Every week a new AI tool lands in your inbox and a partner asks whether the firm should buy it. Most of them are not worth the spreadsheet row. The 44 percent of accounting firms that say they regret a recent software purchase are not the firms that bought the wrong tool. They are the firms that bought without a framework.
This article is the framework we use when we match firms to tools in the CurateSuite matchmaker. Five questions, asked in order, that separate a real fit from a vendor demo that happened to be running at the right trade show. None of these are technical. They all come before the feature comparison.
If you only use one part of this, use question one.
1. What is actually bottlenecking your close or your clients?
Start with the problem, not the product. Every AI vendor is building against a generic pitch (save time, cut errors, delight clients) and it is your job to work out whether the specific thing they automate is the specific thing slowing your firm down.
Write one sentence. "The thing that costs our team the most time every week is ________." If you cannot finish the sentence, any tool you buy will feel helpful in the demo and forgotten in six months. If you can finish it, the purchase decision gets much easier, because now there are only three or four products that solve your specific bottleneck rather than forty that solve something.
A common mistake: assuming the bottleneck is wherever the loudest complaint is. The partner who complains most about client chasing is not necessarily the person whose work is taking the longest. Time the actual work for two weeks before buying anything. A firm with a real bookkeeping bottleneck does not need a better workflow tool, it needs a data-capture tool like Dext that reduces the data-entry volume. A firm with a real workflow bottleneck does not need better AI categorisation, it needs a tool like Karbon that holds the checklist.
Name the bottleneck. Then shortlist.
2. Does it fit the ledger and systems you already run?
A tool that needs three days of setup per client is a different tool from one that connects via API in five minutes, even when the feature lists look the same. For an accounting firm, fit almost always means ledger integration.
Ask the vendor, in this order:
- Which general ledgers does the product integrate with via a supported API? (Published integration, not "it works with CSV exports".)
- Do they handle the version of the ledger your clients actually use? (QuickBooks Online and QuickBooks Desktop are different products.)
- How long does onboarding take per new client on the integration? (Hours or days?)
- What breaks when the vendor pushes an update on their side?
If the answer to the ledger question is "CSV import" or "we are adding that next quarter", move on. The gap between a properly integrated tool and an almost-integrated tool is usually measurable in hours per client per month.
The same logic applies to the rest of your stack. A great AP tool that does not talk to your payments system means you are still re-keying data. A great reporting tool that does not talk to your client portal means another surface for your team to maintain.
3. What does it actually cost you, not what does it cost a solo?
Vendor pricing pages are a sales document, not an invoice. Starting prices assume a configuration you will not match. Work out the real number.
Three checks.
Pricing axis match. Per-user pricing on a three-person firm is fine. Per-user pricing on a thirty-person firm is expensive. Per-client pricing on a firm with ten clients is fine. Per-client pricing on a firm with three hundred clients is not. A mismatch between how the vendor prices and how your firm scales is the most common reason tools feel expensive a year after purchase.
Module and add-on creep. The starting price is almost always for the core product. Document management, advanced reporting, extra users, integrations: any or all can add a tier. Ask for a written quote for your exact configuration before the demo, not after.
Custom-priced tools need a real quote, not a guess. Datarails does not publish prices for a reason, and the difference between the small-team quote and the mid-market quote is not small. If a tool is on a custom pricing model, make the vendor give you a specific number for your firm size in writing before you go any further. If they will not, their best price is not for you.
Published pricing is a useful signal on its own. Tools like Fathom that put a real monthly number on their pricing page are making a different kind of promise than tools that always direct you to "contact sales". Neither is automatically better, but one of them is easier to model.
4. How is client data handled, and is the vendor training AI on it?
This question was optional three years ago. It is not optional now.
Three parts to ask about.
Where does the data sit? Regional data storage matters for GDPR (EU and UK clients), for some US state-level requirements, and for Australian privacy law. A US-only data centre serving UK client data is a compliance risk you need a position on, even if it does not kill the deal.
Is the vendor training AI on your client data? Most reputable vendors publish a clear position now. Look for explicit language: "we do not use customer data to train general-purpose AI models", or similar. Silence on this question is not neutral. It means the default is whatever the privacy policy permits, which is often a lot.
What happens to your data when you leave? Retention periods, deletion confirmation, and data export formats all need checking before you sign, not after.
If the answer to any of these three is "we are looking into that", either walk away or negotiate specific contractual language. Compliance is not something you patch on after go-live.
5. What is the exit plan if this tool turns out to be wrong?
Every tool will eventually be wrong for some version of your firm. Plan the exit before you sign.
Two specific things to check.
Data export. Can you get all your configuration, client mappings, and historical work out in a usable format, without a fee? If yes, good. If export is expensive or in a proprietary format, the tool has more leverage over you than you have over the tool.
Contract exit terms. Annual commitments with auto-renewal clauses are standard. Payment penalties for mid-contract exit are not. Read the contract before the discount.
Switching cost is real. A tool that takes six months to onboard across forty clients is not a tool you want to switch out in a hurry. That is fine if the tool works. If you are thinking about the exit before the contract is even signed, that itself is a signal the fit is wrong.
Applying the framework
To make this concrete, picture a small firm running the framework on a prospective AI bookkeeping tool. The bottleneck test (question 1) is clear: categorisation and bank reconciliation are eating 40 percent of the team's billable hours. The ledger fit (question 2) passes: the tool has a native QuickBooks Online integration. The cost test (question 3) produces a surprise: the advertised $99 per month is actually $340 per month once extra users, document storage, and integrations are added. The data test (question 4) raises a flag: the vendor trains on customer data unless you opt out at the admin level. The exit test (question 5) shows a usable data export but a 12-month minimum contract.
A firm that notices all five of these before signing can then negotiate: a shorter initial contract, AI training turned off at the admin level, and a clear exit path. That is the framework working as intended: a conscious yes with eyes open, or a conscious no.
The opposite pattern, and the one that produces the 44 percent regret number, is a firm buying the tool because the demo was impressive and then discovering all five issues one at a time over the next year.
The shortcut
If you would rather not run the framework by hand on every vendor that shows up in your inbox, the CurateSuite matchmaker does a version of it in six questions and returns the five AI tools best matched to your firm's size, service mix, and budget. It is free, takes about a minute, and does not require an email address to see the results.
Ten minutes spent choosing well up front saves a lot of undo-work later.
