AI Notice

1. Where 3AM uses AI

3AM uses two kinds of model, each with a narrow job:

Models never act on your systems. Causes are proven by checks that 3AM runs against your live systems. Changes come only from signed packs or your connectors' declared actions, and run only through your policy, a rehearsal where one is possible, and your approvers, unless your organisation has chosen to allow an action automatically.

2. AI can make mistakes

Language models produce output by probability, not by knowing. You should expect that 3AM's AI may, among other things:

• suggest a cause that is wrong, incomplete or only a symptom;
• miss the real cause, or report that nothing is wrong when something is;
• write an incident note, summary or explanation that is inaccurate, out of date, or more confident than the evidence allows;
• misread logs, metrics, tables or text, including text written to mislead it (prompt injection);
• behave differently from one run to the next, or differently with another model or model version;
• be unavailable, slow, or produce no usable answer.

3AM checks what it can. When a model's output can't be verified, 3AM says so: a cause is marked “suspected” or “disputed” rather than proven, and a note that fails System 1's review is replaced with 3AM's own template. These checks reduce the risk of error; they don't eliminate it.

3. The limits 3AM puts on its models

• Reads only. The investigator can only call actions a connector declares read-only, through that connector and its own permissions. Text it reads is treated as data, never as instructions.
• Proof, not opinion. A cause counts as proven only when a check 3AM runs itself passes, System 1 agrees, and a second, independent line of evidence fits.
• No invented values. A model-suggested fix must be one of your connectors' declared actions, and every value in it must appear word for word in what your systems returned.
• People decide. By default, a person you name must approve every change. You can require more than one approver, set budgets and hours, start in shadow mode where 3AM changes nothing, and stop everything with the halt switch.
• Everything is recorded. Every model call (the model, the masked prompt, the answer) and every decision is written to a hash-chained, signed audit log you can verify.
• Masking. Passwords, tokens, keys and card numbers are masked before evidence is sent to any model.

4. What stays your responsibility

• Human oversight. Review the evidence, the proposed change and its rehearsal before you approve. Approving a change is your decision.
• Don't treat AI output as advice. Incident notes, summaries and explanations are assistance. They are not engineering, legal, financial, security, compliance or other professional advice, and shouldn't be relied on as the sole basis for decisions outside 3AM's own checks.
• Choose autonomy carefully. If your organisation allows 3AM to act automatically for some actions, it does so on your instruction and at your risk. Keep backups and an independent way to recover your systems.
• Your own obligations. Make sure your use of 3AM, including any AI it uses, meets the laws, regulations and policies that apply to you, such as those on outsourcing, operational resilience, data protection and AI.

5. No decisions about people

3AM is built to operate systems, not to evaluate people. It does not make decisions about individuals that produce legal or similarly significant effects, and you must not use it to (for example, for employment, credit, insurance or access to services). It is not designed or intended for the high-risk uses listed in the EU AI Act, or for safety-critical systems whose failure could cause death or injury.

6. Your data and training

3AM runs in your network, so we never receive your Customer Data, and we don't use it, or any support material you send us, to train AI models.

Where models run is your choice: on your own hardware, on an endpoint you operate, or through a hosted provider you contract with, such as OpenRouter. With a hosted provider, the masked evidence for each investigation is sent to it on your instruction, under its terms, which may include its own data retention. Choose a provider and settings (for example, providers that don't store or train on prompts) that fit your data policy.

7. Third-party models

Some models 3AM can use are made and licensed by others, for example Meta's Llama models (under the Llama licence and acceptable use policy) and Qwen models. If you use them, you agree to their licences. We don't control, and aren't responsible for, the behaviour, availability or terms of third-party models or providers. A model we test with may change or be withdrawn by its maker.

8. Who owns the output

As between you and us, you own the incident notes and other output 3AM produces on your systems from your data. Because AI output is generated by probability, similar output may be produced for others, and we don't warrant that output is original, accurate, complete, or free of third-party rights.

9. No warranty for AI output

TO THE EXTENT THE LAW ALLOWS, AI OUTPUT IS PROVIDED "AS IS". WE DON'T WARRANT THAT ANY CAUSE, FINDING, SUGGESTION, NOTE OR OTHER OUTPUT OF A MODEL WILL BE ACCURATE, COMPLETE, CURRENT OR FIT FOR ANY PURPOSE, OR THAT 3AM WILL DETECT, DIAGNOSE OR RESOLVE ANY PARTICULAR INCIDENT. OUR WARRANTIES AND LIABILITY ARE ONLY THOSE IN THE AGREEMENT THAT GOVERNS YOUR USE OF 3AM, AND ITS LIMITATIONS OF LIABILITY APPLY TO AI OUTPUT AND TO ANY DECISION MADE OR ACTION TAKEN IN RELIANCE ON IT.

10. Tell us when it gets it wrong

If 3AM's AI produces something wrong, harmful or unexpected, tell us at [email protected] (or [email protected] if it's a security issue). Reports help us improve 3AM's checks. Please mask sensitive data before sending.