How they learn · one piece of work, followed honestly
This page follows a single VAT reminder through an agent’s week at an accounting firm — drafted, corrected, remembered, counted. We picked the Ops & Finance candidate because deadlines make the cleanest story, but the loop is the job for all four on the roster — near the end you’ll see the same correction land on a clinic, a website and a product backlog. Every artifact below is shown the way your team would actually see it, and in the margin, quietly, what our engineers call each part.
The scene is a walkthrough, not a recording of a real client — same rule as everywhere on this site: no invented performance claims, and specimens are labelled as specimens.
Al Marwan Trading’s VAT return is due in nine days, and their bank statements still haven’t arrived. No one briefed the agent. It sits inside the firm’s shared inbox, deadline calendar and client list, so the gap is simply visible to it — the way it would be visible to your best senior, if your best senior had time to look every morning.
Ahmed’s WhatsApp · only he and his partners see this chat
Good morning. Al Marwan Trading’s VAT return is due in 9 days and we’re still missing their Jul–Aug bank statements. I’ve prepared a reminder for Mr. Khalid — may I show you?08:14
Every proposal the agent makes has the same four parts: what’s happening, why it matters, the message itself, and what happens next. Not for ceremony — so that Ahmed can approve three parts and correct one, instead of rejecting the whole thing and starting over.
VAT return due 12 Sep. Jul–Aug bank statements missing; chased once on 21 Aug, no reply.
Filing late costs the client money — AED 1,000 penalty, then 2% of unpaid tax.
“Dear Mr. Khalid, a reminder that your VAT filing is due on 12 Sep. Please note late filing carries a penalty of AED 1,000…”
Send today · follow up on the 8th if silent · flag Sara on the 10th either way.
The facts were right. The plan was right. The tone was wrong — you don’t quote penalty figures at a client of twelve years. Ahmed doesn’t write instructions or open a settings screen. He says so, the way he’d say it to a junior across the desk.
“Don’t mention penalties to Khalid — he’s been with us 12 years. Warm tone. And offer to send our driver for the documents.”08:22 ✓✓
“Dear Mr. Khalid, I hope you are well. When convenient, we still need your Jul–Aug bank statements for the VAT filing — happy to send our driver to collect them.”
Approve ✓08:23 ✓✓
Sent ✓ — and noted for good: no penalty figures for long‑standing clients unless you ask.08:23
That correction took ten seconds — and it’s the part worth feeling for yourself. Try making one in the interview →
Every firm keeps a training file on a junior; this one writes itself. Each lesson records who taught it — because who said it matters. Ahmed’s word on how to speak to clients outweighs anyone else’s. On bookkeeping detail, Sara’s might outweigh his. The agent learns your firm’s actual chain of judgment, not a flat average of everyone’s opinions.
Statement requests go out with the exact period named, never “recent statements”.
Approval recorded — deadline reminders 10 days out, chase once before escalating.
No penalty figures for long‑standing clients unless asked. Warm tone for Mr. Khalid and clients like him; offer the driver for document collection.
The following week it isn’t a VAT return — it’s Mr. Khalid’s trade‑licence renewal. Nobody repeats the instruction.
You taught it once. It stayed taught.
That single sentence is most of what we’re selling.
— The following week —
Mr. Khalid’s trade‑licence renewal is due 14 Oct. Draft — warm tone, no penalties, driver offered:09:02
“Dear Mr. Khalid, I hope the family is well. Your trade licence renewal is due on 14 Oct — shall we send our driver for the paperwork as usual?”
Approve ✓09:04 ✓✓
We followed the accounting candidate because deadlines make the cleanest story. The loop is the same job on every desk — one correction each, taught once, still there the following week:
The correction — taught once
“Never quote implant prices in chat. Invite them in, and bring Dr. Rania into the conversation.”
The following week — unprompted
A new patient asks about veneers at nine at night. The agent books the consult, names Dr. Rania — and quotes nothing.
Interview this candidate →The correction — taught once
“We never say ‘cheap’ — we say genuine value. And Ramadan hours are 9–3, not 9–4.”
The following month — unprompted
The Eid banner arrives drafted with the right hours and the right word. Nobody repeated the instruction.
Interview this candidate →The correction — taught once
“Bugs from paying customers outrank feature requests from trials.”
The next sprint — unprompted
The queue arrives in Slack already ordered that way — with the reasoning attached.
Interview this candidate →Different desks, same machinery: the four‑part draft, the ten‑second correction, the log with a name on it — and, next, the Friday count.
One number decides the probation: how much of the agent’s work your reviewers approve without touching. Every tap Ahmed makes all week — approve, correct — is already the measurement. We can’t polish it, because we don’t make it. Your partners do.
Illustrative shape only — your firm’s scorecard shows your firm’s numbers.
The probation clause reads off this card: if the approved‑untouched share hasn’t improved on its day‑one baseline by day 30, you don’t confirm the hire — and the balance of the recruitment fee is never invoiced.
The standing rules
Learning changes what the agent knows. It never changes what the agent is allowed to do. Four rules, in force from day one, in writing:
Every outgoing message waits for a named person’s tap — and for the Web candidate the same rule reads nothing goes live unapproved: every change is staged first. By construction, not by promise, and loosening it later is your decision, made only when the measured numbers have earned it.
Nothing you teach it trains anyone else’s agent, and nothing another firm teaches touches yours. A data‑processing agreement is part of every contract, on every tier. The privacy policy →
The tools it works with, and what each one is allowed to change, are set out explicitly — and anything irreversible sits behind an approval. It cannot quietly acquire new reach.
Stop any month, no notice‑period games — and the written record of everything your firm taught it leaves with you, not with us. Making that record portable is the point of our research.
For the person you’ll forward this to
Everything above is real machinery with published methods behind it. If your IT partner, your technical co‑founder, or your most skeptical colleague wants to lift the hood — this is our open research programme, in its own language:
The feedback ledger, the authority graph and the calibration record — a versioned, machine‑readable spec, portable across chat surfaces and executors.
Read the spec → the harness The loop, end to endFour‑part answers, confidence that has to mean what it says (“98 has to mean 98”), and what an executor is allowed to touch.
See the harness → open problems What we haven’t solvedMulti‑principal feedback is an open research problem and we run it live, on real teams. The unsolved parts are stated precisely, in public.
See the open problems →The week above took Ahmed three taps and one voice note. Tell us the workflow your team complains about most, and we’ll tell you honestly whether an agent can take it.
And the risk is written down: 30 days’ probation, measured on your partners’ own approvals — if the number hasn’t improved by day 30, the balance of the recruitment fee is never invoiced. The offer, in writing →
Not ready for the full hire? Start with the Practice AI Sprint → — or interview a candidate first.