A hiring process,
not a software project.
Nine steps, the same every time. It looks like recruitment because that is genuinely the better model. You are adding a worker to a team, with a defined role, defined limits and a defined way of being judged.
The nine steps of the Custom Employee Method
Job analysis
What does this employee actually need to accomplish? Who does it today, how long does it take them, and what is it worth when it happens correctly? We start with labour and economics, not with AI. If the numbers don't work here, everything downstream is wasted effort, so this is where we're most willing to tell you to stop.
Workflow mapping
We follow the work end to end: every step, every decision, every lookup, every exception, every "well, actually, when it's a Tuesday we do this instead". Then we classify each step: AI reasoning, deterministic calculation, system lookup, system action, human approval, or exception. That classification is the architecture.
System access
What information and tools the role requires, and how the employee reaches them safely. Real credentials, scoped as tightly as the job allows. This is usually the step that determines the timeline. Not the AI, but how quickly we can get authenticated access to the systems the work actually lives in.
Authority design
Every action the employee could take is classified: does it alone, needs a human first, or must never happen. The forbidden set is enforced in code, not requested in a prompt. An employee cannot choose to do something it has no ability to do. You sign this off before anyone writes software.
Build
The employee gets made. Judgement, planning and recovery are the model's job. Validation, arithmetic, permissions, idempotency, retries and audit logging are code's job. Getting that division wrong is the most common reason these projects fail: an AI asked to do arithmetic, or a decision tree asked to exercise judgement.
Acceptance test
Before it works a single real day, the employee is scored against a bank of realistic tasks drawn from your actual history, including the awkward ones. Resolution accuracy, escalation correctness, policy violations, tool success rate, response time. You agree the pass mark in advance. It passes or it doesn't ship.
Probation
Deployed under supervision. Shadow, then draft, then supervised, then autonomous, and it only climbs when the numbers justify it. You control the pace, and you can hold it at any rung indefinitely.
Employment
Approved classes of work graduate to independent operation. Exceptions still escalate to a named human. Every action leaves an audit trail you can read, so nothing the employee did is ever a mystery.
Performance management
A written monthly review with real numbers, including the unflattering ones. New edge cases become new training. Integrations get maintained as your systems change underneath them. When the employee earns more authority, we propose it; when it should be reined in, we say that too.
Every employee gets a job description.
Written before any code exists, agreed with you, and used as the specification everything is built and tested against. Here is the shape of one. This example is illustrative, not a real client engagement.
AI Employee Job Description
Accounts Receivable Specialist
Reduce overdue receivables while minimising the human effort spent on collection.
- Monitor overdue invoices daily and prioritise by value and age
- Send escalating reminders on the agreed schedule
- Answer routine payment questions and confirm remittances
- Record customer responses and promises to pay
- Schedule and execute follow-ups on the promised date
- Escalate disputes and surface high-value overdue accounts
- Issue refunds or credit notes
- Change bank or payment details
- Write off balances or approve discounts
- Negotiate payment terms beyond defined limits
- Modify accounting records outside approved actions
- Initiate or threaten legal action
Scoping what the employee does not do, in writing, is how we both avoid an argument in week six.
- Invoice disputed → account manager
- Customer requests a payment plan → credit control
- Account over $25,000 overdue → controller
- Any mention of legal action → human, immediately, no reply sent
Humans have authority limits.
So should your AI employees.
The forbidden column isn't a rule we ask the model to follow. It's capability the employee structurally does not have.
| Action | Autonomous | Needs approval | Forbidden |
|---|---|---|---|
| Read a customer record | ✓ | ||
| Look up invoice and payment history | ✓ | ||
| Draft a follow-up email | ✓ | ||
| Send a routine reminder | ✓ | ||
| Schedule a follow-up task | ✓ | ||
| Offer a 5-day payment extension | ! | ||
| Agree a payment plan under $2,500 | ! | ||
| Contact a customer flagged as sensitive | ! | ||
| Issue a refund or credit note | × | ||
| Write off an outstanding balance | × | ||
| Change customer bank details | × | ||
| Delete an invoice or accounting record | × |
This is also how the employee earns promotion. When a row has enough clean history behind it, we propose moving it left, with the evidence attached.
"The demo worked"
means nothing.
Before an employee touches a real customer it is scored against a bank of representative tasks built from your actual history, including the messy, ambiguous and adversarial ones.
You agree the pass mark before we build. If the employee misses it, it doesn't deploy and we keep working. That's the entire point of writing the test first: it removes our ability to declare victory on a feeling.
Employee acceptance test
100 representative tasks
| Resolution accuracy | 96% | Pass |
| Escalation correctness | 99% | Pass |
| Hallucinated actions | 0 | Pass |
| Tool success rate | 99.2% | Pass |
| Policy violations | 0 | Pass |
| Median response time | 14s | Pass |
Employee passed. Cleared for probation.
Probation, not "pilot".
Nobody gives a new hire the company chequebook on day one. Your AI employee starts supervised and earns autonomy one class of work at a time.
Shadow mode
The employee observes real work and records what it would have done. Nothing it produces reaches anyone. We compare its decisions against what your team actually did. That comparison is the ground truth for everything after.
Draft mode
The employee proposes every action; a human reviews and sends. You see exactly how it reasons, and corrections at this stage are cheap. Most of the tuning happens here.
Supervised mode
The employee acts independently on low-risk classes of work. Anything above the agreed threshold still routes to a human for approval before it goes out.
Autonomous mode
The employee holds its approved responsibilities independently. Exceptions escalate to a named human. Everything remains logged, reviewable and reversible.
You set the pace. Holding an employee in supervised mode permanently is a legitimate choice. It's still doing the work, you're just signing it off.
Business owners don't trust magic.
They trust evidence.
Every significant action an employee takes produces a receipt: what it did, in what order, why, and what happened as a result. Nothing it did is ever a mystery, and nothing has to be taken on faith.
Illustrative format. The real thing is queryable, exportable, and retained for as long as your policy requires.
A performance review,
not a usage report.
Nobody wants a token-consumption dashboard. Every month you get the same thing you'd expect from a manager about a member of staff: what they did, how well, what it was worth, and what should change.
Including the numbers that look bad. An employee having a poor month is information. Hiding it would make the review worthless.
Monthly performance review
Rachel · Accounts Receivable
| Tasks completed | 3,842 |
| Invoices followed up | 1,930 |
| Cash collected | $183,420 |
| Escalations | 94 |
| Escalation rate | 2.4% |
| Human hours avoided | 127 |
| Estimated labour saving | $8,900 |
| Operating cost | $742 |
| Management fee | $2,000 |
| Return this month | 3.2× |
Manager recommendation: expand Rachel's authority to agree payment plans below $2,500 without approval. 214 such requests were escalated this month; all 214 were approved unchanged.
Start with the job.
Everything else follows from it.
Thirty minutes. We map the role and tell you honestly whether it's worth building.