There is a persistent assumption that using AI seriously at work means learning to code. For a finance professional that is the wrong entry point. The skill that pays off is the ability to break a business process into steps that can be described, checked, and handed over. That is workflow literacy, and it transfers directly from controlling.
AI workflows are not the same as programming
Programming is about writing the logic yourself. A workflow is about arranging existing components so that information moves reliably from one place to another, with a decision point somewhere in the middle. Most workflow tools are built precisely so that the person who understands the process — not the person who understands the language — can assemble it.
The comparison that lands with finance people is a month-end checklist. Nobody calls a closing checklist software, but it has the same properties: a trigger date, defined inputs, a sequence of steps, owners, exception handling, and a sign-off. An AI workflow is a closing checklist where one of the steps happens to be performed by a model.
The typical shape of a workflow
Almost every useful AI workflow, regardless of the tooling, has the same six-part structure. If you can write these six lines for a process you already own, you have done most of the specification work.
- Trigger: a new email in a shared mailbox, a file dropped in a folder, a scheduled date, a form submission
- Data: the document, the ledger extract, the request text, the previous month's figures
- AI step: exactly one narrow job — classify, extract, summarise, compare, or draft
- Rules: deterministic checks that do not involve the model — thresholds, mandatory fields, allowed values, duplicates
- Output: a fixed structure, so the result can be validated rather than merely read
- Human approval: an explicit accept, correct, or reject before anything leaves the building
The most common design mistake is asking the AI step to do four things at once. A step that classifies, extracts, judges materiality, and writes the reply is impossible to debug, because a wrong output gives you no information about which part failed. Narrow steps chained together are slower to build and far easier to trust.
Concrete examples in finance
Invoice review. Trigger: a PDF arrives in an accounts payable mailbox. AI step: extract supplier, invoice number, date, net, VAT, gross, currency, PO reference. Rules: check the PO exists, the supplier is in the master data, VAT arithmetic is consistent, and the invoice number is not a duplicate. Output: a structured row with a confidence flag. Human approval: the clerk sees only exceptions and low-confidence extractions instead of typing every field.
Management summary. Trigger: the fourth working day. Data: closed ledger, order intake, open actions from the previous month. AI step: summarise into a one-page narrative with a fixed section structure. Rules: every figure in the narrative must reconcile to a source cell, otherwise the workflow stops. Human approval: the controller edits and sends. The model saves the first ninety minutes of writing, not the judgement.
Budget explanations. Trigger: a cost centre owner opens a request form. Data: their actuals versus budget, prior explanations, project codes. AI step: draft a plausible explanation using last month's stated reasons plus the current variance. Human approval: the owner confirms or overwrites. The value is that the owner starts from a draft instead of a blank field, so responses actually arrive.
Variance comments for reporting. AI drafts, a rule blocks any comment that contains a figure not present in the source data, and a human approves. This single rule removes the failure mode that worries finance most: a fluent sentence containing an invented number.
Why human approval is not optional
A language model produces the most plausible continuation, not the verified truth. In a marketing context a plausible sentence is usually fine. In finance, a plausible sentence containing a wrong figure is worse than no sentence at all, because it is confident, well-formatted, and likely to be believed.
Human approval also solves an organisational problem, not only a quality one. It keeps accountability with a named person, gives auditors something to follow, and gives the team a mechanism for improvement: every correction is evidence about where the workflow is weak. Workflows that log corrections improve; workflows that run silently drift.
- Approve at the point where an error would become expensive or externally visible
- Show the reviewer the source next to the output, not the output alone
- Log every correction, then fix the step that produced it
- Keep a documented manual fallback for when the workflow is down
How this changes the finance role
Finance already thinks in controls, thresholds, exceptions, materiality, and audit trails. Those concepts are exactly what makes an AI workflow safe to deploy. In a design session, a person who instinctively asks who approves this and how we would detect an error six weeks later contributes more than someone who only knows the platform.
The practical consequence is that finance moves from being a consumer of systems to being a specifier of them. Start small: pick one repetitive task you personally perform, that involves text or documents, and where a mistake is visible and cheap. Map it on paper. Build a rough version with ten realistic examples and see where it breaks. The failures are the point — they show you which parts of the process genuinely depend on judgement, and those are the parts worth keeping human.