compute, and the client will renew it happily because the arithmetic is on the invoice.
This only works if you own the measurement. Whoever produces the number that proves the value is the vendor who cannot be swapped out at renewal. Whoever merely builds the thing is a supplier, and suppliers get repriced.
The risk nobody is pricing
69% of companies are planning AI-related layoffs and 60% plan to cut non-adopters, while a third have no agent supervision plan. Removing the people who understood a process, then handing that process to an unmonitored agent, is how quiet, expensive failures happen — the kind discovered two quarters later in a reconciliation. If you sell into these organisations, that risk is your opening. If you work inside one, it is your warning.
Frequently asked questions about enterprise AI ROI
Why is enterprise AI ROI so hard to demonstrate? Because most deployments never established a baseline. Without a measured before, there is no defensible after, and the conversation collapses into anecdote. Establishing the baseline before anything is built is the single highest-leverage step in the whole engagement.
Is a 23% success rate evidence the technology does not work? No. It is evidence that implementation is unsolved. The same survey shows super-users operating at roughly five times the productivity of laggards inside the very same companies, which is not what a broken technology looks like.
What is the fastest path to a defensible number? Pick one process, measure it for two weeks, change one thing, measure again. Narrow and honest beats broad and hopeful, and it produces a case study you can sell against.
Who should own the measurement? Whoever owns it owns the renewal. If the client measures your work themselves, you are a supplier; if you produce the shared number, you are the partner.
The bottom line
Near-universal deployment with a 23% success rate is not a story about AI being overhyped; it is a story about enterprise AI ROI going unmeasured. It is a story about implementation being unsolved, and unsolved implementation is what services businesses are for. The firms that grow through 2027 will be the ones that stopped selling the technology and started selling the result. Primary sources: the WRITER 2026 enterprise survey and ongoing enterprise research from McKinsey.