In February 2026, Anthropic published a study of how AI agents are actually being used in the real world, drawn from millions of human-agent interactions across its API and its coding tool. One chart from that research — which drew attention this week when Andreas Horn shared it on LinkedIn — maps where agents are being deployed, by domain.
There are sixteen named categories on it: software engineering, back-office automation, marketing and copywriting, sales and CRM, finance and accounting, data analysis, academic research, cybersecurity, customer service, gaming, document and presentation creation, education, e-commerce, healthcare, legal, and travel and logistics.
Procurement is not one of them.
Let me be precise, because precision matters here. The data is measured at the level of individual tool calls, so it is possible that some procurement-related activity sits inside “back-office automation,” or inside the unlabeled “other” category. What the data shows is not that procurement activity is literally zero. It is that procurement does not appear as a domain of its own — in a study built specifically to map where agents are working.
That absence is worth sitting with. Not as a grievance. As a question: why?
The domains agents reached first have something in common
Figure: “In what domains are agents deployed?” Software engineering accounts for 49.7% of tool calls; procurement does not appear among the named domains. Source: Anthropic, “Measuring AI agent autonomy in practice” (February 2026).
Software engineering alone accounts for nearly half of all agentic activity — 49.7% of the tool calls in Anthropic’s sample. That is not an accident, and Anthropic’s own explanation is the important part. Software, the report notes, is “comparatively easy to test and review.” You can run the code and see whether it works. The compiler returns an instant, unambiguous verdict: it runs or it doesn’t; the test passes or it fails.
That verdict is what makes software hospitable to agents. Every action an agent takes can be checked — immediately, and almost for free — by the machine itself. The domains where agents arrived first are, broadly, the domains where verification is cheap. The economics of verification matter as much as the capability of the model: AI moves fastest where the cost of proving an answer is low.
Anthropic makes the corollary explicit: in domains where confirming an agent’s output takes significant effort, the trust required to hand agents that work develops more slowly.
Which brings us back to procurement.
Procurement is one of the domains where verification is hardest
A supplier decision does not compile.
A negotiation has no unit test.
Whether the right vendor was selected — or the right risk accepted — may not be provable for twelve or eighteen months, and even then the outcome is entangled with a dozen other variables.
The core work of procurement is judgment applied to conditions that cannot be verified by running them.
That is not a weakness. It is the nature of the discipline. Procurement is an orchestration function more than an execution function: it coordinates suppliers, stakeholders, risk, and commercial terms across relationships that live outside the four walls of the enterprise. And orchestration is precisely the kind of work a machine cannot yet grade for itself.
So procurement’s absence from the chart is not evidence that procurement is behind. It is evidence of where procurement’s value actually lives — in the layer of judgment that is hardest to automate, and therefore among the most difficult domains to hand to an autonomous agent.
A note from 2019
Seven years ago, I wrote a post here called “Take Your Seat,” prompted by an IDC prediction that “digital thinking” could no longer be delegated to the IT department. At the time, a McKinsey survey it cited found that 23% of companies had a digital strategy in place — and just 2% had one for their digital supply chain.
I framed the problem then as one of adoption: procurement needed to stop waiting for a seat at the table and engage the technology.
Reading it again against this week’s data, I would frame it differently now. The gap was never really about adopting tools faster. The 2% figure in 2019 and the absence from the chart in 2026 point to the same underlying thing — that procurement’s contribution is not one a tool or an agent can simply replace. It is one that AI can amplify, but not stand in for. The seat procurement needs, then, is not solely the one that operates the technology. It is also the one that governs the judgment the technology cannot replace.
The facts, I think, speak for themselves. Agents go first where the machine can check the machine. Procurement is where a human still has to be involved. That is not a reason for procurement to rush to catch up. It is a reason to understand, clearly, what procurement is actually for.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Related
AI Agents Are Being Deployed Across Sixteen Domains. Procurement Isn’t One of Them.
Posted on July 10, 2026
0
In February 2026, Anthropic published a study of how AI agents are actually being used in the real world, drawn from millions of human-agent interactions across its API and its coding tool. One chart from that research — which drew attention this week when Andreas Horn shared it on LinkedIn — maps where agents are being deployed, by domain.
There are sixteen named categories on it: software engineering, back-office automation, marketing and copywriting, sales and CRM, finance and accounting, data analysis, academic research, cybersecurity, customer service, gaming, document and presentation creation, education, e-commerce, healthcare, legal, and travel and logistics.
Procurement is not one of them.
Let me be precise, because precision matters here. The data is measured at the level of individual tool calls, so it is possible that some procurement-related activity sits inside “back-office automation,” or inside the unlabeled “other” category. What the data shows is not that procurement activity is literally zero. It is that procurement does not appear as a domain of its own — in a study built specifically to map where agents are working.
That absence is worth sitting with. Not as a grievance. As a question: why?
The domains agents reached first have something in common
Figure: “In what domains are agents deployed?” Software engineering accounts for 49.7% of tool calls; procurement does not appear among the named domains. Source: Anthropic, “Measuring AI agent autonomy in practice” (February 2026).
Software engineering alone accounts for nearly half of all agentic activity — 49.7% of the tool calls in Anthropic’s sample. That is not an accident, and Anthropic’s own explanation is the important part. Software, the report notes, is “comparatively easy to test and review.” You can run the code and see whether it works. The compiler returns an instant, unambiguous verdict: it runs or it doesn’t; the test passes or it fails.
That verdict is what makes software hospitable to agents. Every action an agent takes can be checked — immediately, and almost for free — by the machine itself. The domains where agents arrived first are, broadly, the domains where verification is cheap. The economics of verification matter as much as the capability of the model: AI moves fastest where the cost of proving an answer is low.
Anthropic makes the corollary explicit: in domains where confirming an agent’s output takes significant effort, the trust required to hand agents that work develops more slowly.
Which brings us back to procurement.
Procurement is one of the domains where verification is hardest
A supplier decision does not compile.
A negotiation has no unit test.
Whether the right vendor was selected — or the right risk accepted — may not be provable for twelve or eighteen months, and even then the outcome is entangled with a dozen other variables.
The core work of procurement is judgment applied to conditions that cannot be verified by running them.
That is not a weakness. It is the nature of the discipline. Procurement is an orchestration function more than an execution function: it coordinates suppliers, stakeholders, risk, and commercial terms across relationships that live outside the four walls of the enterprise. And orchestration is precisely the kind of work a machine cannot yet grade for itself.
So procurement’s absence from the chart is not evidence that procurement is behind. It is evidence of where procurement’s value actually lives — in the layer of judgment that is hardest to automate, and therefore among the most difficult domains to hand to an autonomous agent.
A note from 2019
Seven years ago, I wrote a post here called “Take Your Seat,” prompted by an IDC prediction that “digital thinking” could no longer be delegated to the IT department. At the time, a McKinsey survey it cited found that 23% of companies had a digital strategy in place — and just 2% had one for their digital supply chain.
I framed the problem then as one of adoption: procurement needed to stop waiting for a seat at the table and engage the technology.
Reading it again against this week’s data, I would frame it differently now. The gap was never really about adopting tools faster. The 2% figure in 2019 and the absence from the chart in 2026 point to the same underlying thing — that procurement’s contribution is not one a tool or an agent can simply replace. It is one that AI can amplify, but not stand in for. The seat procurement needs, then, is not solely the one that operates the technology. It is also the one that governs the judgment the technology cannot replace.
The facts, I think, speak for themselves. Agents go first where the machine can check the machine. Procurement is where a human still has to be involved. That is not a reason for procurement to rush to catch up. It is a reason to understand, clearly, what procurement is actually for.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Share this:
Like this:
Related