Count Your Agents. Andrew Boyd Will Wait.

An enterprise customer figured its employees were spinning up roughly 500 AI agents a week. A monitoring tool went looking and found 2,000 already built.

The sprawl is not the interesting part; the cost of the miscount is. In a recent Dataiku/Harris Poll, 85% of CIOs said traceability or explainability gaps have already delayed or stopped AI projects from moving into production. Agents nobody can account for don’t blow up. They queue, waiting for someone to certify what they do, while the board asks why nothing shipped.

Andrew Boyd tells the 500-to-2,000 story often. He also sells the instrument that counted. Boyd spent two decades in enterprise technology, including stints at Microsoft and AWS, before becoming Dataiku’s senior vice president and general manager for Asia Pacific and Japan in April 2026. In October 2026, his company will ship Agent Management, a control tower for tracking and governing agents wherever they were built, including on Bedrock, Salesforce and n8n.

The agent count is only the first number he thinks enterprises here have wrong. There are three more.

Nobody knows the count

The Dataiku/Harris Poll surveyed 600 CIOs between Dec. 11, 2025, and Jan. 7, 2026, all at companies above USD500 million in revenue, across markets including Australia, Japan, South Korea and Singapore. 87% said AI agents were already embedded in critical operations. Yet only 25% had full real-time visibility into every agent running in production. And 82% admitted employees create AI agents faster than IT can govern them.

So, how common is this gap between perceived and actual agent numbers in an enterprise? “Massively common,” Boyd says.

His objection isn’t that people are building agents. “There’s a tagline that says, count your agents, sit back, we’ll wait for you to do it. They’re just there. And you don’t really want to block that innovation.”

The real challenge comes later, when somebody asks how a decision was made. “Prior to agents, I could do that traceability through my workflows and my team,” Boyd says. “Now, that traceability end-to-end of where we arrived is so important.” An agent nobody logged is an answer nobody can source. In a regulated business, that is the difference between an agent that ships and one that sits in review. Multiply that by an estate nobody has counted and you get the 85% of CIOs who say traceability or explainability gaps have delayed or stopped projects.

For CDOs, the problem is not the volume of agents. It’s agents that duplicate each other with nothing measuring what they deliver. “You’ve got a lot of proliferation of agentic or AI use cases, many of which overlap with no consistent measurement of what they’re tying back to,” Boyd says. The second symptom is sequencing: trying to solve one business problem at a time, “kind of like whack-a-mole.”

What fixes the count is a discipline borrowed from another department. Is this agent long-running, or scoped to a project? How are you tracking its drift, the way you once tracked model drift? What are its V2, V3, and V4? “You’re almost seeing the best CDOs merging traits from product management and lifecycle into this kind of methodology,” Boyd says. “The AI teams and the CDOs are the product managers of AI within the business now.”

Get the count right, and the next question is what those agents return. Boyd argues that number was imported.

The ROI number is from somewhere else

Most of the world justifies AI on productivity. Scale human capital, take out cost, and book the dividend. Across much of Southeast Asia, that math is weak because labor is not the expensive input.

“In most markets in ASEAN, the return on investment of doing that is not as pronounced because of the nature of the market economics,” Boyd says.

So, rather than squeeze the supply side, Boyd observes that customers in Indonesia and elsewhere in the region point AI at the demand side: new products, faster cycles, and markets that were too costly to enter. One manufacturer he cites used to refresh its roadmap once a year, a cycle built around a handful of people reading market trends and arguing it out. Turn that expertise into agents and the constraint changes shape.

“Agents nobody can account for don’t blow up. They queue, waiting for someone to certify what they do, while the board asks why nothing shipped.” — Andrew Boyd @ Dataiku

For a CDO graded on cost reduction, that is an argument worth having upstairs. “You can cut costs; it’s a one-time dividend. But you can’t grow yourself out of something by cutting costs,” Boyd says.

Whether the argument lands depends on who is listening. Dataiku’s survey of 900 CEOs found 70% claim ownership of AI strategy, 60% sit in more than half of AI decisions, and 6% sit in nearly all of them. Ownership is high but attention is thin. The CDO making the growth argument is making it to an executive who has delegated most of the decisions and kept all of the accountability.

Which is roughly how the third number gets made.

The vendor bet expires

Three-quarters of those same CIOs regret a major AI vendor or platform decision taken in the past 18 months.

Put the lock-in question to Boyd and he answers as an operator, not a rep. “If I were a CIO making that decision, or a CDO making that, I would not be betting on a single provider.”

His evidence is churn. He says he has met companies that swapped their primary large language model provider two or three times inside six months, and that Chinese open-weight models overtook the field in a quarter. Sovereignty compounds it. Certain model providers, he says, have run into limitations outside certain parts of the world, which turns a technical choice into a business risk.

The prescription is architectural: run business-critical work on open-weight models you control, keep public providers for experimentation, preserve the ability to swap either out. It also describes his own product. Dataiku’s LLM Mesh exists to make models interchangeable, and Boyd is candid that this is why he took the job. The logic survives the pitch, and a CDO can point it back at Dataiku.

The same reasoning covers regulation. Korea is prescriptive, with an AI Basic Act that spells out obligations. Singapore, through the Monetary Authority of Singapore, is stable enough that you aren’t redesigning every quarter. Indonesia, Boyd says, is AI-first by government mandate, though its national AI roadmap is still being written into rules.

Most regional enterprises answer this patchwork by designing to the lowest common denominator, so one architecture works everywhere. Boyd calls that a stall. Attest to each regime separately and the differences become options, not obstacles.

Three wrong numbers so far. The fourth is the one the board actually reads.

What the board is actually reading

95% of those CIOs brief their boards on AI performance, roughly half at least monthly. A quarter of them can see every agent in production.

Read those two together. The 500-to-2,000 problem does not stay in the engineering org. It travels upstairs and becomes a monthly report on an estate nobody has finished counting, delivered with a confidence nobody in the room can check.

So what should a board do once it knows that? Boyd’s answer has two halves.

The first is obvious. No governance, no risk framework. “As a board, that is your primary objective,” he says. “If you can’t do that, you’re not meeting your responsibilities as a board.” That is the argument for counting.

The second is what CDOs rarely hear from a company selling governance. “The second thing is then to use that as a premise to stop doing things because you can’t do that. That is actually the second mistake as a board.”

That is the argument against flinching at what the count turns up. Finding 2,000 agents where you expected 500 is not proof a company has lost control. It is the first accurate number it has had.

Both failures end in the same place. One leaves you with agents you can’t see. The other leaves you with a company that stopped counting on purpose.

Image credit: iStockphoto/Christophe Bourloton