Human AI Judgment
Work with me

Job vs AI · Finance

Accountant

Rule-bound work at volume, wrapped around judgment that a regulator assumes a person made. The first half automates well. The second half is the job.

  • reporting
  • reconciliation
  • audit
  • compliance
Get the playbook

Most of what looks automatable in accounting was automatable before AI existed. Knowing which half you are looking at is the skill.

The problem

Start with the problem, not the tool

You already know accounting. The open question is where technology belongs in it.

Accounting attracts AI pitches because so much of the work looks repetitive from the outside. Some of it genuinely is. But the repetitive parts were automatable long before generative AI existed, and the parts that were not automatable are not automatable now for reasons that have nothing to do with model quality.

"Which of my tasks can AI do?" leads nowhere. It starts from a capability and hunts for a use. The question that works is narrower: at this specific step, what am I trying to accomplish, which part needs judgment, and what has to be checked?

Order matters. Problem, then workflow, then decision, then technology, then verification. In this profession the last step is often not yours to design - a standard, a regulator or the audit file has already specified it.

Lane 01

Automate

Rule-shaped work at volume. Accounting has more of this than most professions.

  • What belongs here

    Transaction coding against a stable chart of accounts, matching invoices to orders and receipts within a stated tolerance, reconciliation matching, exception listing, and assembling reports from data that already exists.

  • What it still costs you

    Writing the rules down, approving the treatment of each recurring counterparty, deciding what counts as an exception, and sampling the output against source documents. Automated output is never sign-off.

  • Why this rarely means generative AI

    If the rule is stable and writable, conventional software applies it identically every time. A language model does not offer that guarantee, and reproducibility is exactly what a ledger needs.

Lane 02

Augment

A person still decides; the machine drafts, searches or organises.

  • What belongs here

    Drafting management commentary from numbers you supply, researching candidate treatments for an unusual transaction, summarising working papers, and preparing forecast scenarios for you to weigh.

  • What stays yours

    The causal claim. A model has the numbers, not the reason the plant was down for a week. Draft with it; never let it supply the cause of a variance.

  • Every citation is unchecked until you open it

    Technical research is the shape of task that invites fabricated references: a specific paragraph number, quoted confidently, that does not say that. Use the tool to find candidates, then do the reading.

Lane 03

Human

Judgment under a code of ethics, and the name against the number.

  • What belongs here

    The accounting treatment of a transaction the standards do not settle cleanly, materiality decisions, professional opinion, and anything an independence or ethics requirement constrains.

  • What AI can still do around it

    Organise the facts, lay out the options and the precedent, model the effect of each treatment on the statements, and draft the question you take to someone more senior. None of that is the decision.

Lane 04

Verify

The check that makes the other three safe - and here, mostly one you did not design.

  • What belongs here

    Reconciliation to source evidence, the audit trail, records retention, and the conclusion on whether evidence is sufficient and appropriate to support what is reported.

  • Imposed, not designed

    In most professions you design the verification. Here the first move is to find out what is already required by standard, regulator or the file, and design inside it. Proposing a lighter check than the standard demands is not a trade-off; it is a defect.

  • What counts as evidence

    Sufficiency is quantity; appropriateness is relevance and reliability. An AI assertion pointing at nothing you can open is not evidence, and ten AI summaries of one document are still one document.

Situations

Six situations, worked

Six representative situations, chosen to show different lane mixtures. The counts are of these six examples - not a measurement of the profession.

  • 01Automate2examples
  • 02Augment2examples
  • 03Human1example
  • 04Verify1example

  • Coding transactions to the chart of accounts

    Automate
    continuousStable rule, checkable output, and a solved problem long before generative AI. A rules engine, not a model.

  • Reconciling a control account at month-end

    Automate
    monthlyMatching is arithmetic against a stated tolerance. The judgment is in what the remaining differences mean.

  • Explaining a variance for the management pack

    Augment
    monthlyIt can assemble the movement and draft the wording. The causal claim stays with the person who knows the business.

  • Researching the treatment of an unusual transaction

    Augment
    as requiredUseful for finding candidate standards. Every citation is unchecked until you open the paragraph yourself.

  • Deciding materiality and the accounting treatment

    Human
    per judgmentA regulator assumes a person made this call, and no tool changes who that person is.

  • Forming the conclusion on the financial statements

    Verify
    per reporting cycleAssembling and indexing the evidence is automatable. Deciding it is sufficient and appropriate is not.

Solutions

Choose the kind of solution, then the tool

Most accounting problems are not generative-AI problems. Naming the class is the decision.

  • Conventional software and integration

    The rule is known and exact: coding, allocations, recurring journals, posting through an API.

    For example: Xero Accounting API, QuickBooks Online API

    Watch for: Asking a model to reproduce a rule you could write is the tell that you picked wrong.

  • Deterministic matching in the ERP

    Compare invoice, order and receipt against a stated tolerance and hold what falls outside it.

    For example: Dynamics 365 accounts payable matching

    Watch for: Often a feature you already own and have not switched on.

  • Workflow automation

    Move data between systems on a schedule, and chase what has not arrived.

    For example: Power Automate and comparable tools

    Watch for: Automating a broken process makes it fail faster, not better.

  • Generative AI

    Language is the deliverable: drafting commentary, summarising papers, extracting from documents.

    Watch for: It must never supply a cause or a citation you have not verified, and reproducibility is not on offer.

  • No technology at all

    Treatment, materiality, opinion - and sometimes a process change, or deleting the step.

    Watch for: Before automating a reconciliation, ask why it produces so many differences.

Chatbot → agent

From chatbot to agent

The same task at four levels of support. Most accounting work should stop at the second.

  1. Ask a chatbot

    You paste the variance table in and it drafts commentary. Genuinely useful, and where most people should start.

    Augment

  2. Save the prompt

    The same instructions every month, so the output stops drifting. This is where most accounting tasks should stop.

    Augment

  3. Give it context and tools

    It reads the ledger extract and last month's pack rather than whatever you remembered to paste in.

    Augment

  4. Make it an agent

    Multi-step, branching on what it finds, with a fixed output contract and an explicit escalation. Worth it only when the task repeats, needs live access, and can be checked mechanically.

    Verify

Blueprint

What an accounting agent looks like

One worked example, through the same ten components any profession would use. The playbook carries three.

  • Goal

    Produce the exception list for one account and one period. Narrow enough that you can tell whether it succeeded.

  • Rules

    Never adjust a figure. Never net off unexplained differences. Every exception cites both sides of the difference.

  • Tools

    Ledger read-only, the bank or sub-ledger feed, and a matching function with a stated tolerance. Read-only is a design decision, not a limitation.

  • Memory

    Recurring counterparties, known timing differences and prior-period explanations - the context you would otherwise re-supply every month.

  • Output contract

    One row per exception: ID, both source rows, difference, age, proposed category, confidence. A fixed shape is what makes review mechanical instead of a re-read.

  • Verification

    Every row opens to both source records, and the control account still has to reconcile in total. The check is the same one you already owe the file.

  • Escalation

    Anything that would change a reported figure, and any difference it cannot categorise. An agent with no defined refusal will answer confidently in exactly the case where it should have stopped.

Risks

Where this goes wrong

The familiar failure modes, plus two that are professional-conduct matters before they are technology matters.

  • Fabricated citation

    High risk

    A standard and paragraph number, quoted confidently, that does not say what the output claims it says.

    What to do:Open every standard cited and read the paragraph. A reference you have not read is not a reference.

  • Invented cause

    High risk

    A plausible explanation for a variance the model has no way of knowing the reason for.

    What to do:Let the tool describe the movement and leave the cause blank. You supply causes.

  • Automation bias

    High risk

    A fluent pack discourages the independent look that would have caught the error.

    What to do:Match the strength of review to the consequence of being wrong, not to how convincing the output looks.

  • Confidentiality breach

    High risk

    Client or employee data placed in a service your professional obligations do not permit - correctly processed, and still a breach.

    What to do:Establish which tools you may put the data into before comparing what they can do.

  • Silent misclassification

    Medium risk

    Coding that is wrong the same way across thousands of transactions, inside a process others rely on.

    What to do:Sample against source documents and reconcile the control account; volume makes a small error rate expensive.

Sources

Research basis

What the profession framing and the risk principles rest on. The lane assignments above are Human AI Judgment analysis, not a classification made by any of these organisations.

+Playbook

The full analysis, in a form you can take into work.

The Practical AI Judgment Playbook: Accounting

Six representative situations, the reasoning behind each lane, three agent blueprints, and the failure modes worth knowing before you trust an output.

Cover of The Practical AI Judgment Playbook: Accounting

In your own words. This decides which profession I analyse next.

Optional.

Optional. You will receive the requested playbook whether or not you opt in — this box only controls future emails.

Keep going

Other professions

The same structure, applied to different work.

  • Sales

    Account Manager

    The only profession in this series where a completely correct output can still be the thing that costs you the account.

  • Delivery

    Project Manager

    The profession whose actual deliverable is information other people act on - which makes it the one where faster reporting can quietly make the job worse.