Stop Hiring for Work Agents Can Already Run

Before you approve the next job posting, sit with one uncomfortable question: is this a human-shaped role, or a backlog wearing a job title?

Most SMEs never ask. Workload rises, someone drowns, a requisition opens — the reflex we have been running for a century, because for a century humans were the only available capacity. That era just ended for a specific, well-defined class of work. If the core of a role is running a repetitive, rule-heavy process, an agent can already run most of it — and hiring a person into it anyway is a disservice to your P&L and, less obviously, to the person.

This is a contrarian argument with an empathic core, and we intend to defend both halves. The companies getting this right are not the ones firing loudly. They are the ones quietly redirecting their next three hires — and keeping the people they have by giving them better work.

What agents actually run today — honesty first

No hand-waving; the claim only works if we are specific. Agentic systems in production today reliably run work that looks like this:

• Recurring document production — proposals, quotes, reports, onboarding packets — drafted from your own archive and templates, with a human approving.
• Reconciliation and data movement — matching invoices to purchase orders, syncing records between systems that refuse to talk, flagging what does not line up.
• Status chasing and follow-ups — the polite persistent emails that consume entire afternoons.
• First-line triage — reading, classifying and routing inbound email, tickets and calls, answering what is answerable from documented knowledge, and increasingly the phone queue itself (we covered that front in Voice Agents Will Eat the Phone Queue).
• Recurring reporting — assembly, formatting, distribution, on schedule, without a Sunday evening.

And, with equal honesty, what agents do not run: genuinely novel situations, relationships where trust is the product, negotiation, accountability — a machine cannot own a consequence — and taste. Anyone selling you agents for those is selling vibes. The boundary matters because everything on the first list is exactly the work that fills job descriptions titled "coordinator", "assistant", "back-office analyst" — roles that are, drawn honestly, 60 to 80 percent repetitive layer with a thin, precious band of judgment on top.

This is about the job description, not the person

Here is where the argument turns humane, and where we part company with both camps in the usual debate.

Nobody's ambition was re-keying invoices. No one leaves university hoping to spend Tuesdays copying numbers between systems and chasing status by email. We build roles like that not because humans are suited to them but because, historically, there was no one else to give them to. The repetitive layer is not a career; it is sediment.

And your best people know it. Look honestly at why capable employees leave SMEs: it is rarely the salary alone. It is the discovery that the interesting third of their job is buried under a mechanical two-thirds that never shrinks. The repetitive layer is not just an efficiency problem. It is your attrition engine.

The kindest thing you can do for a good employee is stop paying them to be a courier between systems.

So the empathic move and the economic move point the same direction: take the mechanical layer away from humans — not the humans away from the company.

Automation as retention, not replacement theatre

There is a cynical version of this essay's headline, and you have seen it: a company cuts a department, calls it an AI strategy, and issues a press release about efficiency. We think that version is not only callous but incompetent, for a cold, practical reason: the judgment layer those people carried — the exceptions they knew, the clients they read correctly, the tribal knowledge in their heads — walks out the door with them, precisely when the agents running the mechanical layer need experienced humans supervising the seams. Replacement theatre burns the asset the automation depends on.

The competent version inverts it. Automate the layer, keep the people, promote the work:

1. Decompose the roles you have — with the role-holders, not about them. They know better than anyone which parts of their week a machine should own; they have been complaining about those parts for years.
2. Move the mechanical layer to agents with approval gates, one workflow at a time — a focused, production-hardened workflow is typically a 4–8 week, €18k–45k build: a one-off, where the requisition it replaces recurs every year — exactly the discipline we describe across our automation practice.
3. Redefine the role around what remains: exceptions, client relationships, quality, ownership of outcomes. Put it in writing — a role redesign that lives only in a reassuring speech is a rumour, and your people will treat it as one. Say plainly what the automation is for: growth without grind, not headcount removal. Then behave accordingly.
4. Create the supervision career path. Someone has to own the agent fleet — its workflows, its exception queues, its performance reviews. Your detail-obsessed ops person is frequently the perfect fleet owner; it is a promotion that did not exist three years ago and it is a better job than the one it replaces. The org shape this produces — few humans, many workflows — is the one we mapped in Small Teams, Big Leverage.
5. Aim the freed hours at revenue-side work — the advisory conversation, the follow-up that never happened, the service line nobody had capacity to launch. Retention improves twice: the boring layer is gone, and the growth it funds creates the roles people actually want next.

What this looks like in practice

Our standing rule: no invented clients, no fabricated headcount savings. Our own operation, and an archetype.

Gravitnomad's evidence is the hires that never happened. Our content operation — one brief becomes a blog article, a LinkedIn post and a Facebook post, drafted, formatted and scheduled by agents with a human gate — is a content-ops role we never posted. The multi-tenant engine that renders our websites from structured brand data is the webmaster we never needed. The assistant on this site, answering visitors from retrieval over every page with persistent memory, is the front-desk shift no human should have to cover at 03:00. The humans here do judgment: architecture, client work, editorial approval. That is not a manifesto; it is our payroll behaving the way this article recommends — the systems do the sediment.

The archetype: a regional accounting firm where juniors spend most of their hours on data entry and chasing documents — three juniors at half a week each is roughly 60 hours of courier work a week, one and a half full-time roles hiding inside three job titles. The agents take intake, matching and follow-up; the juniors move up the stack to review and client advisory — the work the partners never had bandwidth to delegate properly. The firm's next hire is a different species entirely: someone hired for judgment, not for throughput. Multiply that shift across a few years and you get the operating model we sketched in The 2030 Back Office.

The transition conversation

Most automation programmes do not fail on technology. They fail in a meeting room, in the first ten minutes of a badly-run announcement. So it is worth being concrete about the conversation itself.

Two scripts reliably backfire. The corporate reassurance — "nobody will lose their job" delivered in the tone of a press release — is heard as its opposite, because everyone has watched that sentence age badly somewhere else. And silence is worse: a pilot run in secret becomes a rumour, and the rumour is always darker than the plan.

What works is blunt specificity. Name the sediment out loud — "the reconciliation, the chasing, the re-keying: we are taking that off your desks" — and let the role-holders lead the decomposition, because they know the mechanical layer better than any consultant and have been complaining about it accurately for years. Put the redesigned role in writing the same month, with the new ownership spelled out. Automate the most hated task first, not the most valuable one: the first win should feel like relief, and relief creates advocates. Then make the first fleet-supervision promotion loudly visible.

Do that, and the question in the corridor changes from "am I being replaced?" to "when is it my workflow's turn?" — which is the only adoption metric that matters.

The hiring filter going forward

Add one test to every future job description: what fraction of this role, honestly decomposed, could an agent run within a year? If the answer is more than half, do not post it — redesign it. The arithmetic leans your way: the automation is a bounded one-off with evolution typically at €2k–6k a month, while the requisition it pre-empts is a salary that recurs, with raises, indefinitely. Hire for the judgment band, automate the sediment, and tell candidates exactly that in the interview. The good ones will pick you because of it. Nobody fights for the privilege of re-keying invoices.

If you are staring at a requisition right now and wondering which kind of role it is, talk to us before you post it. Bring the job description; we will help you decompose it honestly — including the parts where the honest answer is "no, that still needs a human, hire the human". Both answers are wins. Only the unexamined requisition is a loss.