You Cannot Own What You Cannot Count

For three weeks it feels like I have been writing the same sentence over and over. Name an owner. A named person who verifies that the enforced boundary matches the intended one. A named person who owns the response and can act at machine speed. I still think that is the right answer, and I am not hedging on any of it.

Then a company did it and went further than I asked. What happened next is the part I never wrote about, because I did not think anyone would actually try it.

Kantar is a market research firm of more than ten thousand people. Somewhere between last November and this March it changed job titles. Andy Doyle, Chief People Officer since 2022, became Chief People and Agent Officer. On March 26 this year, Narelle Burke announced that she had moved into the same title for the Global Research Group and the Asia Pacific region, which two trade outlets picked up the same day, one of them working from her own post and neither from a company announcement. The confirmation came from Kantar itself: by the Q1 lender call on May 28, chief executive Paul Zwillenberg was using the title the way you use one that has been long established, “Andy Doyle, our chief people and agent officer.”

That is not a steering committee. That is not a policy in a drawer. That is agents written into the C-suite, in the title, next to people. I do not know another company of that size that has gone that far, and I want that on the record before anything else in this piece. On the thing I have spent the month asking for, Kantar is ahead of nearly everyone.

Now the counting.

On March 25, Kantar published its FY 2025 presentation for lenders. On the AI slide: “2,500+ agents built, with focus on 50 ‘Hero’ agents to be deployed in workflows, at scale.” The chief executive said the same thing out loud on the call the next day. “We’ve built over 2,500 agents internally and are focusing on scaling 50 Hero agents across the business.”

Three months later, on June 24, Andy Doyle sat down for Microsoft’s WorkLab podcast. He said this: “we’ve got 15,000 odd agents across the organization today. Some of those agents are just very individual for someone. Some of those, about 150 of them, are big enterprise agents that we’ve built to do big corporate tasks.” He is not talking about licenses, and he says so in the same breath: “we have more agents than people in the organization.”

So the number of agents is 2,500, or 15,000, or 150, depending on which sentence you read.

Every one of those can be true at once. It is a finding, not an objection. Two and a half thousand in March and fifteen thousand in June is a plausible curve for a company that handed a license to every employee and told them to go play. The one hundred and fifty is tracking a different metric altogether, the ones built well enough to be pointed at the whole company. I do not read any of this as Kantar contradicting itself. I read it as Kantar doing what this entire market does, which is use one word for three different things, because nobody has published a definition of the unit.

Here is the part to be careful about when reading. During that same call on the 26th, Kantar’s chief financial officer, Michael Uzielli, described “the rollout of Copilot to over 15,000 people.” I am not saying the fifteen thousand people and the fifteen thousand agents are the same fifteen thousand. I do not know that, and I have no basis to say it. That is exactly the thing I want you to work through. Two numbers landing on the same figure, from one company, three months apart, one counting humans and one counting software, and nothing published anywhere that lets a reader, a lender or an auditor tell them apart from the outside.

Now look at what the record does not contain, which is the more interesting piece.

Nothing public, from Kantar or from Microsoft, describes what happens to an agent nobody uses. Doyle is not asked about it. There is no retirement process anywhere in the public record, no review cycle, no expiry, no published count of how many of the fifteen thousand ran this week. Kantar may well run all of that internally, and I would not expect a podcast to be where it announced if it did. The point is narrower and it is about what a person outside the company can establish, which, in this case, is nothing.

Two smaller things take the same shape. Doyle refers to a head of the agent factory, who he says came from the business rather than from technology, and does not name her, which is an ordinary thing not to do on a podcast and is only worth mentioning because it means there is no publicly identifiable operator of the factory itself. And Kantar’s published AI Explainability Statement, which is its formal public document on the technology, does not contain the word “agent” once. It is a document about Copilot rather than about agents, it carries no version date, and it has simply not been extended. I bring it up because it is the one place a customer would go looking.

He is candid about the state of the factory, and that is worth mentioning. There is “a lot of duplication,” he says, “a lot of repetition,” and “a lot of very diverse quality in those agents.” He is describing the problem accurately and out loud, on a podcast, with his name on it. What is not anywhere on record is what happens to any of them next.

None of that is a Kantar failure. It is what the category looks like right now.

Gartner, on April 28, put the number of agents in the average global Fortune 500 enterprise by 2028 at over 150,000, “up from less than 15 in 2025.” It is published as a growth forecast. Read it as a definitional one as well. Either agents are appearing at a rate nothing in the history of enterprise software has matched, or the word is being applied to a good deal more than it was a year ago, and my guess is that it’s both. “Agent” has become more than a technical term; it has become marketing buzz. Either way, the count is not measuring a stable object. The same release says only 13 percent of organizations believe they have the right AI agent governance in place.

Microsoft counts what it calls monthly active agents and they define them very precisely: an agent counts if it had at least one day of user-initiated usage in a rolling 28 days or completed one autonomous run. Kantar’s deck counts agents built. Those numbers are not answering the same question. One is an inventory of what exists, the other is a census of what moved. A company can answer both honestly and hand you two numbers that will never reconcile.

You cannot own what you cannot count.

I want to show you what that looks like on the other side of the arc, in a case that is finished, litigated and no longer in dispute, so nothing here rests on predicting anything.

In late January 2012, JPMorgan’s Chief Investment Office put a new value-at-risk model into production. In the words of the Senate subcommittee’s staff report, it “cut in half the SCP’s purported risk profile.” It had been built by a credentialed quantitative specialist who, by the bank’s own account, “had not previously developed or implemented a VaR model, and was also not provided sufficient support, which he had requested.” He asked the trader he reported to for resources “on a number of occasions,” and “did not receive any.”

The bank’s own Task Force report describes how the thing actually ran: “the model operated through a series of Excel spreadsheets, which had to be completed manually, by a process of copying and pasting data from one spreadsheet to another.” In one of those spreadsheets a calculation “divided by their sum instead of their average, as the modeler had intended,” an error the report says “likely had the effect of muting volatility by a factor of two,” while hedging twice, in the same sentence, that nobody could say by exactly how much.

Here is the part that matters for this article.

None of that was discovered afterward. The bank’s own independent model reviewers wrote it down before the model was approved. Their review, an exhibit in the Senate record, reads: “The VaR computation is currently done off spreadsheets using a manual process. Thus it is error prone, and not easily scalable.” They attached an action plan requiring the calculation be automated. And at the end of the action plan, printed on the page, is this:

“Item Owner: Patrick Hagan; Completion Date: TBD”

A named owner, and a TBD. On the same exhibit a colleague writes to him: “Pat, please let me know if you agree with the formulation of the action plans, and what are your suggested completion dates.” The dates never arrived. The automation was never done. The Task Force’s explanation of why runs to one sentence: risk management “viewed themselves more as consumers of the model than as responsible in part for its development and operation.”

I am going to be very careful here with the man whose name is on that line, because he is a real person and the record is kinder to him than the line looks. The bank’s own report says he had never built a model of this kind, that he asked repeatedly for help and was refused, that he was carrying his existing job at the same time, and that the review process was never designed to check whether anything on the action plan actually got done. The Task Force apportions the failure to resourcing, to the review process and to risk management. It does not apportion it to him, and neither do I. That is rather the point. Putting his name on the line was the only part of the arrangement that worked, and it changed nothing.

On May 10, 2012, the bank withdrew the model and reinstated the old one. The losses reached at least $6.2 billion.

Kantar is not JPMorgan and to compare the two is not completely fair. Nothing in the Kantar record suggests risk-taking of any kind, and a research firm running HR agents is not a bank running a credit derivatives book. The parallel is narrower than that, and more useful because of it. In both cases a capable person outside the technology function built something the organization came to lean on. In both cases the risk was visible and written down at the time. In both cases a name was attached. And in the one whose ending we can already read, the name was attached to a line with no date, no inventory and nobody checking.

A named owner with a TBD is a name, not an owner.

So, what does the sentence I have been writing all month actually require, once you take it seriously?

There are three additions to every “name”.

A definition. Decide what one agent is in your company, write it down, and use the same word for the same object in the board deck, the lender call, the risk register and the podcast. If you cannot define it, you are not counting, you are estimating.

An inventory. Not how many were built. How many exist, how many ran this month, who made each one, what it can reach, and what becomes of it when the person who made it leaves. The gap between built and active is itself a number worth knowing, and almost nobody publishes it.

A date. Every open item on your agent estate carries a completion date or it carries a TBD, and the record is now unambiguous about what a TBD is worth.

Earlier this month I wrote that the answer was a named owner who verifies that the enforced boundary matches the intended one. A couple of days later I wrote that the answer was a named person who owns the response. Both are still correct. Neither one said what that person has to be handed on the first morning. Neither one said how that person is supposed to know what they are responsible for. I asked for a name and never asked for a number.

Kantar is not a cautionary tale here. Kantar is the company that got the first part right. They simply got ahead of themselves. Doyle says as much himself, that the ways “people can track and monitor and manage agents” did not really exist six months ago, and that managing the governance before those products arrived “was really tough.”

The cautionary tale is everyone who reads about a Chief People and Agent Officer, concludes that naming someone is the whole job, and stops. Naming an owner is the easy half. It is also the half you get to finish in an afternoon and have a press release.

The other half is arithmetic, and nobody gives simple math a microphone. Count them. Define them. Date them. Until that is done, the title is the only thing anyone owns.

SOURCES

Kantar Group, FY 2025 lender presentation, 25 March 2026 (the “2,500+ agents built” slide). https://www.kantar.com/investor-relations/financial-announcement/kantar-groups-fy-2025-lender-presentation

Kantar Group, FY 2025 lender presentation, direct PDF. https://www.kantar.com/-/media/8E19905351894200AA50B06824247DEE.ashx

Kantar Group, FY 2025 lender call transcript, 26 March 2026 (Zwillenberg on 2,500 agents; Uzielli on Copilot to over 15,000 people). https://www.kantar.com/-/media/A35687D8B9AB4475A8F600F4C7A7F469.ashx

Kantar Group, Q1 2026 lender call transcript, 28 May 2026 (the chief executive uses the title “chief people and agent officer”). https://www.kantar.com/-/media/68DDC88DA2D8416C8B460DA6BEAA4B0F.ashx

Microsoft WorkLab, “Andy Doyle: Let 15,000 agents bloom,” 24 June 2026. https://www.microsoft.com/en-us/worklab/podcast/andy-doyle-let-15000-agents-bloom

Microsoft customer story, Kantar and Copilot Studio, 18 November 2025 (Doyle still titled Chief People Officer). https://www.microsoft.com/en/customers/story/25674-kantar-microsoft-copilot-studio

Kantar, AI Explainability Statement for Copilot in Microsoft 365 (contains no instance of the word “agent”). https://www.kantar.com/AI-explainability-Copilot

People Matters, Kantar elevates Narelle Burke to Chief People and Agent Officer, 26 March 2026. https://sea.peoplemattersglobal.com/article/appointments/kantar-elevates-narelle-burke-to-chief-people-and-agent-officer-for-global-research-group-and-apac-48972

MrWeb Daily Research News Online no. 39611, 26 March 2026, corroborating the same appointment. https://www.mrweb.com/drno/news39611.htm

Gartner, “Gartner Identifies Six Steps to Manage AI Agent Sprawl,” 28 April 2026. https://www.gartner.com/en/newsroom/press-releases/2026-04-28-gartner-identifies-six-steps-to-manage-artificial-intelligence-agent-sprawl

Microsoft 2026 Work Trend Index, 5 May 2026 (the active-agent definition). https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization

Report of JPMorgan Chase & Co. Management Task Force Regarding 2012 CIO Losses, 16 January 2013. https://ypfsresourcelibrary.blob.core.windows.net/fcic/YPFS/JPMorgan%20Management%20Task%20Force%20Regarding%202012%20CIO%20Losses%201-16-13.pdf

U.S. Senate Permanent Subcommittee on Investigations, JPMorgan Chase Whale Trades, staff report, March 2013. https://www.hsgac.senate.gov/wp-content/uploads/imo/media/doc/REPORT%20-%20JPMorgan%20Chase%20Whale%20Trades%20(4-12-13).pdf

S. Hrg. 113-96, JPMorgan Chase Whale Trades hearing record, Volume 1 (Exhibit 44 at printed page 774). https://www.govinfo.gov/content/pkg/CHRG-113shrg80222/pdf/CHRG-113shrg80222.pdf