The AI Wrapper: Same Model, New Inputs

When I started at PwC, I inherited a tax cash forecasting model with more than a thousand rows of entity-level data feeding into it: income, capital gains, capital losses, foreign tax credit, foreign tax paid, net operating losses, all rolling up to one number: consolidated cash tax. Every time we wanted to stress-test something: what happens if Japan has a catastrophic loss year? what happens if we change an assumption across a whole class of entities? someone had to go through and manually update the rows by hand.

AI’s best job isn’t running the model — it’s talking to it

One of the most useful ways I’ve found to apply AI to a process like that is as a wrapper. You take a process you already trust (built in Excel, Python, etc.) and you put AI on the input side, the output side, or both. The underlying calculation logic never changes. AI’s job is to translate a plain-language request into the structured change the model already knows how to run, or to translate the model’s output back into something you can act on without staring at a spreadsheet.

That’s a different job than asking AI to rebuild or rewrite the model itself. Changing a model’s logic is where the risk sits, new formulas, new structure, new places for a mistake to hide. Translating a request into an existing model’s inputs is comparatively low-risk, and it’s also where AI is genuinely good.

A simple example

Say you’re running a scenario model for an accounting or advisory firm. Trailing three-month averages by project, jurisdiction, contract type, rate, hours, revenue, cost, EBITDA. Nothing exotic. Now leadership asks: We’re planning on raising our fees by 5% across the US. Can you model the first quarter of next year?

With the wrapper approach, AI reads that sentence, identifies the rows that match: United States, time-and-materials contracts specifically and updates the rate on just those rows. It leaves fixed-fee contracts alone. It leaves every other jurisdiction alone. The model runs exactly like it always has, just with new inputs.

Run it, and you’ll typically see a modest revenue lift with a proportionally larger EBITDA lift, because the cost side never moved.

Why this reduces risk, not adds it

In my experience with automation and modeling, when something goes wrong, it’s almost always the inputs, not the formulas. Someone updates 40 of 43 rows and misses three. Someone changes the wrong column under deadline pressure. That’s the actual failure mode, and it’s exactly the kind of work AI is well-suited to take off your plate: identifying the right records and applying a change consistently, instead of a person copying and pasting by hand.

It also sidesteps a problem I’ve written about before in The Session You Can’t Actually Audit: the thing doing the actual math here is still the deterministic model your team already knows how to audit. AI never touches the calculation, it only touches the inputs going in or the summary coming out. None of this replaces review. You still need the professional skepticism to check the output before it goes to a client or a partner. AI just removes the copy-paste step where most of the actual errors happen.

This scales with your data

The example here is simple on purpose: project level, a handful of columns, one change. The value shows up as you add granularity: staff level instead of project level, individual accounts instead of jurisdictions, leverage assumptions instead of just a rate. If your input data is structured enough to support that level of detail, AI can generate and adjust it at scale just as easily as it can for three rows, and firms are already doing versions of this. Journal of Accountancy recently profiled four small firms building exactly this kind of practical AI tooling around the systems they already have, not replacing them.

Key Takeaways

• Wrap AI around the input or output of a process you already trust — don’t ask it to rewrite the process itself.
• The underlying calculation logic never changes. AI’s job is translating language into structured inputs, or structured output into plain language.
• Most scenario-modeling errors live in manual data entry, not in formulas — that’s exactly the failure mode this approach reduces.
• Changing a model’s logic with AI carries meaningfully more risk than wrapping its inputs or outputs. Know which one you’re doing.
• The pattern scales: the same approach that updates three rows works whether you’re modeling one change or fifteen, at the project level or the individual level.

Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: The AI Wrapper: Same Model, New Inputs.