AI Agent Adoption Surveys Say 70% to 97%. Read the Sample Before the Headline

Creatio says 70% of organisations are past AI experimentation. WRITER says 97% of executives deployed AI agents in the past year. Placed side by side, those two AI agent adoption survey results look like a range running from 70% to 97%. They are not a range. They measure different things, in different people, with fieldwork months apart.

This post is written for the ops or delivery lead who gets asked why their team is behind a 97% figure. That reader needs a checklist to run on the next statistic, not a counter-statistic to wave back.

The short answer. The 70% and the 97% cannot be combined. Creatio's 70% counts 608 decision-makers whose organisations use AI of any kind beyond experimentation. WRITER's 97% counts 1,200 executives, screened to firms that permit generative AI, who say their company deployed agents in the past year.

What "deployed" means in each AI agent adoption survey

The word doing the most work in both headlines is the one neither survey defines publicly. "Deployed" can mean a feature switched on in one team's software, or an autonomous system running a business process end to end. A survey that does not say which one it counted cannot be compared with one that does.

Creatio counts AI in any function, not agents in production

Creatio's release of 20 October 2026 defines its headline precisely. "Seventy percent of respondents are beyond experimentation, having deployed AI in selected functions, begun scaling it across the organization, or embedded it enterprise-wide." The subject of that sentence is AI, not agents.

The report's landing page breaks the stages down. 14% are embedded across the organisation, 25% are scaling, and 31% have deployed in selected functions. Those three sum to the 70%. The remaining 30% split into experimenting (15%), still evaluating (13%) and no plans (2%).

So the single largest group inside the 70% is "deployed in selected functions". That is 31 of the 70 points, or 44% of everyone counted as past experimentation. A company with one AI tool in its marketing team qualifies for the headline on the same terms as one that runs AI across finance, service and sales.

The agent figures in the release are forward-looking. More than three-quarters of respondents "expect to scale AI agents across departments or use cases within the next 12 months". A contributed CIO.com article by a Creatio senior vice president puts that at 78% "already scaling AI agents across departments or use cases, or expect to within a year". That number merges doing with planning, so it measures intent as much as adoption.

WRITER counts any agent deployment in the past 12 months

WRITER's blog summary of its survey states that "Nearly all executives (97%) say their company deployed AI agents in the past year, with 52% of employees already using them." The unit is the company. The window is 12 months. The threshold is any deployment at all.

That 97% is also used for a different claim. WRITER's press release of 7 April 2026 says "Nearly all executives (97%) say AI has been beneficial." Two statements, one number, published the same day. They may come from separate questions that landed on the same result, but without the questionnaire a reader cannot check.

No WRITER page I reviewed defines "AI agent", which I think is the most important missing fact in either survey. An assistant feature that drafts email inside software a company already owns might count. So might a system that files purchase orders without a human. The headline treats them as one thing.

A stricter definition produces a different headline

For contrast, a third survey shows what happens when "deployed" gets split into grades. Monte Carlo surveyed 260 practitioners and engineering leaders at organisations with 1,000 or more employees. As reported by IT Brief on 1 May 2026, 46% said agents are in full production and 39% in limited production.

Add the two together and you get 85%, close to the other headlines. Keep them apart and the honest headline is 46%. The same survey found 64% put agents into use before they felt ready. "In production" and "ready for production" are separate claims.

The methodology table: sample, sponsor and wording

This is the comparison the headlines skip. Every row comes from the sponsor's public pages or the coverage cited. Every blank is a blank in the public record.

Creatio from its 20 October 2026 release and report landing page. WRITER from its 7 April 2026 press release, blog and report landing page. Monte Carlo from IT Brief coverage, 1 May 2026. Both full reports sit behind registration forms and were not reviewed for this post.

Read across the "question wording public" row first. It is "no" in every column, so none of these surveys can be checked by an outside reader. That is the norm for vendor research, not the exception.

Then read the "what the headline counts" row: AI in a function, agents in a company over 12 months, agents in production. Any article that quotes two of these numbers in one sentence is comparing different objects as if they were one.

Who answered each AI adoption survey, and who was filtered out

A survey result is a statement about the people who answered it. Change who answers and the result moves before a single question is asked.

Creatio: 608 decision-makers with no published breakdown

Creatio's report page says "The study surveyed 608 IT and business decision-makers." The release calls them "global". The public pages publish no split by role, country, company size or industry, and no fieldwork dates.

WRITER: executives screened for permission to use AI

WRITER is more transparent about its screen. The press release states that "executives were required to be working at a company that permits the use of these tools", and employees had to be using generative AI at work. Companies ranged from 100 to over 10,000 employees across nearly 30 industries.

A company that bans generative AI could not have supplied an executive respondent. So the 97% describes executives at companies that had already said yes to AI: the converted, not the market.

The size mix tilts small. The report page shows 42% of respondents at companies with 100 to 999 employees and 13% at companies above 10,000. At a 200-person company, one support team switching on an agent in its help desk software is a company-level deployment.

The geography is also unclear. The press release lists the US, UK, Ireland, Benelux, France and Germany. The report landing page lists the US, UK, France, Germany and Australia. The public material does not explain the difference.

What the sponsors sell, and why it matters less than you think

Both headline surveys were paid for by companies that sell AI agent software. That is the first fact readers reach for, and the least useful one.

Sponsorship does not make a number false. Workplace Intelligence ran WRITER's fieldwork, and buyers who know their own companies will test any published finding. A sponsor's real influence shows up somewhere else: the problem it measures is often the product it sells.

Look at what each survey chose to highlight beyond adoption. Creatio's release stresses that organisations with established governance are more than 4 times as likely to feel very confident scaling AI. A Creatio executive's CIO.com article gives the underlying figures as 71% against 17%. Creatio sells a governed platform.

WRITER's survey highlights that 36% of companies lack a formal plan for supervising agents. It adds that 35% of executives are not very confident they could "pull the plug" on a rogue agent, and that only 23% see significant ROI from agents. WRITER sells agent infrastructure built around control and oversight.

None of that is improper. It is the normal shape of vendor research: measure adoption, then measure the gap your product closes. My rule: treat the adoption headline as context and the gap finding as the sales argument, then check whether the sample could support either. For the same trap applied to return figures, see why NVIDIA and McKinsey report such different AI ROI numbers.

WRITER's 36% supervision gap, recomputed

The 36% figure is the most quoted governance number in WRITER's survey. Read with the 97%, it puts a floor under how many companies run agents with no formal oversight plan.

Assume both figures share one base, the 1,200 C-suite respondents. The report landing page labels the supervision figure "C-suite: 36% have no agent supervision plan", while the blog phrases it as "companies", so this assumption is reasonable but not confirmed.

97% say their company deployed agents, so 3% did not. Even if every one of that 3% is among those without a plan, at least 33% of all respondents (36 minus 3) deployed agents and have no formal supervision plan. At most it is 36%. That leaves between 61% and 64% who deployed agents and do have a plan.

As a share of deployers only, the no-plan group is between 34% (33 of 97) and 37% (36 of 97). Roughly 1 in 3 executives who say their company deployed agents also say there is no formal plan to supervise them.

The pages disagree with each other

The control figures do not line up across WRITER's own pages either. The blog and press release say 35% of executives are not very confident they could "pull the plug" on a rogue agent. The report landing page shows "C-suite 23% can't stop a rogue agent". Those could be two different answer cuts of one question, such as "not very confident" against "not at all able". The public pages do not say.

I would not quote either control figure without the full questionnaire. The supervision gap above survives that doubt because it uses two numbers that appear consistently. For what a supervision plan should actually contain, the breakdown of human-in-the-loop architecture covers the design choices, and the agent governance maturity model covers how to grade your own.

Why the 70% and the 97% cannot be combined

The tempting move is to report a range, average the two, or treat the 97% as an update on the 70%. Each is wrong, for reasons that stack.

First, the unit differs. Creatio's figure is about AI in general and WRITER's is about agents. Second, the respondents differ: IT and business decision-makers against C-suite executives screened for AI permission. Third, the measure differs: a six-stage maturity scale against a yes or no on any deployment in 12 months. An average of 70% and 97% is 83.5%, and that number describes no population and no question that anyone asked.

WRITER's fieldwork closed on 25 January 2026. Creatio published on 20 October 2026, 268 days later, and has not published its own fieldwork window. Reading the 97% as newer or the 70% as older gets the order backwards: the 97% came from earlier fieldwork, and the October report is the more recent of the two by publication.

In a category moving this fast, 9 months is long enough for both numbers to have moved. That is a reason to date every figure you quote, not to assume either has risen.

Neither survey publishes a margin of error. If you treated each as a simple random sample, the 95% margin on Creatio's 70% with 608 respondents would be about plus or minus 3.6 points. On WRITER's 97% with 1,200 executives, about plus or minus 1.0 point.

Both samples are recruited panels, not random draws, so the formula does not strictly apply. Differences in definition and screening move the result by tens of points. A 3-point sampling error is noise next to the gap between "AI in one function" and "agents in production".

This is the same failure that makes the AI use-case count a vanity metric: a loose unit of counting flatters the total. It also explains why adoption headlines and the ledger of cancelled agentic AI projects can both be accurate at once. One counts starts. The other counts what survived.

Where this argument is weakest

This post criticises two surveys for missing methodology, so it should pass the same test. It falls short in three places.

The full reports are gated and were not read

Both full reports sit behind registration forms. The full reports may well publish the question wording, agent definition and fieldwork dates that the public pages omit. If they do, some "not published" cells above become "published but buried", a weaker criticism. The table describes what a reader sees without handing over contact details, which is as far as most people quoting these numbers get.

The direction is probably right

Strip away the loose definitions and the evidence still points one way. Even Monte Carlo's strict reading finds 46% of large-company practitioners with agents in full production. My judgement is that agent adoption in large firms is genuinely broad by late 2026. The honest headline is "widespread and shallow", not "rare".

So this is not an argument that adoption is low. It is an argument that a 97% figure and a 70% figure tell you almost nothing about depth, and depth is what decides whether your team is behind.

Executives may undercount, not overcount

Self-reported adoption is usually assumed to run high. WRITER's own data cuts the other way on one point: 67% of executives believe their company has already suffered a data leak because an employee used an unapproved AI tool. If usage runs ahead of what leadership knows about, an executive survey could understate how many agents are actually running. That does not rescue the 97%, but the error is not all in one direction.

A checklist for reading any AI adoption survey

Run these seven questions on the next AI adoption statistic that reaches your inbox. Each can be answered in minutes from the public pages or not at all, and "not at all" is itself an answer.

The second row is worth one more sentence. The 56-point gap is Creatio's 70% headline minus its 14% embedded share, which shows how far a single survey can move on the bar alone. If a vendor gives you an adoption number without the bar, ask for the stage split.

For the same discipline applied to pilot outcomes, the sample behind MIT NANDA's pilot failure figure runs the identical test on the year's most quoted failure rate.

Frequently asked questions

Where to start

Take the last AI adoption figure someone used to tell your team it was behind. Find its source page and fill in the seven-row table above for it, leaving blanks where the source is silent. This takes about 20 minutes, and the count of blanks is the most useful thing you will learn about that number.

Then answer the same seven questions about your own team, which you can do completely. Write down which grade of use you are actually at, using Creatio's six stages as the scale, and whether you have a written supervision plan for every agent in production. That comparison you can defend, because you know both the definition and the sample.

Related reading