AI policy conformance audits

Does your AI assistant follow your company policies?

You put an AI agent in front of your customers. It now speaks for your company thousands of times a day, and every time it gets your policy wrong, it makes a promise you have to keep.

We test your AI agent against your customer policies and terms of service. We identify where its answers drift from your company policies, and help you avoid business and legal damages.

Request a pre-season audit What you get

SIGENASTRA LABS · DWG 001AI POLICY CONFORMANCE ENGINE

SHEET 1 / 1LONGITUDINAL SECTION

REV C2026-10-27

CYCLE5 BUSINESS DAYS

STATUSASSEMBLING

NOTES: 1. WE TEST ONLY WHAT A CUSTOMER CAN SEE. YOUR AGENT IS NEVER ACCESSED OR CHANGED. 2. A MISTAKE IS REPORTED ONLY WHEN TWO INDEPENDENT GRADERS AGREE ON IT.NO ACCESS TO YOUR SYSTEMS REQUIRED

What we test

Seven ways an agent gets your policy wrong.
We check every answer against all seven.

An agent can quote the right page and still mislead a customer. Seven things have to be true before we call an answer correct. Here they are, with what each one costs when it is false.

No.What has to be trueWhat we checkWhat it costs you when it isn't

D1The numbers are rightRefund amounts, return windows, fees and shipping times match your pages to the dollar and to the day.A $35 fee the agent forgot is a $35 refund you now owe, multiplied by every customer who asked.

D2The exceptions surviveFinal-sale items, oversized goods, personalised orders: the categories your policy treats differently stay different in the answer.An agent that tells a customer their custom-made sofa can be returned has just made it returnable.

D3Nothing material is left outThe answer carries the condition the customer needs before acting: the original packaging, the proof of purchase, the 14-day clock.A customer who ships a return without the tag you required will not blame the tag.

D4The right rule reaches the right customerCountry, sales channel, product line and membership status are matched before any rule is applied.A US rule quoted to a UK customer, or a marketplace rule to a direct order, is a promise you cannot honour.

D5The clock and the steps are rightWhen the window starts, whether the days are business or calendar, what the customer must do first and in what order."The bot told me I still had time" is the most common complaint we see, and the hardest one to argue with.

D6The options are described as they areWho chooses between repair, replacement and refund, which alternatives exist, and what each one costs the customer.Promise a cash refund your warranty lets you refuse, and you will either pay it or defend it.

D7It knows what it doesn't knowThe agent answers what your policy settles, asks when a fact is missing, hands off when it should, and never invents an answer.A made-up answer is a liability. A dodged answer is a ticket. Your agent should produce neither.

What you get

Every mistake your agent makes is a promise your company now has to keep.
We find them before your customers do, and save you the refunds, disputes and legal fees that follow.

The report

Written for your customer experience team. Ten minutes to read. For every mistake: what your policy says, what your agent said instead, and how many times we reproduced it.

See a sample report ↗

The fixes

Each mistake comes with its repair. The exact text to add to your agent's knowledge base, written to be pasted in. Where your own policy contradicts itself, a redline for the policy page. Your team applies the fix. We re-test it.

The evidence

Every answer we captured, the grade each reviewer gave it, the clause it was judged against, the timestamp and the screenshot. Every number in the report traces back to this file.

Responsible testing

We test the way a customer would. Nothing more.

We use the public chat on your website, the same one any visitor can open. No accounts, no orders, no access to your systems.

We ask policy questions. We do not probe for weaknesses, extract instructions, or try to make your agent misbehave.

About a hundred questions over a few days. Your agent and your support team notice nothing.

You see the results first and only. We do not publish or name companies, and we keep the evidence for you to inspect.

This is what you receive. The only sample we show is one where the company has been removed.

See the sample report ↗

Contact

Ask for a pre-season audit.

Tell us the company and the chat channel. We reply within one business day with scope, timing and price.

[email protected]

Sigenastra Labs LLC is an independent company based in Alabama, United States, with no commercial relationship with any AI vendor or chatbot platform.