For teams shipping AI-drafted text

Fact-check an AI-drafted tax or legal memo.

Lenz is a fact-checking service for tax and legal memos. It identifies the factual statements it can check against independent public sources — what a statute requires, what a court held, a rate, a deadline, a form number — and returns a result for each. It is not a citation checker.

€10,000 warranty per qualifying fact check

Example: a tax memo

“A calendar-year C corporation files its federal income tax return on Form 1120, due April 15, and can extend the deadline by six months by filing Form 7004. An S corporation files the same Form 1120 by the same date.”

US federal, as of 2 September 2026

What a memo asserts about the law.

A memo states things about the law before it argues from them: what a statute requires, what a court held, which rule governs and from when.

Lenz checks each against the published rule and the public record — what the memo asserts, not whether a rule applies to your client.

Statutes and the rules made under them

“The statute requires X.” “The rule took effect in March.” Lenz looks for the published text and the authority’s own guidance, and returns what it finds, including when it finds too little.

Case law: verify what a court held

A model can name a real case and misstate its holding. Lenz treats “the court held X” as a factual statement and checks it against the public sources it can reach. When Lenz finds no public evidence for a statement, the finding says so and the result scores low. Treat that statement as unverified.

Forms, deadlines, rates and thresholds

The mechanics the example above checks. They turn on the jurisdiction and the tax period, so a memo that names both gives Lenz something to check against; one that leaves them implied does not.

AI tax advice and legal research, before you rely on it.

An assistant’s answer reads the same whether it is right or a year out of date, and the memo drafted from it carries the error forward.

Lenz checks each statement in the draft against the public record — a regulatory claim, an effective date, a threshold — and reports each one separately.

Hallucinated holdings and legal AI accuracy

A holding a model invented is checked as a statement, not as a citation: the question Lenz answers is whether the case held what the memo says it held.

Regulatory claims in a legal document

What a rule requires, when it took effect, the threshold at which it bites, who it covers: statements with a published answer, each returned with the regulation or guidance it was checked against. Whether the rule reaches your client is analysis, outside the check.

Check the memo in Claude, Cursor or a script.

Connect Lenz to the assistant that drafts the memo (Claude: one sign-in; Cursor and Claude Code: one command with your key) and ask for a check before the memo leaves the window.

Hand over the memo whole; each statement comes back with its own result.

Memos a product generates, checked before delivery

When the memos are the product, the check belongs in the pipeline, not in someone’s habit. A script or the API sends each memo as it is drafted: extract the statements, assess them in bulk, send the uncertain ones to a full verification, route what needs a reviewer to one before delivery.

Connect Claude →

What this does not do.

It checks the factual statements the memo makes, against public sources.

It does not validate a list of citations, check citation format, authenticate a quotation, or assess the quality of the legal analysis, and it does not judge whether a model wrote the memo.

A record you can keep.

Open a full verification to inspect the source a statement rests on and to keep the reasoning with the memo: the statement as framed, the verdict and its score, every source named and dated, and the arguments for and against.

That record comes from a full verification: Lenz searches independent public sources, weighs what they say and returns the verdict with those sources and the reasoning, in about 90 seconds per statement. An assessment is the other check: a panel of models returns a result and a confidence level in about 10 seconds, for triaging a long memo.

Warranty

Vendors disclaim accuracy. Lenz warrants the verdict.

A verdict can qualify when a full verification returns True or False with high panel confidence and enough independent sources behind it. Its certificate is a document you can hand to the client with the memo.

On a paid Pro or Scale plan, every full verification that qualifies carries a contractual warranty, at no extra charge.

How the warranty works →

Pricing: start free, no card.

Free tier, no card: 100 assessments or 10 full verifications a month, in any combination. The Pro tier is $99 a month for 5,000 assessments or 500 full verifications, the same way, and the Scale tier is $399 a month for production volumes. Every plan includes 1,000 extractions a day, an extraction being one call however long the text.

Compare plans →

Frequently asked questions

Published statutes, regulations and the tax authority’s own guidance, plus independent commentary. A full verification includes the sources it used and the reasoning behind the result.

Only what the statement says. Include the jurisdiction and the tax period; Lenz searches public sources for the rule applicable to that statement and returns what it finds.

Lenz searches the public record when it runs the check, so a published change is in scope; every source and the check itself are dated. Where superseded guidance is still online beside the new rule, a full verification returns both sides and the reasoning. Put the tax period or the date in the statement to anchor the check.

What you send through the API or a connector is private to your account, and each adviser’s checks are their own. The example on this page is a memo paragraph Lenz wrote and published.

Not answered here? Talk to us →

Check the memo before it reaches the client.