A sample report, unredacted

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

Prepared 9 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/

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 is the one you told me you were most keen to automate.

Order is not the same as saving. Internal reporting ranks third despite a smaller saving because it is the safest place to learn what your oversight process actually 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 four hours a day on it, it feels repetitive, and the volume is steady. I understand why it looked like the obvious first candidate.

It scores 17 out of 100, and it is the one workflow here where the numbers and the risk point the same direction.

The three failure modes that would actually bite you:

• A tax or filing position stated as advice when it needed a qualified sign-off
• 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 is not a hypothetical — it is the standard failure mode for any AI with access to incoming correspondence, and it is the reason this workflow sits at delegation level 1 rather than 2.

What would change this verdict. Splitting it. Roughly 40% of what you call “client queries” is document chasing and status updates — genuinely repetitive, low judgment, low risk. Carve that out as its own workflow and it would score in the sixties. The remaining 60% is advice, and advice stays with a person. That split is the recommendation, not automation of the whole.

3. Start here

Data entry and reconciliation Liability-gated

The cost per item halves. That is the number to hold onto — the headline saving assumes you keep producing 1,162 items, and if your volume grows the saving grows with it.

Oversight contract. Delegation level 3, act and sample.

Steady state is roughly 115 items a month, about 16 hours of checking. The oversight premium already built into the $6,138 covers it with room. Step down on measured performance, never on the calendar date.

Watch for: a transaction matched to a plausible but incorrect counterparty, leaving the ledger balanced and wrong. That is the one that survives a cursory review, because nothing looks out of place. Your sample should be drawn at random, not from the items that look unusual.

Invoice processing Liability-gated

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 against.

That last one is why your sample should deliberately include invoices from your least common suppliers, not a random draw across volume. A random 8% sample is 95% routine invoices and tells you nothing about the edge.

Internal reporting Capability-gated

Run this one third, despite the modest saving, because delegation level 2 means 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.

Note the gate. This is the only capability-gated verdict in your set, which means it is the one expected to move. If the next frontier release lifts it past 65, the recommendation changes and you will get that in writing. The other four rest on accountability rather than capability, and those will hold.

Onboarding paperwork Liability-gated

4. Sequence

Do not run two of these at once. The constraint is not the technology, it is whether one person can genuinely sustain the review load during two overlapping ramp periods. They cannot.

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 tells you whether the task is worth automating, not who should build it.
• Whether the figures you gave me are right. Every number here descends from your inputs. If the 2.2 reconciliations per person-hour is optimistic, everything downstream shifts.
• Your actual error rates. The sampling schedule is calibrated from task shape, not from measurement. Your first six weeks will produce real numbers and those beat the model.
• Redundancy, retraining or management time. Not priced. These are real costs of any transition and they are not in the savings figures.
• Anything legal. This is an estimate and a template. It is not legal, financial or employment advice, and the onboarding caveat above needs a real adviser.

6. Assumptions used

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