Delegating to an AI agent: where to start

Start with the work you can check in ten seconds. A follow-up sent on the right channel, an interview write-up, an updated candidate record: you know whether it is good by reading it. That is the only criterion that matters for the first thing you hand to an agent.

The classic mistake is starting with the most impressive. Sourcing is the demo that lands, five profiles in thirty seconds, but it is also the task whose quality you cannot judge without spending an hour on it. You delegate, you have no idea what to make of the result, you redo it by hand. The agent has saved nothing.

Why verifiability comes before time saved

An agent does not settle into a team because it is fast. It settles in because the first three results were good and somebody noticed.

That is a trust mechanism, not a productivity one. As long as you are reading everything closely, you have not delegated: you have added a step to a process that already had too many. The switch happens the day you stop re-reading, and you only stop re-reading what you have been able to check easily, several times over, without an unpleasant surprise.

Hence the order: first what a glance settles, then what takes judgement.

The same logic applies to volume. Manual CV screening runs to roughly 106 hours a month for eight open roles, and automation brings human review down to 15–20% of that. The gain is real, but it only reaches whoever first learned to trust the agent on smaller work.

The order that works, week by week

Week 1, follow-ups. Nothing is more mechanical and nothing is more expensive to forget. A placement is rarely lost on skills; it is lost on a message that never went out. The agent spots conversations with no reply, picks the channel where the contact actually answers, email, LinkedIn, WhatsApp, and writes a follow-up that picks up the thread rather than a template. You approve before it sends. You can tell instantly whether the message is right, because you would have written roughly the same one.

Week 2, interview write-ups. The agent listens to the call, pulls out skills, expectations, availability and reservations, then updates the record. The gain is double: you stop taking notes while you talk, so you listen better, and the record actually exists instead of existing “as soon as I get five minutes”. It is also the first time the agent writes into your ATS, which teaches you a lot about what it understands of your work.

Week 3: the morning digest. What moved since yesterday, the three priorities of the day, and the reason behind each. This is the first task where the agent does not merely execute: it sorts, so it ranks, so it can get the order wrong. You are ready to let it sort because two weeks of follow-ups and write-ups have taught you what it grasps and what it misses.

Week 4, anticipation. Missions ending within sixty days, consultants coming free, matching client needs. This is where the money is: a consultant on the bench costs a French staffing firm around €5,480 a month. Anticipating a roll-off by three weeks is worth more than ten hours of sourcing, and it is exactly the work nobody has time for on a Tuesday afternoon.

The three questions before delegating anything

Before handing over a task, ask three questions. They take ten seconds and head off most disappointments.

Would I know it was done well just by reading it? If answering takes an investigation, the task is not a first-month task.

Is the mistake still repairable? A bad draft gets thrown away. A sent email does not.

Could I explain it to a new colleague in three sentences? If your criteria live in your head and have never been written down, the agent has no better odds than an intern you told nothing. That is not a limit of the technology, it is a limit of the brief.

Tasks that pass all three are, without exception, the ones in the first four weeks.

What it assumes about your tools

One thing only: that the agent reaches the place where the work already happens.

An agent living in its own interface gets delegated very little, because you have to remember it, go there, and re-type the context. An agent you tag in the channel the team already uses gets delegated by reflex.

The rest is a question of connections: your ATS to write into records, your inbox to send, your job boards to search, your calendar to offer a slot. That is what the list of tools Balt plugs into covers, and the useful test is not how many there are, but whether the four or five you open every morning are among them.

What it should not be given in the first month

Anything that leaves the company without you having read it. An inaccurate write-up gets corrected; a message to the wrong candidate does not, and it gets retold.

And any decision that closes a door: rejecting a profile, choosing between two consultants for a mission, delivering a no. An agent can prepare those decisions, gather the material, lay out the options, flag what is missing, but taking them is a different job. Handing it over too early is the surest way to stop trusting the agent with everything else, and it is a regulatory question too: we go into who has to approve what, and why the line is not where instinct puts it.

Finally, do not delegate what you cannot describe. If you cannot state your sourcing criteria in three sentences, the agent will not guess them, and the problem is not the agent.

The mistake of delegating a decision first

There is one shortcut worth naming, because it is the most tempting and the most destructive: handing over a decision before you have handed over a task.

It looks efficient. Rather than asking for a shortlist and choosing yourself, you ask the agent to pick. Rather than reading follow-ups before they go, you let them go. You have skipped the boring part and gone straight to the payoff.

What happens next is predictable. The agent makes a choice you would not have made, not a wrong one necessarily, a different one, and you have no way to tell which, because you never built the sense of how it reasons. So you take everything back, and the agent is now something you supervise rather than something you use.

Trust is built on tasks and spent on decisions. Spend before you build and you run out immediately. This is also the shape of the approval line every team ends up needing: tasks flow, decisions do not.

How long before it becomes a reflex

Three to four weeks. Not because the tool takes time to learn, but because you take time to think of it.

That is the part nobody plans for. The agent is available on day one; the habit of asking is not. For a fortnight you will do half of it by hand out of pure muscle memory, then catch yourself thinking “I could have given it that”. That sentence is the real adoption metric.

The sign that it has landed is not a usage statistic. It is the day somebody on the team asks the agent for something nobody suggested, and is annoyed that it is not possible yet. That day the question has changed: it is no longer “does this work”, it is “what do we hand it next”.

And that question leads straight to a less comfortable one: who approves what, when the agent is writing to your candidates.