Check AI-written drafts against your company facts before they ship

The problem: AI drafts say things your company never said

An assistant writes fluent text whether or not the statements in it are true. For marketing and founder content, the risky parts are the specific ones: a growth figure nobody measured, a customer story that never happened, a promise your product cannot keep. Three sources show how this plays out.

• A company is held to what its AI says. In Moffatt v. Air Canada, decided on 14 February 2024 by the British Columbia Civil Resolution Tribunal, the airline's website chatbot told a customer he could claim a bereavement fare after travel, which was not the airline's policy. Air Canada argued the chatbot was a separate entity. The tribunal rejected that and held the airline responsible for all information on its website, whether it came from a static page or a chatbot. Source: American Bar Association, Business Law Today, February 2024.
• Professional work ships with invented citations. Australia's Department of Employment and Workplace Relations published a Deloitte review in July 2025 that contained fabricated citations and footnotes, found by a University of Sydney academic. The corrected version removed more than a dozen references and disclosed that a generative AI tool chain had been used, and Deloitte agreed to repay the final instalment of its contract. Source: The Register, 6 October 2025.
• B2B software copy already leans on claims it does not source. In our own study of 104 B2B SaaS homepages read on 10 October 2026, the rules found 600 checkable claims (numbers, customer counts, superlatives). Only 24.3% had any visible source next to them, 88.5% of homepages carried at least one unsourced claim, and 3 homepages sourced every one. Source: We checked 104 B2B SaaS homepages for claims they can't back up, with the method, limits and dataset.

The first two are not marketing drafts, and the third studies published pages, not AI output. Read together they make one point: a person or company stays answerable for the statement, so the check has to happen before publishing, not after. Generic advice is a human fact-check checklist. That does not scale across every draft an assistant can write in a day, and it has no record of what was checked.

How the Klype claim check works

1. Saved facts. You save your product facts, buyer and proof once per workspace. That includes limits: a fact that states what the product does not do never counts as support for a claim that it does.
2. Claim extraction. The check reads any draft and picks out what can be checked: figures (digits, written-out numbers and percentages), outcome promises and guarantees, vague authority such as unnamed studies, and stories told as fact.
3. Unsupported-claim flags. Each claim is matched to your saved facts and to any source text you supply with the draft. What matches is marked as verified. What does not is flagged with the sentence and the reason: a figure with no source, a figure not in your business facts, vague authority, or a story the check cannot confirm happened. The check is rule-based, so the same text gives the same answer.
4. Verdict and fixes. The draft gets a score and a verdict of ready, needs changes or blocked, with the specific fixes listed. Your publishing rules are set as block or warn: a block rule stops approval until the problem is fixed.
5. Approver. The owner approves, or sends the draft to a reviewer for sign-off by a private link. Reviewers need no account.
6. Audit record. Each approval is kept in a tamper-evident record of what was approved and how much it was edited, which you can export for a client, an auditor or a buyer.

Klype never publishes. A person approves and posts every post. See the feature pages for Check and Governance, or try the free post checker, which flags unsourced figures and invented stories without sign-up.

Use it from Claude or ChatGPT

The same check runs inside your assistant through the Klype connector (an MCP server, with Cursor and VS Code supported too). Connect once, sign in to Klype and approve the permissions on the consent screen. Then the assistant can:

• read your saved facts with get_workspace_context before it writes;
• run review_draft on text it wrote itself, so unsupported claims come back as fixes in the same conversation;
• record your approval with approve_draft only when you say you approve, and read the record back with list_approvals.

The rules you set apply to every connected assistant, so a draft written in another tool meets the same rules. Setup steps and the full tool list are on the MCP server page.

Who it is for

• Founders who write their own content and use an assistant to draft it, and who cannot afford a made-up figure under their name.
• Small marketing teams where several people prompt several assistants and nobody holds the facts in one place.
• Agencies and ghostwriters who draft for clients and need a sign-off trail they can show.

What the check does not do

It checks figures, promises, vague authority and stories against what you saved. It does not browse the web to verify an outside claim, and when it detects nothing of those kinds it says nothing was verified. Other statements are not fact-checked, so read every draft before it goes out. The check is only as good as the facts you save, and it does not replace legal or compliance review.

At a glance