How Procurement and Risk Decisions Are Quietly Splitting the Talent Intelligence Workforce Into Three

I had three conversations this week with talent intelligence leaders. Each was wrestling with the same problem. Each walked away from it with a completely different outcome. And the variable that decided which outcome they got was not their talent, their ambition, or even the vendor they happened to be using. It was the procurement decision their organisation had made, and the risk posture wrapped around it.

That should worry you more than it probably does. But before I introduce them, we need to settle a definition, because the entire story turns on a word that has been stretched well past the point of usefulness.

---

First, Let’s Be Clear About What an Agent Actually Is

The word “agent” is doing far too much work right now. It has become a marketing term, applied to everything from a saved prompt to a genuinely autonomous system, and that ambiguity is a meaningful part of the problem this piece is about.

So, some grounding from the people building this technology.

IBM defines an AI agent as a system capable of autonomously performing tasks on behalf of a user or another system. Google describes AI agents as software systems that use AI to pursue goals and complete tasks, exhibiting reasoning, planning, and memory. Anthropic frames it as systems where the language model dynamically directs its own processes and tool usage, maintaining control over how it accomplishes tasks. AWS lands in the same territory: a program that interacts with its environment, collects data, and uses that data to perform self-directed tasks in pursuit of a predetermined goal.

Microsoft draws a deliberate line between two things. A copilot, in their framing, is an AI-powered assistant that supports tasks, offers insights, and boosts productivity. An agent is a specialised tool built to handle a specific process or solve a business challenge. Their analogy is a neat one: agents are the apps of the AI era, and the copilot is the interface you reach them through.

The consistent thread across all of these is agency. The capacity to determine the best path to a goal, rather than to execute a path a human has already mapped out in full.

An agent is not a prompt with a name. It is not a chatbot with a clever system instruction. It is not a workflow where every step was written by a person and the model simply fills in the gaps. A genuine agent receives a goal, reasons about how to approach it, selects and uses tools, adapts when it hits an obstacle, and delivers an outcome. The human sets the destination. The agent works out the route.

Hold that distinction in mind, because the three people I want to introduce you to were all told they had agents. Only one of them actually did.

---

Three Practitioners, One Brief

All three lead a talent intelligence function inside a large organisation. All three were handed the same mandate this year, and it is a mandate every TI and TA leader will recognise: the central team cannot scale to meet demand, so we need a self-service capability. Let the hiring managers run their own labour market queries. Let recruiters generate their own market maps and competitor views. Let the business pull intelligence without forming an orderly queue outside one overworked analyst’s desk.

Same problem. Same goal. Same week. Here is what each of them was actually able to build. Names changed for anonymity but the situations are all true.

Maya works inside a large, heavily regulated enterprise. Her organisation runs Microsoft Copilot across the M365 estate, and the rollout was announced internally as the company’s “AI transformation.” Maya was genuinely excited. She built what the platform let her build, which it calls agents, and which turned out to be saved prompts with knowledge grounding and a name attached. Her best effort was a labour-market assistant her stakeholders could open in Teams and ask questions of. It returned plausible answers in clean prose with data derived from their sharepoint. But it could not reliably reach live data, could not run anything on a schedule, and could not do a single thing unless a human opened it and typed. It helped people write better questions. It did not touch the scaling problem, because every output still depended on a person sitting in the seat. Maya’s honest assessment, offered quietly, was that she had shipped a better search box.

Daniel works somewhere more permissive. His enterprise agreement gives the team access to Claude, and the difference was immediately obvious. The outputs were richer. The artifacts were genuinely useful: interactive dashboards, formatted market reports, visual skills breakdowns that stakeholders could actually use. Daniel was able to give the business something close to real self-service, and for a while it felt like he had won. Then he hit the wall. No agent builder access. No orchestration layer. No way to connect outputs to other systems or to let anything run without him. His tool was powerful, but it stopped at the edge of his own attention. Every analysis still began and ended with Daniel. He had a magnificent instrument and no orchestra.

Priya works in an organisation that took a different view of risk. She was able to build properly. Specialist sub-agents, each owning a defined slice of the problem: one for labour supply, one for competitor monitoring, one for compensation benchmarking. A conductor layer that routes a stakeholder’s request to the right combination of them. Automation connecting the outputs into the systems the business already uses. Pipelines that trigger on events rather than waiting for a human to start them. When a hiring manager submits a request, the system reasons about what is being asked, gathers what it needs, and returns a finished piece of intelligence. Priya genuinely solved the brief. She built a function that scales without her in every loop.

Three skilled practitioners. One identical problem. The distance between Maya’s better search box and Priya’s autonomous intelligence function is not a difference of effort, talent, or ambition. It is the difference between three decisions made in rooms none of them were in.

---

We Predicted This. In 2023.

In December 2023 I published a piece arguing that the era of bring-your-own-device was giving way to something more significant: bring-your-own-AI. The argument was that high performers would begin to accumulate personalised AI capability, tuned and portable and proprietary to the individual, carried from role to role the way an elite athlete carries their own coaching and conditioning team rather than relying on the club’s.

What I did not see coming was how quickly the procurement decision itself would become the binding constraint. The split I expected was between people who embraced AI and people who didn’t. The split actually unfolding is between people whose organisations give them room to build and people whose organisations have made choices, about contracts, licensing tiers and risk, that quietly define their ceiling before they ever touch a tool.

Maya, Daniel and Priya are all true believers. The divergence between them has nothing to do with mindset.

---

The Three Tiers Are Hardening

What I saw this week was not a spectrum. It was three distinct tiers, and the walls between them are setting like concrete.

Tier one is the branded prompt interface. Maya’s world. Saved instructions with a friendly name, deployed across an estate and announced as transformation. There is real value here, but it is the value of an assistant in the most limited sense. It helps a person do what they were already doing, a little faster. It does not compound, and it does not remove the human from the loop, which means it never truly scales.

Tier two is the powerful but bounded individual tool. Daniel’s world. A capable model producing genuinely strong analysis, synthesis and artifacts. This expands what one person can produce, sometimes dramatically. But the boundary is the person. Nothing builds and runs without a human present, so the ceiling is the user’s own time and focus. It is a better instrument, played one note at a time.

Tier three is agentic infrastructure. Priya’s world. Orchestrated systems where specialist agents own specific functions, a conductor layer manages their interaction, and automation connects the whole thing to the systems a business actually runs on. Work happens overnight. Pipelines trigger on events. Analysis no longer waits for a person to begin it. This is not AI used as a tool. It is AI built into the operating model.

The gap between tier one and tier three is not a feature gap. It is a generational gap in capability. And the people in tier three are not waiting for the others to catch up. They are accelerating away.

---

It Was Never About the Vendor

Here is where it would be easy to be lazy, and where the truth is far more interesting.

You might assume Maya was stuck because Microsoft cannot do tier three. That is simply not the case, and the point matters.

Microsoft can absolutely build genuine agents. Copilot Studio supports autonomous agents that perceive events, make decisions, and execute tasks independently through triggers, instructions and guardrails, operating continuously in the background rather than waiting for a prompt. As of mid-2026 it offers generally available computer-using agents that interact directly with websites and desktop applications through the interface, automating processes that previously relied on brittle scripts because the underlying systems lacked APIs. It has moved from standalone agents toward composable, multi-agent systems. By any of the definitions above, that is real agency. Priya’s tier-three world is entirely buildable inside the Microsoft ecosystem. The capability is genuinely there.

So why was Maya stuck with a better search box?