A sample report, unredacted

This is the paid deliverable in full, prepared for Meridian Accounting, a fictional fifteen-person firm. It uses real arithmetic and real oversight schedules, and it recommends against automating one workflow.

Order a report for your own task list — $150 →

What the report adds beyond the free tool

The free audit scores one task at a time. The report looks at your whole list together, which is where most of the useful decisions are:

• The order to go in, based on risk and what you’ll learn, not just on size of saving
• What not to run at the same time, because your review capacity has limits
• Where to split a task, when part of it should be automated and part should stay human
• Where the law or your contracts may override the score
• Mapping your real tasks to the right reference tasks, which is where most self-run audits go wrong

In this sample, the most valuable finding isn’t a number. It’s section 2: why client queries should stay human, and how to split off the part that shouldn’t.

About this sample

This sample shows the essentials every report covers: verdicts, the order to automate in, oversight schedules, failure modes and what the audit can’t tell you.

Your own report goes further. It’s built around your real numbers and your real situation, so it’s more specific than a fictional firm can be. The more you tell us about what’s causing problems right now, such as a backlog, errors that keep slipping through, a task that eats your team’s week, or a tool that isn’t working, the more closely the report focuses on those problems.

---

SAMPLE: Meridian Accounting is a fictional firm, but the arithmetic is real. This is the deliverable in full, unredacted.

Oversight audit: Meridian Accounting

Fifteen staff · five workflows assessed · what to automate, in what order, and what to keep human

Prepared 26 September 2026 Framework v4.2 · last recalibrated 9 September 2026 against GPT-6 Astra Every figure below comes from the public model at aiworkaudit.net/how-it-works/

Combined monthly saving, at today’s output: $13,032

This covers the three workflows this audit recommends automating, plus one to run as a hybrid. It means producing the same volume you produce now, with fewer people on it, and it already includes platform fees and setup costs.

One of the five is recommended against. On paper, automating it would save about $1,011 a month, and this audit tells you to leave that money on the table. That recommendation is the most important page in this document.

1. The short answer

Three of your five workflows clear the bar. One is worth running as a hybrid. One should stay human, and it’s the one you told me you most wanted to automate.

Order isn’t the same as saving. Internal reporting goes first, despite one of the smallest savings, because it’s the safest place to learn what your oversight process really looks like before you point it at the ledger.

2. What not to automate, and why

Client queries · Liability-gated

You flagged this as the priority. Three people spend five hours a day on it, it feels repetitive, and the volume is steady. So I understand why it looked like the obvious first candidate.

It scores 17 out of 100.

On paper, it saves money. That isn’t the problem. The problem is what it would take to trust it.

At high judgment and serious risk, you’d review every AI-drafted reply for the first two weeks, and 95% for the next four. After that, the schedule settles at 38% and stays there. With two people on the hybrid, that’s about 204 replies and 111 hours of close reading every month.

Even then, the sample can’t prove the process works. To show the error rate is below the 1% ceiling, you need about 300 reviewed replies a month, and you’d have 204. So you’d be paying for a heavy supervision load that still couldn’t tell you whether it was catching the mistakes that matter.

A better model won’t fix this. The thing you’d be automating is the judgment, and the judgment is what clients pay you for.

The three failure modes that would actually bite you:

• A tax or filing position stated as advice when it needed a qualified sign-off
• A tone mismatch on a complaint or a sensitive thread
• Acting on an instruction embedded in the incoming message itself

The third one deserves a sentence. If a client email contains text that reads as an instruction, an agent with mailbox access may act on it. That isn’t hypothetical. It’s the standard failure mode for any AI that can read incoming correspondence, and it’s why this workflow sits at delegation level 1 with read-only access.

What would change this verdict: splitting it. Roughly 40% of what you call “client queries” is document chasing and status updates. That part is genuinely repetitive, low in judgment and low in risk, so it should score far higher on its own. Run it through the audit as a separate workflow to confirm. The remaining 60% is advice, and advice stays with a person. So the recommendation is that split, not automating the whole thing.

3. The four workflows to automate, in order

1st: Internal reporting · Capability-gated

Run this one first, despite the modest saving. At delegation level 2, a person approves every report before it goes anywhere. That makes it the cheapest possible place to find out whether your review process works, on a workflow where a mistake is embarrassing rather than expensive.

What approval means here. Approving a report is a read-through. The sampling rate is different: it’s the share you check in full against the source data. At steady state, that’s about 12 reports and 9 hours a month. The oversight premium already budgets about 28.

Sample adequacy. Twelve reports a month is too few to prove the 3% error ceiling, which needs about 100. At level 2 this matters less, because nothing leaves without approval. Still, treat the full checks as the real test, and expect it to take months, not weeks, to build real evidence.

Watch for: a figure carried forward from a prior period without being refreshed.

Note the gate. This is the only capability-gated verdict in your set, so it’s the one expected to move. Standardising the task would also move it: raising repeatability from 78% to 93% would lift the score to 67, into Automate. If a future frontier release moves it instead, the changelog will record it. The other four rest on accountability rather than capability, and those will hold.

2nd: Data entry and reconciliation · Liability-gated

The cost per item nearly halves. That’s the number to hold onto. The headline saving assumes you keep producing 1,540 items. If you keep three people on it instead, output rises to about 2,295 items a month at $4.92 each, and the oversight schedule below is sized for that.

Oversight contract: delegation level 3, act and sample.

At steady state, that’s about 186 items a month, or 28 hours of checking. The oversight premium budgets about 92 hours, so it covers this with plenty of room. It’s also a large enough sample to detect a 3% error rate. Always step down on measured performance, never on the calendar date.

Watch for: a transaction matched to a plausible but incorrect counterparty, which leaves the ledger balanced and wrong. That’s the one that survives a quick review, because nothing looks out of place. So draw your sample at random, not from the items that look unusual.

Liability-gated. If a reconciliation is wrong, the exposure sits with Meridian. No model provider’s terms transfer it, and none indemnify you for the output. Keep the audit log, and plan for six months minimum.

3rd: Invoice processing · Liability-gated

With two people kept on, output rises to about 1,815 invoices a month. At steady state, you’d check about 148 of them, roughly 16 hours. The premium budgets about 53, and the sample is large enough to detect a 3% error rate.

Watch for:

• Tax treatment or currency carried over from a similar invoice, rather than read from this one
• Fields left blank rather than flagged when the source is ambiguous
• Confident output on a document type the process was never tested on

That last one is why your sample should deliberately include invoices from your least common suppliers. A random 8% sample is mostly routine invoices, so it tells you little about the edge cases.

4th: Onboarding paperwork · Liability-gated

One caveat outranks the score. The main failure mode here is right-to-work or tax documentation accepted without the checks the law requires. Some of those checks may have to be done by a person as a matter of law, whatever this score says. So confirm that with your own adviser before you move this to level 3. If any part of it must be human-verified, carve that step out and automate only the rest.

Sample adequacy. At steady state, 8% of about 192 files is 16 a month. Proving a 3% error ceiling takes about 100 reviewed files, so at this volume the sample can’t prove it. Treat the monthly check as an early warning, not as proof. The simplest fix is also the legal one: keep a person verifying every right-to-work and tax check, and sample only the rest.

4. Sequence

Don’t run two of these ramps at once. The constraint isn’t the technology. It’s whether one person can really sustain the review load across two overlapping ramp periods, and they can’t.

5. What this audit does not tell you

• Which vendor to use. Tool selection is a separate exercise, and the answer changes every quarter. This audit tells you whether a task is worth automating, not who should build it.
• Whether the figures you gave me are right. Every number here comes from your inputs. If 2.0 reconciliations per person-hour is optimistic, everything downstream shifts.
• Your actual error rates. The sampling schedule comes from task shape, not measurement. Your first six weeks will produce real numbers, and those beat the model.
• Redundancy, retraining or management time. These are real costs of any transition, but they aren’t priced here and they aren’t in the savings figures.
• Anything legal. This is an estimate and a template, not legal, financial or employment advice. The onboarding caveat above needs a real adviser.

6. Assumptions used

Salaries are loaded, full-time equivalent. Setup costs are spread over 12 months, and every saving above already includes them. The model assumes 22 working days a month.

The model used is mid-tier. Across all five workflows, the gap between the cheapest and most expensive model is under 1% of the monthly figure, so model choice changes no verdict here.

Every formula behind these numbers is published, not proprietary. If you disagree with a parameter, you can check it and argue with it, and that’s the point.

Meridian Accounting is fictional. The arithmetic, parameters and oversight schedules are real, generated by the framework at v4.2.

Parameters (v4.2.json, CC BY 4.0) · Method · Changelog

An estimate and a template, not legal, financial or employment advice.

Order a report for your own task list — $150 →