How Release Assurance Checks a Claim

Why “Verified” Needs a Definition

“Get the numbers checked before it goes out” is good advice for any deck, website or proposal. It is also close to meaningless until someone says what checking involves. The word is used for everything from a colleague reading a draft to a formal audit.

Most informal checking is really a plausibility test. Someone reads the claim and decides whether it sounds right. That has value, but it has a structural weakness: the reader is not required to find the original source, is not accountable if they are wrong, and, in the case of a general-purpose AI assistant, has a documented tendency to agree with the person asking.

Release Assurance exists to replace “sounds right” with a defined sequence. This article follows one claim through that sequence, from the moment an asset arrives to the moment a named person signs the result.

Step 1: Agree What Is Being Checked

Before anything is researched, the asset is read and every checkable claim is listed: numbers, comparisons, “first” and “only” statements, customer results, regulatory statements and anything a buyer or investor would rely on. The scope is confirmed in writing before the work starts, so it cannot quietly shrink later to meet a deadline.

Take a single example claim from a B2B website: “Trusted by hundreds of UK accountancy firms, we cut month-end close time in half.” That sentence holds three separate claims: a customer count, a sector, and a measured result.

Step 2: Find the Evidence Before Judging the Claim

The order matters. If you judge a claim before you know what evidence exists, you start deciding which claims “feel” true. So the evidence comes first: official registers where they apply (such as Companies House or the FCA register), the client’s own records, and named, dated public sources.

For the example, that means the customer list behind “hundreds”, how many of those customers are accountancy firms, and the data behind “in half”: how many customers were measured, over what period, and compared with what baseline.

Step 3: Test It Against the Strongest Objections

Each claim that has evidence is then tested against the objections a careful reader would raise. Three are always asked:

• The buyer’s or investor’s objection. “Half of what, measured how, for whom?”
• The regulatory objection. In the UK, the Advertising Standards Authority expects marketers to hold evidence for objective claims before they publish them, and comparisons with identifiable competitors have their own rules.
• The most likely factual alternative. Perhaps the result came from a handful of early customers with unusually messy processes, and the typical result is smaller.

A claim has to survive all three, not only the one you expected.

Step 4: Cite It or Rewrite It

A claim that survives is cited to a standard: the source is named and traceable, the date is stated, the method behind any number is on record, the sample size is explicit, and a confidence level is given. A citation is a package, not a link.

A claim that does not survive is not simply deleted. It comes back with a rewrite the evidence does support. For the example, that might be: “Used by more than 120 UK firms, including 40 accountancy practices. Across 18 customers measured in 2025, month-end close time fell by a median of 38%.” (The figures here are illustrative.) Less dramatic, and much harder to challenge.

Step 5: A Named Person Signs

Every claim in scope is checked against its source and marked as verified, failed or unverifiable. On top of that, the named verifier personally re-checks claims according to risk: every decision-critical and high-impact claim, at least 30% of medium-impact claims, and a spot-check of around 10% of low-impact claims. The Assurance Note states which claims received a full human re-check.

Then a named person signs. If they will not sign it, it does not ship. Their name stays on the document, and if an error is later confirmed it is corrected publicly under our errata policy.

Why an AI Assistant Alone Is Not Enough

General-purpose AI assistants are useful for finding sources and drafting. They are a weak final check. Research published at EMNLP 2025 found that language models tend to agree more with a position the more forcefully it is argued, regardless of whether it holds. Asking an assistant “my market is worth this much, right?” is asking an agreeable system to grade a claim you have already committed to.

We use AI to speed up retrieval and drafting. A person reads the sources, tests the reasoning and signs. Every paid deliverable says which steps used AI, as set out in our AI transparency policy.

What You Get, and How Long It Takes

The Release Assurance Pack checks one asset, such as a website, a sales deck or an investor deck, and returns a signed Assurance Note in 3 business days. The Release Assurance Suite covers several assets together, with a check that they agree with each other, in 5 business days. If you want to see which of your claims a careful reader would doubt before you commit, start with a free First Signal in claims mode from the Release Assurance page.

Pranav Unni

Founder · ThriveFinity Connect on LinkedIn →

Pranav Unni is the founder and lead verifier of ThriveFinity. He reads and signs every paid deliverable personally, and writes about go-to-market decisions for established UK B2B companies.