There is a lot of noise about AI and tax. Some of it promises that the provision will soon run itself. Some dismisses AI because ASC 740 requires professional judgment. After nearly twenty years across CPA-firm and in-house tax roles, and now building provision automation, I see a more practical question: which work should the tool do, and what does the team need to check?
AI assistance, rules-based automation, and professional judgment serve different purposes. Treating them as interchangeable is where the trouble starts.
The provision is not one task
The provision is a chain of tasks: gathering data, mapping accounts, calculating balances, documenting positions, reconciling outputs, and preparing workpapers for review. Extracting an amount from a document is different from deciding its tax treatment. Applying an approved rate is different from determining which rate applies.
The opportunity is to reduce repetitive preparation without hiding those distinctions. I wrote about the rework behind that preparation in Why the Tax Provision Is a Fire Drill.
Where AI genuinely helps
- Extracting and organizing information. AI can pull information from documents and help organize inconsistent formats. The team checks extracted amounts, periods, and entity names against the source before using them.
- Proposing a first-pass mapping. AI can suggest account classifications from descriptions and supporting workpapers. Those suggestions are a starting point, not approved tax treatments. New accounts, ambiguous descriptions, and changes in treatment need review.
- Drafting documentation. Turning approved notes into a workpaper narrative gives the preparer something to edit. The reviewer checks the facts, sources, and reasoning rather than treating polished language as evidence.
- Helping investigate exceptions. AI can suggest explanations for an unexpected movement or organize the supporting information. The team verifies the explanation against the underlying transactions and workpapers.
These are examples of potential uses across the provision process, not a list of current Polaris features. Their value depends on the task, the source material, and the checks around the output.
Where rules-based automation fits
Defined reconciliations, arithmetic, and tie-outs do not need a generative model to decide the answer. Rules-based automation applies specified logic to approved inputs and flags differences against defined criteria. AI can help investigate a flagged item; it should not substitute for the check itself.
For provision calculations, the foundation is controlled inputs, tested calculation rules, linked formulas, and validation checks. Reviewers need to follow how a number was produced and identify the assumptions behind it.
Repeatable does not automatically mean correct. A formula can produce the same wrong answer consistently.
That is why the rules need testing, the inputs need completeness checks, and changes need review. A successful run establishes that the process ran, not that every tax conclusion is right.
Where AI should not be relied on
- Owning technical positions and conclusions. AI can help research an issue and organize supporting material. Evaluating the evidence, determining the treatment, and approving the conclusion stay with the tax team and its advisors. Sources and citations must be verified.
- Providing unverified provision numbers. A plausible answer is not calculation support. Amounts used in the provision need a traceable basis, tested calculations, and appropriate review, not just a model's response.
- Resolving novel or ambiguous facts on its own. An unfamiliar transaction or uncertain interpretation calls for technical analysis. Use the tool to organize questions, not to turn missing evidence into a confident conclusion.
- Deciding whether sensitive data is safe to share. Do not upload company or client data to an unapproved AI service. Approval depends on the actual service, terms, access, retention, and data-use settings, not whether the interface looks professional.
The controls that matter
“A human reviews it” is not enough on its own. The workflow needs to make clear what is being checked, by whom, and how exceptions are resolved.
- Check the source. Confirm extracted amounts, entities, periods, and completeness against the original documents or reports.
- Approve the mapping. Separate proposed classifications from approved mappings. Record changes and the basis for material decisions.
- Test the calculations. Check the rules and formulas against expected results, including relevant exceptions. Retest affected calculations when the logic changes.
- Verify the narrative. Check factual statements, citations, and explanations against the supporting evidence. A well-written paragraph is not a substitute for that evidence.
- Resolve and document exceptions. Keep the input version, checks performed, outstanding differences, and review decisions with the workpapers. Do not treat a completed run as automatic approval.
- Agree on data handling. Confirm permitted use, access, retention, and any AI-provider involvement before sharing confidential information. Retaining your data to run your own future work is different from using it to train a model. Ask any vendor which one applies.
What this means for Polaris
Polaris produces formula-linked Excel provision workpapers from controlled inputs and approved mappings, with validation and tie-out checks. The provision calculations are not delegated to a generative AI model. Technical judgment and final review stay with the team.
The goal is practical: less repetitive preparation, with workpapers the reviewer can follow. Whether a particular step calls for AI assistance, conventional automation, or an experienced professional depends on the work itself.
For any proposed engagement, the conversation about company data comes before the upload: what information is needed, where it will be processed, who has access, and what will be retained. Those terms need to be explicit, not inferred from the word “AI.”