Using AI to Support Accruals, Provisions & Estimates — Without Losing Judgment

By Christian Matumona, Kyootek AI Financial Controller

Accruals, provisions, and estimates are the part of the close where AI's strengths and its real limits show up most clearly. The calculation — applying a method to a set of inputs — is mechanical and AI handles it well. The judgment — whether that method still reflects the facts of this period — is exactly the kind of decision that should never be delegated to a model that has no way of knowing what changed. Getting this distinction right is the difference between using AI to support estimates responsibly and quietly letting it make decisions it isn't equipped to make.

This is the deep dive on accruals, provisions, and estimates referenced in the Complete Guide to AI for Financial Controllers.

The line that matters: method versus calculation

Every estimate has two parts. The method is the judgment call — how should we estimate this, given what we know this period. The calculation is mechanical — given the method and the inputs, what's the number.

AI should only ever touch the calculation. The moment a controller asks AI to decide whether last period's method is still appropriate, AI will answer fluently and confidently — and that confidence is not evidence the answer is right, because the model has no visibility into the facts that would make a method obsolete: a customer that went bankrupt, a warranty claim pattern that shifted, a litigation matter that escalated.

Where AI genuinely helps

Mechanical accrual calculations

Given a stable method and the current period's inputs, AI calculates the amount and drafts the entry — the same pattern covered in more depth in AI for Journal Entries.

Method: accrue bad debt reserve at [X]% of receivables aged over 90 days, per our standing policy.

This period's aging schedule:
[PASTE: Customer | Amount | Days Outstanding]

Calculate the required reserve using the stated method. Show the calculation by customer and the total. Flag any customer balance that individually exceeds [YOUR MATERIALITY THRESHOLD] for separate review, since large individual balances may warrant a specific assessment rather than the standard percentage.

The flag for large individual balances is the safeguard — a standard percentage method can understate risk on a single large, specifically troubled account, and that's exactly the kind of fact pattern that needs a human look rather than a formula applied uniformly.

Completeness checks against prior periods

Compare this period's accruals to the prior period's:

Prior period accruals (Account | Amount):
[PASTE]

Current period accruals (Account | Amount):
[PASTE]

Flag any account that appeared in the prior period but is absent or materially different in amount this period, without explanation. Do not assume the omission was an error — only flag it for review.

This is the same completeness-check pattern covered in the Month-End Close Playbook, applied specifically to recurring estimates — a category of entry where a quiet omission is easy to miss because nothing about its absence looks wrong on its own.

Drafting the documentation

Once an estimate is calculated and the method confirmed appropriate, AI can draft the supporting memo — useful for audit and review purposes.

Draft a brief supporting memo for this period's [ESTIMATE TYPE] provision.

Method: [STATE THE METHOD]
Inputs used: [STATE THE INPUTS]
Calculated amount: [STATE THE RESULT]
Confirmation that method remains appropriate: [STATE WHY — e.g. "no change in customer base composition or payment behavior versus prior periods"]

Write a clear, factual memo explaining the method, the inputs, the result, and the basis for continuing to use this method this period. State only the information provided above.

Notice the last input — confirmation that the method remains appropriate — is something you provide, not something AI determines. The memo documents your judgment; it doesn't substitute for it.

The estimates that should stay furthest from AI

Litigation provisions, restructuring provisions, and any estimate that depends on probability assessments of future events outside historical patterns — these depend on facts and judgment that don't reduce to a formula, and AI has no reliable way to assess them. Use AI here, if at all, only for formatting or drafting language around a number and method that a qualified professional — often alongside legal counsel — has already determined.

A simple test before using AI on any estimate

Before applying AI to an accrual, provision, or estimate, ask: is the method this period the same as last period, and do I have evidence that nothing has changed that would make the old method wrong? If yes, AI can safely handle the calculation. If you're not sure, that uncertainty is itself the signal that this estimate needs your judgment before any calculation happens — not after.

That discipline — knowing precisely which part of the work is safe to delegate and which isn't — is the practical center of the AI-Augmented Financial Controller framework: AI extends capacity for the mechanical work, and makes the judgment work more visible, not less.