Counted, not measured

Dataiku and The Harris Poll asked 685 CIOs across eight countries how their AI agents are actually doing. Two answers from the same people, in the same survey, sit right next to each other. 90% say they have complete tracking of all their agents. And 72% say they can't consistently confirm whether those agents deliver the business outcome they were built for.

Read that twice. Nine in ten know where their agents are. Seven in ten can't tell you whether they work.

A count is not a measurement

Tracking an agent means you know it exists, who built it, what it touches. Useful. But it answers the question "how many do we have," not "is the work better than it was." The survey says 67% of these CIOs estimate 51 or more agents running in production. That's a lot of things running with no scoreboard.

And the count itself is softer than it sounds. 81% of the same CIOs say they lack complete oversight of agents created outside approved systems. 84% agree employees build agents faster than IT can govern them. So the 90% confidence is really confidence in the part of the fleet they can see. Only 21% report full visibility into what AI costs, broken down by team or use case.

Twenty agents in the bin

Here's the number I'd sit with longest: 47% have already decommissioned more than 20 agents this year. That can be healthy — pruning is part of running anything. But ask the obvious question. If 72% can't confirm what an agent delivers, how did they decide which twenty to switch off? Not on results. On cost, on noise, on whoever complained loudest.

That's the pattern I see in every operation that bought before it mapped. Build fast, deploy wide, count the fleet, and then argue about value with no "before" to compare against. The model isn't the weak point. The missing baseline is.

The pressure is real, too. 72% say their AI budget is likely to be cut or frozen if targets aren't met by the end of 2026, and 76% believe their own role is at risk if the gains don't show up by the end of 2027. Those CIOs will be asked to prove a return. Most of them will have an inventory to show instead.

Write the "before" first

In transportation and logistics, nobody would add a truck to the fleet without knowing what a lane pays and what it costs to run it. You'd laugh at a carrier that could list every unit it owns but couldn't say which ones make money. That is exactly where most AI fleets sit right now.

When we took on escalations, we wrote the number down before anything changed: 7 to 10 a day landing on one desk. Six weeks later it was 3 to 5, and we knew it because the "before" existed. No baseline, no claim — that's the rule.

So before the next agent goes live, one line per agent: the job it does, the number that job produces today, who reads that number, and the date you'll check it again. If nobody can fill in the second column, you're not ready to build. You're ready to measure.

Nine CIOs in ten can count their agents. Seven in ten can't say if they work. Tracking tells you what's running — only a baseline tells you whether it should be.

Sources: Dataiku, "81% of Global CIOs Say They Have Lost Oversight of Their Own AI Agents" (press release, September 24, 2026; The Harris Poll for Dataiku, survey of 685 CIOs in the US, UK, France, Germany, UAE, Japan, South Korea and Singapore, July 9–29, 2026) · The Next Web, "81% of CIOs lack full oversight of AI agents built outside IT, Dataiku says" (September 24, 2026)