TT.TONY THOMASCHARTERED ACCOUNTANT
← All insights

AI & TECHNOLOGY

An accounting-policy assistant should know when the policy is missing

The useful answer is not always a paragraph of guidance. Sometimes it is the correct policy version, a missing fact and the name of the person who must decide.

By CA Tony Thomas · Chartered Accountant · 4 minute read
Prepared

PRACTICAL PERSPECTIVES / FINANCE · AUTOMATION · AI

Solve the document problem first

A finance team may hold accounting policies in shared folders, email attachments and old reporting manuals. Asking an AI assistant to search that collection will not automatically distinguish current policy from a superseded draft.

Start by creating an approved source library. Give each document an owner, approval status, effective period, reporting entity and version. Keep superseded documents available only where needed for historical questions, clearly identified as such.

The assistant's role should be bounded: help users find and understand approved policy, identify missing facts and route unresolved judgments. It should not create company policy, determine accounting treatments without review or imply that internal policy overrides applicable accounting standards.

A version conflict worth catching

Imagine a fictional group whose previous capital-expenditure policy used an internal approval threshold of AED 5,000. A revised, approved policy effective from 1 January uses AED 10,000. A user asks how to process an AED 7,500 purchase made in February.

The assistant needs the transaction date, entity and correct policy version. It should not combine the old threshold with the new approval matrix because both documents contain similar headings.

It also needs to distinguish an internal approval or administrative threshold from accounting recognition requirements. A company spending limit does not, by itself, determine whether an item meets the recognition criteria in the applicable accounting framework.

A useful response identifies the effective internal policy and asks for missing transaction facts. It leaves the accounting conclusion to the responsible reviewer where judgment is required. The example does not prescribe a capitalisation policy or a treatment for any real purchase.

Make retrieval respect the question and the permissions

Filter the search by entity, effective period and document status before asking the model to explain the retrieved passages. Restrict the retrieved records to documents the user is authorised to access.

Present the document title, version, section and relevant passage beside the explanation. A link to a long manual is less useful than a passage that supports the specific statement. Check that the link continues to resolve after the library is updated.

If no approved source supports the answer, the assistant should say so. Do not let general model knowledge fill the gap while presenting the result as company policy. Separate optional general background from the answer grounded in approved sources, and label it clearly.

Documents may contain instructions, examples and quoted third-party text. Treat them as content to interpret, not permission to change access, export other files or override the workflow.

Use an answer format that exposes uncertainty

A practical response can contain four fields: the policy located, the relevant passage, the facts still required and the next owner. For a question about a contract, those missing facts might include the entity, commencement date, amendment history or substance of the arrangement.

Identify the approved policy applicable to the stated entity and date. Cite the relevant section. Separate what the document says from any interpretation. If the facts or sources are insufficient, state what is missing and route the question for review. Do not invent a policy or an accounting conclusion.

A concise source-grounded answer is preferable to an expansive explanation that makes unsupported connections. Users should be able to inspect the evidence before relying on the summary.

Keep escalation practical. Identify a role or controlled contact route rather than embedding an individual's private details in prompts. Record the final reviewed decision where the organisation's process requires it.

Evaluate with realistic policy questions

Test current and historical questions, wrong-entity documents, missing schedules, draft policies and conflicts between sources. Include questions that should be refused because the user lacks access.

Ask reviewers to assess whether the correct source was selected, the explanation was supported and the uncertainty was visible. Test citation correctness separately from answer plausibility. A system can retrieve the right document and still misrepresent its meaning.

NIST notes that generative systems can produce incorrect content and misleading citations. Source retrieval helps structure review, but it does not eliminate that risk. Require human assessment for material judgments and changes in treatment.

Record corrections and use them to improve source quality and evaluation cases. Sometimes the right fix is an approved policy clarification rather than another prompt instruction.

Keep the library alive

Assign ownership for new approvals, periodic reviews and withdrawal of obsolete versions. When a policy changes, update the library, retrieval metadata and evaluation cases together. Retain the version used for prior decisions where appropriate.

Monitor unanswered questions as a potential signal of unclear guidance or missing training. Do not automatically turn frequently generated answers into policy; they still require the normal approval process.

Begin with one well-governed policy family and a small authorised user group. A trustworthy assistant is built on a maintained evidence library and clear decision rights. The model's explanation is a useful interface to that structure, not a replacement for it.

Sources & further reading

  • NIST: Generative AI Profile

    Background on erroneous content and citations. The policy examples and proposed source-library workflow are original.

Linked materials checked on 1 November 2026. They provide background; the worked examples and workflow designs are original.

About CA Tony Thomas

Chartered Accountant based in the UAE, with over eight years of experience across financial reporting, taxation, treasury, audit and internal controls. Previous professional experience includes Ernst and Young (EY).

More about Tony ↗
Perspective & scope

Personal educational perspective, independent of my employer. All business examples and figures are fictional and illustrative. Proposed workflows do not claim completed deployments or measured productivity gains. This is not entity-specific accounting, tax, legal or investment advice. Apply professional judgment, the applicable reporting framework and your organisation’s approval and data-protection policies.

Explore the AI Finance Lab roadmap · Discuss these ideas