AI accuracy you can defend
This blog was originally posted on 20th November, 2026. Further developments may have occurred after publication. To keep up-to-date with the latest compliance news, sign up to our newsletter.
Authored by Eric Farr, Chief Technology Officer, Adherent
Why we built this release around what goes into an assessment, what comes out, and who has the final say.
Picture a compliance manager at a power tool company. She opens a battery regulation in Adherent and finds it marked as applicable to one of her cordless drills. She does not sell batteries. She sells drills. Until recently, the explanation under that verdict would have told her, in slightly more words, that the regulation applies to her drill. She knew that. What she wanted to know was why.
Today the explanation reads differently. The regulation does not name power tools. It covers batteries placed on the market, and this drill ships with a removable lithium-ion pack, so the drill is reached through the battery it contains rather than through anything about the tool itself. An exemption exists for batteries built in so they cannot be removed. It does not apply here, because this pack is user-replaceable. Under that sit two short lists. One is headed “What mattered about your product”. The other is headed “What we had to assume”, and it holds a single line: that the pack is supplied with the drill rather than sold separately.
If that assumption is wrong, she does not argue with the AI. She corrects the product record and runs the assessment again. An assumption you can see is an assumption you can fix.
The wrong question
“How accurate is your AI?” is the first question I hear in almost every evaluation. It is a fair question, and it is the wrong one.
A percentage on a benchmark tells you how a model performed on someone else’s products, against someone else’s regulations, on a particular day. It cannot tell you whether you can stand behind this verdict, on this product, when an auditor asks about it two years from now. Compliance decisions have to hold up long after the demo ends.
So I have stopped thinking of accuracy as a score the model earns. It is a property of the whole system: what went in, how the reasoning was recorded, and who had the final say. We built this release around those three things.
Before the AI answers, you see what it thinks your product is
When a customer disagrees with a verdict and we look into it, the cause is more often a context failure than a reasoning failure. The assessment reasoned soundly about a product that was not quite theirs: a material missing from the record, a wireless feature never written down, a sales channel left blank. Two similar products can land on different outcomes because one record says something the other leaves out.
Until now, the profile Adherent built from your product description came together behind a loading screen and went straight into your first assessment. You saw the results before you ever saw the inputs.
In the new setup flow, that order is reversed. Ari researches your product while you choose your market and topics, then stops and shows you what she found: the regulated attributes, grouped by category, covering the materials, components, claims and classifications that decide which rules apply. You add anything she missed and remove anything that does not belong. Only then does the first assessment run.
The same review lives under Context for every product you add afterwards. Change an attribute and Adherent marks the product as needing assessment, records who changed what and when, and waits for you to run it.
After the AI answers, you can read how it got there
“Tell me more” sits under an applicable verdict. It used to restate the outcome as bullet points. It now explains the analysis: what about your product brought it within the regulation’s reach, whether the regulation names your kind of product or catches it through a broader duty on the manufacturer or importer, which exemption was weighed and why it did not take you out of scope, and what had to be assumed because your record was silent.
The part I care about most is a constraint. The explanation is held to the record the assessment left behind. It is not allowed to add a market, a date, a threshold or a legal consequence that the analysis did not use. If the assessment did not record it, the explanation cannot say it. That is the difference between an account of a decision and a plausible story about one.
You can also ask Ari why a regulation applies to one of your products. She answers from the same record.
Then you decide
AI makes the first pass, and it can be wrong. When you know better, you have two options. If the AI missed information about your product, you can adjust it and have the system run the assessment again. If the system had the right product details, you can change the verdict, write a justification your colleagues will see, and the result carries an Overridden badge. A later assessment does not quietly put it back. When something looks off, but you are not ready to change it, you send feedback with a reason, and that feeds how we improve future assessments.
Put those pieces together, and you have something you can hand to an auditor, a customer’s procurement team, or your own leadership: the product as you reviewed it, the reasoning as it was recorded, your decision with your reason, and the history of every change. That is what I mean by accuracy you can defend.
Showing our work
In June I wrote about the architecture underneath Adherent: every regulation distilled into a structured decision model that is checked before it is used, and every verdict stored with the evidence that produced it. This release does not change that foundation. It opens windows onto it. “Tell me more” reads from a record that already existed. The setup review puts the one input that record depends on in front of you first.
We did not make the model talk more. We made the system show its work.
What I will not tell you
I will not tell you our AI is always right. No one’s is, and in this field a vendor who claims otherwise should worry you. What I will tell you is that when it is wrong, you will be able to see where, fix it at the source, and keep the correction. Your experts make the final call. The system is built so that they can.
Join us at the launch webinar to see the new setup flow and “Tell me more” in action, or book a demo at adherent.com.

See Adherent in Action
Discover how agentic AI is reshaping product compliance for global enterprises.
