Academy

Hire, coach, and learn to trust your AI employees

You install software; you hire an employee. That small difference changes everything about how you build a digital organization.

The Felesh team 7 min

For years we have bought AI the way we buy a tool: a new capability bolted onto the product, a button we press to make some text appear. But step back, and the thing actually taking shape does not look like a tool. It looks like a workforce.

We have come to believe that agentic systems should not be treated like software — they should be treated like staff. This is not a marketing metaphor; it is an engineering decision that determines how we build these systems, how we evaluate them, and how we come to rely on them over time. You install software and expect it to work flawlessly from the first second. You hire an employee, coach them, and over time learn where you can lean on them and where you can’t yet.

Hire, don’t install

When you install software, you buy a specification: defined input, defined output, deterministic behaviour. When you hire a person, you buy something else — a capacity. A capacity that has to take shape inside the context of your work, get tuned to your real tasks, and settle inside the boundaries of responsibility you set for it.

Installing software versus hiring an employeeA two-column comparison in three rows. Installing software, you buy a specification: on day one an invoice goes in and fields come out, the same every time; in week six it is identical to day one, by design; it does what the spec says and nothing more. Hiring an employee, you buy a capacity: on day one it reads invoices slowly and asks when unsure; in week six it knows your suppliers and your formats; it has a role, accounting, and knows where its job ends.install softwareyou buy a specificationhire an employeeyou buy a capacityday oneweek sixits edgeinvoice in, fields outsame every timeidentical to day oneby designdoes what the spec saysnothing morereads invoices slowlyasks when unsureknows your suppliersknows your formatsrole: accountingknows where it ends
Software is bought as a specification and stays as it is; an employee is hired as a capacity that takes shape inside your work.

AI agents are the same. A good agent is not a general capability that will accept any job; it is a role with a defined domain. Just as no healthy organisation hires someone for “everything” — you hire for a specific role: accounting, support, analysis — the boundary of the role is exactly what makes an agent trustworthy. You know what you are asking of it, and more importantly, you know what you are not asking of it.

This reframing changes the design. Instead of one giant model expected to be an expert in every domain, you build a team of focused roles, each of which does its own job well and knows where that job ends.

Roles with edges hand over what is not theirsThree roles, each inside its own dashed edge: support owns order status and the returns policy and does not handle payments; accounting owns invoices and refunds and does not chat; analysis owns churn by plan and the monthly report and takes no action. A customer asks where the delivery is, and support answers inside its edge. The next request, refund a deposit of 840K, moves money, which is outside support’s edge, so support hands it to accounting instead of guessing.customer: “where is my delivery?”customer: “refund my deposit, 840K”supportorder statusreturns policynot: paymentsaccountinginvoicesrefundsnot: chattinganalysischurn by planmonthly reportnot: any actioninside its edge, so support answersmoves money: outside support’s edgehanded to accounting, not guessed
Each role has an edge, so a request inside it is answered and one outside it is handed over rather than guessed.

Coach, don’t configure

A new hire is not at peak performance on day one. They get good because they get feedback. Someone watches the work, says this part was right and that part wasn’t, and they do better next time. That simple loop — work, observe, correct — is what turns a novice into a specialist.

We do exactly this with our agents, just in engineering vocabulary. Instead of “feedback,” we have evaluation: a set of real examples that show where the agent performs well and where it slips. Evaluation plays the role of the coach — it lights up precisely where the weakness is, not by guessing, but with evidence.

And once the weak point is identified, the training is focused too. We don’t rebuild the whole agent; we correct the specific capability that is leaking, with the smallest intervention that does the job. That is the difference between coaching and rewriting. A coach doesn’t replace the person; they strengthen exactly the point that needs it.

Evaluation finds the weak spot, one rule fixes itAn evaluation of the accounting role on five real examples: match invoice to order, flag duplicate invoice, pay Nova 3.4M, post monthly entry, reconcile bank line. Four pass. Paying Nova 3.4M went out alone, and it is lit as the weak spot. The fix is one rule: payments over 2M go to a person. On the next run that row goes to a person, and the other four rows were not touched.evaluation · accounting · 5 real examplesmatch invoice to orderflag duplicate invoicepay Nova 3.4Mpost monthly entryreconcile bank linepaid aloneto a personthe fixpayments over 2Mgo to a personone rule onlythe other four rows were not touched
Evaluation shows exactly which example went wrong, and the fix is one rule on that spot, not a rewrite.

Trust is earned

No sensible manager hands a new hire the keys to the vault on the first day. Trust is not granted all at once; it is earned along a path. First small tasks, supervised. Then larger tasks, with less oversight. And eventually full responsibility — but only in the domain where the employee has repeatedly shown they can handle it.

Trust earned per domainA grid of three trust levels by three domains. Small tasks under supervision: support drafts replies, accounting drafts entries, analysis pulls figures. Larger tasks with less oversight: support sends replies, accounting posts entries, analysis not yet. Full responsibility: accounting pays up to 2M, support and analysis not yet. Trust is per domain: accounting has it, support is partway.supportaccountinganalysissmall taskssupervisedlarger tasksless oversightfullresponsibilitydrafts repliesdrafts entriespulls figuressends repliesposts entriesnot yetnot yetpays up to 2Mnot yettrust is per domain: accounting has it, support is partway
Trust is earned per domain: accounting has reached full responsibility, support is partway, analysis has only started.

This is perhaps the most important lesson of the “staff, not software” view. We do not assume trust in our agents. We measure it. Any agent that takes on real responsibility must have shown that, within that responsibility, its behaviour is stable and inspectable. We are on the side of the glass box, not the black box: you should be able to see what the agent did, why it did it, and where it should be stopped.

Black box versus glass boxTop, a black box: a Nova invoice goes in, an opaque box with no trail, and paid 840K comes out, so you only see what went out and cannot answer why 840K or who stops it. Bottom, a glass box shows three readable lines: it paid Nova 840K; the reason, the invoice matches its order; and where it stops, payments over 2M go to a person.black boxNova invoiceno trailpaid 840Kyou see what went out: why 840K? who stops it?glass boxdidpaid Nova, 840Kwhyinvoice matches its orderstops atpayments over 2M go to a personreadable at every step, stoppable at the last
A glass box shows what the agent did, why, and where to stop it; a black box shows only what went out.

Earned trust is different from blind trust. The first is durable, because it stands on evidence. The second is brittle, because it stands only on hope. We build systems that earn trust — not ones that merely ask to be trusted.

A digital organisation that grows

When you see AI as a workforce rather than a tool, what you are building is no longer a product; it is an organisation. One that can hire, learn, and grow more trustworthy over time. And an organisation, unlike a static feature, compounds: every role that settles in, every evaluation that reveals something, every small correction that lands, accumulates.

This is the view Felesh builds with. We don’t build chatbots; we build digital colleagues — focused, coachable, inspectable roles that earn trust. If AI is going to genuinely become part of your work, it deserves to be treated with the same care you give to hiring and growing a human team.

A good employee doesn’t perform miracles on day one. Over time, they become something you can lean on. A system built this way travels the same path.