The idea in one paragraph
A self-improving finance function is a team whose AI workflows get better every cycle, because the team feeds its own best output back into the prompts and context the AI works from. Each month-end produces fresh examples of good commentary and clean reconciliations; those become the reference material for the next month, so the first draft needs less editing each time. The saving compounds instead of staying flat. The loop is a discipline, not a piece of software.
What a Self-Improving Finance Function Is
A self-improving finance function is a finance team whose AI-assisted workflows get measurably better with each cycle, because the team deliberately captures its own best work and feeds it back into the context the AI uses. The model does not learn on its own. The team makes it learn, by treating every reviewer-approved output as training material for the next round.
The mechanism is simple. When a financial controller edits an AI-drafted variance commentary into its final, sign-off-ready form, that final version is the single best example of what good looks like in your house style. If it goes back into the workflow as a reference, the next draft starts closer to the finish line. Do that across every recurring artefact and the function quietly gets better at producing its own work.
Flat Savings vs Compounding Savings
This is the distinction that matters. Most finance teams that adopt AI set up a workflow, get a useful one-off time saving, and stop there. The AI is no better in month twelve than it was in month one, because nothing fed back into it. The saving is real but flat.
A self-improving function turns that flat line into a curve. In month one the AI saves a controller perhaps a third of the time on commentary. By month six, with the best past examples in context and the prompts refined around the edits people kept making, the first draft is closer to final and the saving is larger. The output also gets more consistent, because the reference set anchors the model to your standard rather than a generic one. Same tools, very different trajectory.
Want help turning a flat AI saving into a compounding one?
Build the loop with usThe Feedback Loop in Four Steps
1. Draft. The AI produces the first version of the artefact, commentary, reconciliation note, board narrative, from the data and the current context. This is the step most teams already have, usually built in Claude Projects.
2. Review and edit. A qualified person edits the draft to final, sign-off quality. Nothing new here either; this is the human control point that should always exist.
3. Capture. This is the step almost everyone skips. The final, approved version is saved back into the project context or a curated example library, tagged as a strong example. Crucially, you capture the reviewer's version, not the AI's first draft.
4. Refine. Once a month the team looks at where the most editing happened and adjusts the prompt and example set accordingly. If reviewers keep softening the same overconfident phrasing, the prompt gets a rule about it. The loop closes, and the next cycle starts from a better baseline.
A Worked Example
A 50-person accountancy practice in Leeds ran its clients' management commentary through AI but kept the setup static for a quarter. The partner saved time, but each month felt like starting over: the same generic phrasing, the same edits across every client file. The fix was not a better model. It was adding steps three and four.
They began saving each finalised commentary back into the Claude Project as a reference, and spent twenty minutes at each month-end noting the edits that recurred. Within three cycles the first drafts were arriving in the practice's own voice, citing the drivers each client's board actually asks about, and the partner's editing time on commentary fell from around two hours to under forty minutes. The model did not change. The loop did the work.
What to Capture, and What Not To
Capture the finished artefacts that represent your standard: approved commentary, model reconciliation notes, the board narrative the CFO signed off, the reviewer questions that proved most useful. These are the examples that teach the workflow your house style and your reviewers' priorities.
Do not capture raw financial data, anything containing personal or sensitive information beyond what the workflow needs, or unreviewed AI drafts. The point of the loop is to reinforce good final output, and feeding back unverified drafts does the opposite. Keep the same governance discipline described in our AI in finance guide: a person reviews everything, nothing sensitive goes where it should not, and the ledger remains the source of truth.
Free 30-minute consultation
Map your AI opportunities
Book a free 30-minute consultation. We’ll review your workflows, identify the highest-ROI AI opportunities, and tell you whether you need consulting, training, or a fractional CAIO. No pitch, just direction.
Book a 30-minute callGetting Started
Start with one artefact, not the whole function. Pick the single most repetitive written output your team produces, usually month-end commentary, and add only steps three and four to the workflow you already have. Capture the finalised version each cycle and hold a short monthly refinement review. Give it three months and measure the editing time against where you started. Once the loop proves itself on one artefact, extend it to the next.
Building this well is most of what our AI consulting work involves: not just standing up AI workflows but designing the feedback loop that makes them compound. An AI readiness assessment identifies the highest-value artefacts to start with, and our AI for finance training teaches the team to run the loop themselves.