Legal/Acceptable use

Acceptable use

Short, and mostly obvious. The one clause worth reading properly is the last section: using the product to remove human accountability is the thing we care most about, and the thing a list of banned words would not catch.

Last updated 1 September 2026

Don't use Xalios to

Break the law

Including sanctions, export control, fraud, money laundering, and anything that would put you or us on the wrong side of a regulator. If a process exists to evade an obligation, it is not a process we will host.

Process data you have no right to

Documents obtained without authorisation, personal data you have no lawful basis to process, or another organisation's confidential material you happen to hold.

Attack the service

Probing another tenant, attempting to bypass permission enforcement or trust levels, denial of service, or load testing without asking. Good-faith security research is different and welcome: contact us first.

Interfere with the record

Attempting to alter, suppress or fabricate decision records, or to make an action appear to have been approved by someone who did not approve it. The record's integrity is the product.

Resell or reverse engineer it

No reselling access, no sharing logins outside your organisation, no scraping the product or benchmarking it for a competitor without telling us. Ask and we will usually say yes to the honest version.

Deceive people about who acted

Presenting an AI employee's output as a named person's individual judgement where that misleads someone relying on it: a client, a regulator, a court. Attribution can be to your organisation; it should not be a lie about a human.

The clause we actually care about

Don't use autonomy to make responsibility disappear.

The product is built so that a person grants what an AI employee may do, and the record shows who granted it. That only works if the grant is a real decision by someone accountable. Configuring it so that no human can be identified as having decided, and then pointing at the software when something goes wrong, is a misuse of it, whatever the terms elsewhere say.

Keep an owner named

Every AI employee needs a current owner and a fallback. An unowned one running in production is a breach of this policy in substance.

Don't automate what must be judged

Where a decision has a legal or similarly significant effect on an individual, keep it supervised so a person decides it.

Approval means reading

Automating approvals, or routing them to an unread queue to satisfy a control on paper, defeats the purpose of having one.

What we do if something breaches this

We talk to you first. Most of what looks like a breach is a misconfiguration or a process that grew sideways, and a conversation fixes it. If there is a live risk to other customers or to individuals, we may suspend the affected part immediately, with a written reason and a named person to speak to, not a generic notice.

Termination is a last resort for a serious or repeated breach, under the terms of service. Even then you get your export window: we will not withhold your records to win an argument.

Not sure if your use case is fine?

Ask before you build it. We would much rather tell you no in a twenty-minute call than discover it in a review six months in.

Ask us→