AI agents vs hiring: what a 200-person company should automate first

There is a pattern that repeats in companies with about two hundred people. They sit in the middle: large enough that nobody sees everything anymore, yet small enough to pretend someone still can.

Someone asks for a report. Three people spend two days compiling it from four different systems. The report is useful for about three days. Then someone asks for it again. Nobody questions why it exists. Nobody asks what decision it supports. It just runs, like a ritual.

This is the first thing to understand before you hire someone or buy an agent: most of the work you are considering automating is not work. It is the organizational equivalent of walking water uphill.

The hiring reflex is understandable. A manager notices that their team is drowning in manual reconciliation, reporting, and routing approvals. The team complains. The manager asks HR for another headcount. HR says the budget is tight but maybe one more person. The company hires someone with a generic title like Operations Coordinator. Six months later, that person is drowning too.

The problem was never headcount. The problem was that three separate workflows were silently consuming twenty hours a week across four people, and nobody had mapped them because they were too small to justify a project and too embedded to question.

You do not have a headcount problem. You have a workflow problem that you keep trying to solve by hiring another person to do the same invisible work.

Organizations behave this way after growing beyond one person's field of view. Informal processes become the real processes. Undocumented spreadsheets become the system of record. Only the people doing the work can still see it.

Decide which workflows deserve an agent, which need a person, and which should disappear entirely.

The companies that get this right start with an inventory of what people actually do on a Tuesday morning. The org chart and the SOP usually tell a cleaner story.

In the War Map engagements, the first week is usually spent answering five questions about each candidate workflow. How much volume? How many handoffs? Where does the context live? What is the actual cost when it breaks? And what kind of intervention does this actually need: deterministic automation, supervised agent, human-owned decision, or redesign first?

The answer is rarely obvious. The workflow that looks like the biggest time sink is often the one with the most undocumented exceptions. The one that seems minor turns out to be a bottleneck that blocks twelve other processes. The thing everyone complains about is usually a symptom of something else.

Here are the three patterns that show up most often in companies of this size, and what tends to work for each.

Reporting. Not strategic reporting. The operational kind. The weekly sales dashboard compiled from three CRMs and a spreadsheet. The monthly financial reconciliation that takes two people four days. The compliance report that everyone hopes is accurate but nobody has time to verify. These are not creative tasks. They are extraction, transformation, and assembly. They are almost always the first place to put an agent, because the work is structured enough to automate and voluminous enough to justify it. The risk is low: the agent produces a draft, a human reviews it. The time saved is real and measurable.

Reconciliation. Invoices against purchase orders. Inventory counts against system records. Timesheets against project budgets. Two systems contain overlapping information and a person checks whether they match. Pattern matching at this scale exhausts people. A controlled system can read both sources, flag discrepancies, and present them for human judgment.

Routing approvals. Purchase requests. Expense reports. Contract revisions. Something moves from A to B to C while each person checks, approves, or sends it back. Routing creates the delay. Predictable cases can follow deterministic rules while exceptions go to a person.

Leave workflows involving company direction under human judgment until leadership agrees on what a good answer looks like. If a client approval takes three weeks because nobody owns the decision, an agent will only accelerate a bad process.

McKinsey's MGI research published in 2024 found that while roughly 90% of European companies report using AI, fewer than 40% report seeing measurable results from it. The gap is not technology. It is scope. Companies deploy agents on workflows that are not ready, skip the mapping step, and then wonder why the agent produces errors or gets abandoned after two weeks.

A process problem automated is just a faster process problem.

This is why the first question in any engagement is never which tool should we use. It is what work are we actually doing, why are we doing it, and what would happen if we stopped?

The hiring decision and the automation decision are the same decision viewed from different angles. Both are responses to a capacity gap. The question is whether that gap comes from too much work or from work that should not exist in its current form.

If you hire someone to do a process that should not exist, you have now institutionalized the waste. That person will build their identity around it. They will defend it. They will train someone to replace them when they leave. And three years later, you will be paying two people to do something that could have been eliminated with one afternoon of honest questioning.

If you automate a process that should not exist, you waste money on tooling and integration for something that produces no value. The agent works perfectly. The output goes to nowhere. The cost is lower than a person, but it is still a cost.

The sequence matters. Map first. Question second. Decide third. Build fourth.

There is a cost question underneath all of this, and it deserves a direct answer.

Hiring a full-time operations person to handle reporting, reconciliation, and routing typically costs between 35,000 and 55,000 euros per year in a mid-market European company, depending on seniority and location. That market cost still cannot guarantee correct, consistent, or sustainable execution. People leave. People get sick. People have good days and bad days.

A system intervention for the same scope of work requires someone who understands the workflows, the systems, and the exceptions. It is not a plug-and-play purchase. It begins by deciding whether the case needs a War Map or a bounded Feasibility Sprint, then agreeing scope, evidence, timing, and risk controls for the actual operation.

The comparison is not cost versus cost. It is what each option actually produces. A person produces capacity and variability. An agent produces capacity and consistency, if it is built on a real understanding of the work. Neither is inherently better. The wrong choice for your situation is the expensive one.

The question to ask before you post a job requisition or sign an agent vendor contract is the same one that matters for any operational decision: what work are we actually doing, and what would happen if we did it differently?

Not hypothetically. Specifically. Which workflow. Which volume. Which handoffs. Which person currently owns it. Which system contains the context. Which decision requires human judgment. Which step is just waiting.

The companies that answer these questions before they hire or automate are the ones that see results. The ones that skip them are the ones that end up with a new employee doing old waste, or a new agent doing old waste faster.

A third option comes first: understand the work.