The CFO Filter | Five Questions Before You Adopt the Next AI Tool

Every week there is a new AI tool that your inbox tells you is essential. Sometimes the pitch comes from a VC who saw a demo. Sometimes from a peer who is genuinely excited. Sometimes from a vendor who wants twelve months on a contract.

Most of these tools will not matter in two years.

That is not a contrarian pose. We have spent the last year building and using AI tools inside a fractional CFO practice, and watching founders lose weeks to frameworks that were unmaintained six months later.

The signal-to-noise ratio is bad, and getting worse. What you need is a filter, not a feed.

Below is the filter we run before we let any AI tool into a client's finance workflow. Five questions. If any answer is uncertain or negative, the default is skip. The cleanest external version of this we have read recently came from a writer named Rohit at the end of April 2026; we have adapted it for the lens a CFO actually works in.

1. Will this matter in two years?

The first question, and the one that rules out the most candidates. Wrappers around a frontier model, CLI flag conveniences, "Devin but for the close," "ChatGPT for FP&A." Almost always no. The half-life of a wrapper is short. The half-life of a primitive is years.

When ChatGPT launched, more than a hundred wrapper tools shipped within six months. Most of them are gone. The pieces that remain are the primitives: the model providers themselves, MCP as a protocol layer, the handful of real agent harnesses. Those will still be running, in some form, in 2028.

For finance specifically, the test is whether the tool ships a primitive or a feature. A tool that automates one specific close process is a feature. A tool that gives you an auditable, retryable connection between your ERP and a model, with auth and logging you can read, is closer to a primitive. Ask which side of the line you are being asked to buy.

2. Has someone you respect built something real on it and written about it honestly?

Marketing posts do not count. Demo videos do not count. Postmortems do.

A blog called "we tried X in production and here is what broke" is worth ten launch announcements. Look for what happened after the honeymoon. Which edge cases hit them. What it cost when it failed. What they would do differently.

If the only people writing about a tool are the vendor and a handful of influencers who got early access, you are looking at a marketing campaign, not a community. The good signal in this space is always written by someone who lost a weekend to it.

For finance, ask whether anyone has used the tool through a real close cycle, a real board meeting, a real audit. "We piloted it for two weeks" is not the same as "we ran the quarterly close on it for two quarters." Those are different products in the buyer's mind. The vendor will not make that distinction. You have to.

3. Does adopting it require you to throw out your tracing, your retries, your config, your auth?

If yes, it is a framework trying to be a platform. Frameworks trying to be platforms have a roughly 90 percent mortality rate. Good primitives slot into your existing system without forcing a migration.

For a finance function, the blast radius matters more than usual. If a new AI tool requires you to migrate your chart of accounts to its specific schema, route every transaction through its API, or replace your existing close process with theirs, the cost of being wrong is enormous. The damage extends past the tool itself into every downstream system you migrated to fit it.

A simple rule: the AI tool should adapt to your stack, not the other way around. Anything else is a re-platforming project in a marketing-tool disguise.

4. What does it cost you to skip this for six months?

For most launches, the honest answer is nothing.

You will know more in six months. The winning version will be clearer. The frameworks that survive their first quarter of production use will still be available. The ones that did not survive will have spared you the migration.

This is the test most people refuse to run, because skipping feels like falling behind. It usually is not. The people who waited six months to adopt LangChain in 2023 saved themselves an entire migration when the abstractions changed. The same pattern has played out in finance technology three times in the last decade. Boring discipline pays. The most expensive thing in this category is being early, not being late.

5. Can you measure whether it actually helps?

If you cannot, you are guessing.

This is the test that most finance teams fail. They adopt an AI tool to "speed up close," watch close cycle time drop by a day, and credit the tool. They never check whether the same drop would have happened from the parallel process improvements they made at the same time. They never check whether output quality dropped. They never check whether the tool gets worse when the underlying model is upgraded.

The instrumented version: before you adopt anything, write down what you would need to see in six months to believe it worked. A specific metric, a baseline, and a target. Then come back in six months and check. If you cannot articulate the answer up front, the tool's value is unmeasurable, and you are making a faith-based decision with finance dollars.

The deeper point

Run any AI tool that hits your inbox through these five questions. Most will fail at question 1 or 4. A small number will pass all five. Those are the ones worth your attention. The rest belong on a six-month watch list, not in the next budget cycle.

The point underneath the filter, though, is that all five questions require judgment. Someone has to decide what "matters in two years" means for this specific business. Someone has to know which operators to trust as references. Someone has to read the contract and find out what it actually costs to swap out. Someone has to design the metric that determines whether the tool is helping.

That someone is the CFO. The person with budget authority, the cross-functional view, and the discipline to ask the boring questions when the rest of the room wants to ship the demo.

This is not a pitch for finance leadership in the abstract. It is a description of the work that actually has to happen for an AI investment to pay back. If the founder is making the buy call without that layer in the seat, and at most $20M businesses they are, the filter does not get run, and the AI line item on the P&L grows without any way to defend it on the next board call.

If any of this landed, let's talk. Not about AI. About what your finance function actually needs right now, and how AI shows up only where it changes the answer. scott@mainstreetiq.com.

Main Street IQ provides embedded fractional CFO services across Southern California: Santa Barbara, San Luis Obispo, Ventura, Los Angeles, and Orange County.