Total Workforce Management Now Includes Agents. HR Should Own It.

At its conference on 1 September, CrowdStrike said a Fortune 500 customer scanned its endpoints and found 18,000 AI agents running. That customer had approved 300. CrowdStrike did not name the company, and the scan was run by Falcon Guardian, the product it launched that morning.

Sixty to one is not a rounding error in a policy. It is the distance between what a company believes it is running and what it is actually running. A company acquired a workforce and did not notice.

We already solved this once, for contractors

Walk around a large company and a lot of the people doing the work are not employees. They are contractors, agency temps, freelancers, and people billing against a statement of work. For years nobody senior could say how many there were, because every business unit hired its own against its own budget and told no one.

So companies built a job to fix it, and the industry calls it Total Workforce Management. One list covering everyone doing work for the company, employees and non-employees together, so that four questions have answers: how many are there, who approved each one, what are they costing, and when does their access get switched off. It runs on software called a vendor management system, often operated by an outside firm.

Agents belong on that list and they are not on it. They do work that would otherwise be staffed. They cost money. They hold logins. Somebody has to switch them off when the work is done. What they do not have is an approval, a manager who answers for what they produce, or a line in anyone's budget.

A company that spent years getting a grip on its contractors now has a bigger workforce it cannot count.

Why the productivity has not reached the P&L

McKinsey's 2026 State of AI survey is the clearest read on this. Around 80 percent of respondents report productivity improvement. Thirty-seven percent attribute any EBIT impact at all, which is roughly flat on last year. About 6 percent qualify as high performers, meaning they credit at least 5 percent of EBIT to AI. Large companies scaling agents rose from 27 to 40 percent, so the deployment is real while the financial result sits still.

That combination stops being strange once you accept that most of the fleet is invisible. Capacity nobody can name cannot be redeployed on purpose, and capacity that is not redeployed on purpose gets absorbed into slightly shorter days and slightly less overtime. It never reaches the P&L, and a CFO is right to refuse to book it.

The survey has one more number that points the same way. High performers are far more likely to run defined human validation in the loop, 65 percent against 23. Validation requires knowing which agent did what, which means somebody built the list first.

Three people it should not be

The CISO gets the blast radius and none of the budget authority over the work itself. Give security the ownership and the first deliverable is a block list. We have seen how that goes: when central IT was too slow and business units bought software on company cards, the block did not stop them, it moved them onto personal accounts where nobody could see anything at all. Most of those unapproved 17,700 agents are engineers shipping faster, and a restrictive policy destroys the 80 percent productivity finding in order to close a governance gap. Security belongs here as a control, not as the owner.

The CIO owns the tooling and the identity plumbing, which is necessary and not sufficient. What the CIO does not own is the work the agents are doing, so this ends as a complete inventory that no business leader has a reason to act on.

The CFO owns the budget line and will eventually force the question, but finance can only count categories somebody else has defined. Ask the CFO to own it and you get a cost centre before you get a definition.

HR was invented for exactly this

The first recognisable HR department was created at National Cash Register in the early 1900s, after a run of strikes and walkouts. The company had a workforce problem its line managers could not handle, so it built a function to handle it. Within twenty years that function had settled into what most people still picture: records, payroll, hiring, attendance, and later compliance.

That is the whole shape of the profession's history. It was created to answer a hard question about the workforce, and then spent a century administering the answer.

The question has changed underneath it. For the first time since the department was invented, the definition of who works here is genuinely open. A company's capacity is now some mix of employees, contractors, and software that does work, and the mix is moving every quarter. Somebody has to hold one number that covers all of it.

HR is the only function that has ever held that number. The people who run the contractor list already know how to add someone, take their access away when the work ends, argue with a vendor about the rate, and report one headcount that counts employees and non-employees together. That is the job. Agents are a new category on a list HR already keeps.

This is the once-in-a-century opening for HR to change what it is. Not a bigger seat at the same table. A different job: the function that owns the company's total capacity, human and otherwise, and can say what it costs and what it produced.

What HR has to become to earn it

HR cannot take this as it stands today, and pretending otherwise would make the argument useless.

Owning agents means being able to say what counts as one, which is a technical definition that has to hold up against engineers. It means writing the equivalent of a job description for a piece of software: what it is allowed to touch, what it may spend, what it must never do without a person. It means reviewing performance of a fleet rather than a person, using output data rather than a manager's opinion. It means answering the redeployment question honestly when an agent absorbs work that used to be a role, which is the conversation the whole organisation is most afraid of.

Most HR functions are not staffed for that today. Neither was personnel management in 1910. The reason to take it anyway is that every other candidate is missing something HR already has, and the gap HR needs to close is skills, which is the one gap a function can actually close on purpose.

What happens if HR passes

The work gets done by someone. The CIO builds an inventory, or the CISO writes a policy, and the company ends up with a technically accurate list of running processes that answers no question a business leader was asking. The agents stay outside the workforce numbers, the capacity stays unbooked, and the McKinsey figures look about the same next year.

HR stays where it is, which is the part that should sting. A function created to solve the hardest workforce problem of its era will have sat out the biggest change to the workforce since it was founded.

The market has already priced the problem. CrowdStrike launched Falcon Guardian on 1 September, which finds agents on endpoints and blocks the unapproved ones, and expanded its OpenAI partnership the next day. HiddenLayer, an Austin security company that protects AI models and agents while they run, raised 100 million dollars on 2 September. JetStream, backed partly by CrowdStrike's own venture fund, launched Clearance the same day with 34 million behind it, which checks each agent action before it runs and stops the ones falling outside what that agent was authorised to do. Three funded answers inside 48 hours to the question of how a company sees and controls agents at runtime.

Tools will not decide this. Whoever ends up answering three questions about the fleet will: how many agents are running, who owns each one, and which budget line pays for it.

Those are workforce questions. They have always been HR's questions. The only thing that has changed is that some of the workforce is now software.

Written in a personal capacity.