There is a lot of noise right now about AI and tax. Some of it promises that the provision will soon run itself. Some of it insists that AI has no business anywhere near something as technical as ASC 740. After nearly twenty years across CPA firm and in-house tax roles, with the most recent stretch spent building automation for the provision, my honest view sits in between. AI is genuinely useful for parts of the provision, and genuinely a bad idea for others. The real skill is knowing which is which.

The provision is not one task

It helps to remember what the provision actually is. It is not a single job, it is a chain of them: gathering data, mapping it, running calculations, documenting positions, tying everything out, and preparing workpapers a reviewer can trust. Some links in that chain are mechanical and repetitive. Others require professional judgment and carry real consequences when they are wrong. AI belongs in the first group. It should stay out of the second.

If you have lived through the manual version of this, the pain is familiar: fragile spreadsheets, constant rework, and tax on the critical path at the worst possible moment. I wrote about why that happens in why the year-end tax provision becomes a fire drill.

Where AI genuinely helps

These are the parts of the work where AI can take real weight off your team, because the human stays in control of the result:

  • Wrangling and standardizing data. Pulling information out of documents, normalizing inconsistent formats, and suggesting a first-pass account mapping that the team then confirms. Mapping itself can involve real tax judgment, so the point is not to hand it off. It is to take the tedious setup off your plate so a person can focus on getting the mapping right.
  • Drafting and documentation. Turning your notes into a first draft of a memo or a workpaper narrative, so you are editing something rather than starting from a blank page.
  • Surfacing inconsistencies. Flagging numbers that do not tie, balances that moved unexpectedly, or items that look out of pattern, so a person knows where to look first.
  • Repetitive reconciliations and checks. Running the tedious, rules-based comparisons quickly and the same way every time, without the fatigue that creeps in at hour eleven.

The common thread is that AI reduces the manual effort while a person still owns the outcome. Nothing here decides a tax answer. It clears the path so a professional can get to the answer faster.

It is worth being precise about one point, because sophisticated provision teams will ask it: AI assists with data extraction, first-pass mapping, drafting, and triage. The provision calculations themselves should be produced through controlled inputs, deterministic rules, linked formulas, and validation checks that a reviewer can trace back to source data. Polaris is not asking an AI model to calculate ASC 740.

Where AI does not belong

Just as important is being honest about where it should not be used:

  • Technical positions and conclusions. Whether a position is supportable, how a specific item should be treated under ASC 740, what the right answer actually is. That is professional judgment, and it stays with your team and your advisors.
  • Anything that must be exactly right and defensible. A tool that is usually right is not good enough for numbers that feed audited financial statements. Provision work has to be reliable and auditable, not probabilistically close.
  • Novel or ambiguous situations. AI is weakest precisely where the law is unsettled or the facts are unusual, which is exactly where you most need an experienced human making the call.
  • Confidentiality-sensitive steps. Sensitive company or client financial data should never be pasted into a public AI tool. Where the data goes matters as much as what the tool can do.

The line I use

When I decide whether to automate a step, I ask one question: is this reducing manual work, or is it making a judgment?

If it is manual work, automation is fair game, with a person reviewing the output. If it is judgment, a person owns it, and the tool is there to inform them, never to decide for them.

That single distinction keeps you on the right side of the line. Automation should give your specialists more time for review, judgment, and planning. It should not quietly take those things over.

The guardrails that make it safe

Used carelessly, AI can absolutely make provision work worse. Used with a few guardrails, it earns its place:

  • Keep a human in the loop on anything that matters. Review is not optional.
  • Use controlled inputs and validation checks, so you can trust what goes in and catch anything odd that comes out.
  • Make it auditable. You should be able to see how a number was produced and trace it back to its source.
  • Protect confidentiality. Know where the data goes, and agree the controls before any work begins.

What this means in practice

This is the philosophy behind how I built Polaris. Polaris is not designed to replace the tax specialist. It was built to take the manual, repetitive parts of the provision off their plate so they can spend their time where it actually counts, on review, judgment, and planning.

Used this way, with the right guardrails, AI does not make tax less human. It gives the humans back the hours the manual work was quietly taking from them.