Every desk drafts to your standard.

Career journalists interview your people. What they say becomes your rules, your review standards and your voice — written into one shared AI workspace per department, inside the seats you already pay for.

We call each one a room. Your own lead holds the keys in ninety days. Social Security, EIN and bank numbers never go in. Your firm’s AI policy decides the rest.

NEVER REQUEST, AND REFUSE IF OFFERED: Social Security numbers, EINs, bank and card numbers, passwords.

Everything else follows your firm’s own AI policy.

Book 20 minutes

Peer-reviewedAI and Ethics, Springer Nature, 2026

Patent pendingDeterministic governance runtime

Inside the models18 months evaluating Google’s Bard and Gemini pre-release

Executive AI coachingCertified coach, AI Advantage

Where every firm is

Everyone is already using it. Nobody agreed on the rules.

One partner spent six months tuning it to his own voice. Nobody else got any of it.

The policy is signed. It never reached a single desk.

You cannot say what anyone pasted into it last week.

What actually changes

A Tuesday in your AR queue.

About thirty follow-ups a week at the firm we are building for now. Today each one gets opened, a template gets found, invoice lines get retyped, tone gets adjusted for a client having a hard year.

You: here's this week's AR list — let's work through it

Room: two questions that apply to everyone. Phone contact attempted? Invoices already resent?

— then down the list, asking only what is specific to each client.

Drafts come back with the balance stated plainly, the payment path obvious, both office numbers, and the resent-for-your-convenience line. A person reads each one and sends it. The room never sends anything.

Volumes are that firm’s own, stated in discovery. We do not claim measured time savings; a timed test runs before the engagement closes and the result posts here.

Where the prompts come from

Your prompt library is written out of your own interviews.

Nobody hands you a generic prompt pack. Every prompt in every room is built from what your people said about how the work actually gets done — including the standard they hold in their heads and have never written down anywhere.

And nobody has to open it. Your staff say what they need in plain words. The room recognizes the job and runs the right prompt, because it has already read the whole library.

WHAT THEY TYPE → WHAT THE ROOM DOES

“The numbers moved since we reported.”
→ checks the financials are closed and finalized, then runs the affected line as a percentage of income and shows the math

“Here’s this week’s AR list — let’s work through it.”
→ asks the two questions that apply to the whole batch, then works down the list asking only what is specific to each client

“Client got a notice for 2023.”
→ asks the six things that change a notice response, then builds the skeleton with every taxpayer-specific fact left as a bracket

“hi”
→ a short menu. Pick a number.

DISCOVERY · we asked the admin lead what the room
should already know, so she never has to explain it twice

“The tone of each email needs to be friendly yet professional, approachable and understanding due to any hardships a client may have. The client needs to know we’re here to help, but each email should be clear and to the point.”

INSTALLED · the check the room now runs on itself
before it hands her a draft

States the outstanding balance clearly.
Makes the payment path convenient and clear.
Offers a way to discuss alternatives — without proposing one.
Professional, not adversarial.
Reflects that the client may be facing hardship.
Concise.

DO NOT PROPOSE A PAYMENT PLAN. That decision is Admin’s judgment, case by case.

That last rule is there because she told us the firm has no formula for it — the call depends on what she knows about how a client pays. The room stops exactly where her judgment starts. A template vendor cannot write that line, because nobody told them.

Delivered as a working file inside the room rather than a PDF nobody opens, so the firm can see and change what the room runs on. It is versioned. In bookkeeping, v2 dropped the question sequence for a single paste, and added the one prompt the department head said he most wished he could just ask for. In admin, v2 added a batch mode, because one client at a time turned out to be the wrong shape for a thirty-client list.

What sits in the firm when we leave

Nine things, all of them yours.

One room per department, one firm-wide

Shared files, shared instructions, and memory that stays inside the room.

A prompt library per department

Eight named categories, written from the interviews, versioned as your people send back corrections.

A Room Guide for each

One page: what it does, what it refuses to do, what it cannot reach, who reviews what.

Your voice guide

Written from your partner’s answers and your own best client emails. Pinned in every room.

The Three Lines

Task, data, review — per department, in your people’s own words. Your carrier can read them.

Three or four custom tools

Chosen by volume from the interviews, not by whoever was most enthusiastic in the room.

A permission review

Folder by folder with your IT contact before anything is connected.

Three recorded trainings

Management, firm-wide, and a follow-up built from your staff’s own first-week questions.

A named owner

Your AI lead, the admin console, and a one-page sheet for adding a seat after we are gone.

Built on ChatGPT or Claude, inside your own subscription. Every account, setting, room and document is yours from day one. Nothing to migrate, nothing to renew, no dependency on us.

Client data

The numbers that identify a client never enter what we build.

Social Security numbers, EINs, bank and card numbers: not during discovery, not during the build, not after handoff. The rooms are configured to refuse them, and the refusal is written where your staff can read it. If someone pastes what they shouldn't, the room stops and names what has to come out before it will continue.

Most vendors answer this with a clause promising not to train on your data. Never receiving them is a stronger promise than not misusing them.

It starts before we are hired. Our own discovery emails say it in writing: do not send returns, taxpayer-identifying information, account numbers, credentials, or notices. Describe the process and the category of information instead. If an example cannot be safely de-identified, describe it rather than sending it.

For your own counsel to weigh: the auxiliary-services provision at 26 CFR §301.7216-1(b)(2)(i)(B), distinct from the definition of disclosure at (b)(5). Nothing here is legal advice.

Method: Smith, C.B. & McCarthy, D. Deterministic governance for generative systems: a policy-aligned runtime for validating AI outputs. AI and Ethics 6, 394 (2026).

Review

The room drafts. A person signs.

A CPA letter, or anything going to the IRS under a partner’s name, routes to that partner before it moves. Routine work goes out when the person doing it has checked the number. The rooms know which is which, because your people told us.

PREPARED BY the room · first draft

REVIEWED BY ____________________

DATE ____________________

— installed at the top of every output

See one

Every room comes with one page explaining it.

What it does. What it refuses to do, on purpose. What it cannot reach. Who reviews what. Written for the person using it, not for the person buying it.

What this is / How to start / What it can do / What it won’t do,
on purpose / What it can't reach / Who reviews what /
When you're not sure / What's still coming

Take a real one. Fully anonymized, from a live engagement, exactly as it was delivered.

Ask for a Room Guide

How it gets built

We interview the people who do the work.

Career journalists, several thousand interviews between us. We are not better at prompting than your smartest associate. We are better at getting people to describe what they actually do.

One question at a time, at their own desk

About twenty minutes. It follows up on the step they skipped, because the boring middle is where the decision nobody notices lives. It refuses regulated data on the spot.

We read it back and let them correct it

Nobody signs off on a description of their job they did not get to fix.

We grade it before anything gets built

An 18-point sufficiency gate, then a 26-point quality read. The delivered documents name their own gaps, which is why they read like a workpaper rather than a brochure.

We install it, train your staff, hand over the keys

At day 90 your own lead builds a new room without us watching. If they cannot, we are not finished.

EXCERPT · the rules the interview runs under

Ask ONE question at a time. Wait for my answer before asking the next.
When a question has multiple parts, ask the parts one at a time — never as a wall of questions.
SKIP RULE: if I type “skip,” move on immediately. No comment, no persuasion. Log it as skipped.

DATA RULE, enforce it actively. I should not share client names, tax IDs, account numbers, credentials, or live client financial data.

1 hryour AI lead

20 mineach department lead

30 minone partner

0group meetings

0homework

Built on ChatGPT or Claude, inside your own subscription. Every account, setting, room and document is yours from day one. There is nothing to migrate and nothing to renew.

In progress

A working accounting firm.

5departments interviewed

6rooms

90days, fixed fee

A strong answer cites an IRC code section or regulation that supports the position… OUTPUT IS FOR THE FILE, NEVER FOR THE CLIENT.

— tax room quality bar, anonymized

REQUIRES REVIEW BY THE PARTNER RESPONSIBLE FOR THE CLIENT BEFORE SENDING.

— admin room output header, anonymized

IT WILL NOT HAND YOU A CITATION YOU CAN RELY ON. Every authority carries VERIFY BEFORE USE, regardless of how confident it sounds. It cannot tell its wrong answers from its right ones, and a fabricated authority can reach a return.

— tax room guide, anonymized

What the rooms cannot do, said out loud. They cannot reach QuickBooks, CCH, the drives, Outlook or Teams. They cannot look up an invoice, check whether ACH is on file, or open last period. They will say so rather than invent a number. They draft, and they refuse. A person reviews everything.

Discovery complete across five departments. The timed result posts here when the engagement closes.

Timing

Start by October and it is in use before January.

Start in January and it waits until May, because nobody at your firm learns anything new during busy season. If the honest answer is May, that is a real answer, and we would rather hear it now than sit in your inbox until spring.

Who builds it

Two people who interview for a living.

C. Brian Smith

Co-founder

Career journalist. Eighteen months evaluating Google’s Bard and Gemini models before release, which is why the platform recommendation you get is not a reseller's. Lead author of the peer-reviewed governance paper cited above. Certified AI Performance Coach, AI Advantage.

brian@theprompttank.com

Daniel McCarthy

Co-founder

Career journalist, and a sitting VP of Marketing running this same method inside a company with a P&L, a board and a warehouse. That is how we know it survives contact with a business that has to hit numbers. Co-author of the paper.

dan@theprompttank.com

We build for accounting and advisory firms first, because that is where our rules and our references are. Law firms, insurance and title agencies run on the same review-and-privilege logic and we take those engagements. Ask on the call which of our references is closest to your work.

Partners ask

Six questions, every time.

What if someone pastes a Social Security number anyway?

The room stops, tells them exactly what has to come out, and will not proceed until it does. That behavior is installed, not hoped for, and your staff can read the rule that produces it.

What if it makes something up?

It will, and the rooms are built on that assumption rather than against it. In the tax room every authority it produces carries a verify-before-use flag no matter how confident the wording sounds, because the model cannot tell its wrong answers from its right ones. In bookkeeping it runs the percentage and refuses to tell you whether something is material, because that threshold was never written down at the firm. The rooms are designed to be checkable, not to be trusted.

We already pay for ChatGPT. Why is this not redundant?

That is why it is inexpensive. We are not selling you software. We are putting your firm’s rules inside seats you already own, which right now run on whatever each person happens to type.

My people won’t use it.

Usually true, and it is why we build from interviews instead of templates. There is nothing to learn and nothing to open — your staff describe what they need the way they would to a colleague, and the room runs the matching prompt itself. If someone is unsure, they type “hi” and pick from a menu. Each department also gets a one-page guide. Total staff time across the whole build is about twenty minutes each.

Why interview inside the AI instead of sending a questionnaire?

A written answer to “walk me through it” is three sentences that skip the middle, and the middle is where the decision they make without noticing lives. The self-interview asks one question at a time and follows up on the step that got compressed.

What does a peer-reviewed paper have to do with my firm?

It is why the rules layer is built the way it is: a rule that cannot be pointed at, versioned and reviewed is not a rule. Nothing in the paper is being sold to you.

What happens at day 90?

Your own lead builds a room without us watching. Everything already sits in your subscriptions under your name, so there is nothing to hand back and nothing to renew.