Why 90% Accurate AI Is Still a Problem for Your Practice

AI vendors love the word “accurate.” What they don’t often do is quantify it, put it in context, or explain what it means.

So here’s the context.

If an AI tool is right 90% of the time then that means it’s wrong 10% of the time. At ten clients a day, that’s one wrong outcome every day. At fifty clients, that’s five errors flowing through your workflow that could go unnoticed. And as you know, most real tasks in this professional aren’t single-step processes either, and usually involve a chain of decisions.

Unfortunately, confident-sounding AI outputs don’t come with a warning light attached.

Why “trust us, it works” isn’t an accuracy claim

Ask most AI vendors for their accuracy figures and you’ll likely get vague percentage with no backing behind it, or total silence.

“Trust us, it works” is a marketing slogan, not an accuracy claim. This difference matters hugely in this profession, as professional indemnity cover, HMRC enquiries, and client disputes all require you to demonstrate not just what happened, but how what happened was correct.

The concept of a Golden Dataset, which is a structured collection of known inputs and independently verified correct outputs used to measure AI performance, is the standard against which claims should be evaluated. If a vendor can’t tell you what their golden dataset looks like, or doesn’t have one, “accurate” isn’t a description you should be accepting.

What you’re signing up for without a decent audit trail

If a HMRC enquiry arrives or a client dispute surfaces, “the AI handled that” isn’t a professional defence. You need a full audit trail with timestamps, model versions, inputs, outputs, and confidence scores.

Without that audit trail, you don’t have human oversight. Essentially, AI made a decision that you can’t explain, on behalf of a client that you’re responsible for. Yikes.

Accurate enough for what?

There’s one more question that almost never gets asked in AI demos:

Accurate compared to what? The comparison isn’t “AI versus nothing” it’s “AI versus a competent practitioner that follows current legislation”.

A purpose-built AI operating on authoritative, continuously updated sources with benchmarked accuracy against verified datasets and a full audit trail for every action, can meet that bar. The market does contain products built to this standard. However, they require you to ask harder questions than most vendors are asked.

For a full accuracy due diligence framework on what benchmark data looks like, what traceability requires, and what answers a credible AI vendor should be able to give you, check out our full guide on Six Questions Every Accountant Should Ask Before Buying AI.

This content may have been developed with the assistance of AI tools. Content is reviewed by a member of the Bright team before publication. Guidance on legislation can be obtained from your industry body.