Sectors · ICT and technology

The runbook is right. The one in Confluence is from 2023.

Your answers live in four places — the wiki, the design docs, the incident channel and a thread nobody linked. AiSU reads across them and cites what it found, so an engineer at 2am can tell which version is current instead of guessing.

Start 7-day free trial Read the architecture notes

No card · Read-only access · EU-hosted by default

A question asked mid-incident

AiSU StrataDrain, then revert the migration before the deploy — the order changed after the March incident. The wiki page still shows the old sequence and has not been edited since 2023.

payments-runbookConfluence · updated 2023#inc-0412 postmortemSlack · 14 Mar

Answers are AI-generated, and dates come from the source. Open the citation before you run anything against production.

Where it earns its keep

Documentation rot, without the documentation project.

During an incident

Ask what the procedure is and get the answer with its source and its date — so the on-call engineer can see that the wiki page is stale and the postmortem is not.

Every answer carries the source's last-modified date.

Why the system is like this

Design decisions live in RFCs, review comments and a Slack thread from two years ago. Ask why a choice was made and get the reasoning, not just the outcome.

Reads design docs and threads, not only the wiki.

Answering the security questionnaire

Procurement sends 200 questions you have answered before. Ask for the previous response and the policy behind it, and reuse what is still true.

Cites the policy document, so the answer is defensible.

Getting a new engineer shipping

A joiner asks the codebase's history in plain English instead of interrupting three seniors — and the answers point at what to read next.

Useful from day one, without a documentation sprint.

For the person doing the review

You will be asked to justify this internally. Here is the short version.

ICT buyers are the ones who read the architecture before the pricing. Nothing here is a claim we would not put in writing.

Document storagenone

Files are read to answer a question and then discarded. We store derived search data — vectors and metadata — not copies of your documents.

Access modelread-only OAuth

No write, send or delete scope is requested. Disconnecting revokes the token and deletes the derived index for that source.

Permissionsinherited from source

Every search runs with the asking person's own credentials, so AiSU cannot surface a document they could not already open.

Model trainingnever on your data

Subprocessor agreements prohibit training on or retaining your content beyond transient processing.

HostingEU by default

Processing sits in the EU. Hybrid on-premise runs inside your own network on Apple silicon, offline, where policy requires it.

Trust centre and subprocessor list How retrieval works

Connects to what you have

Read-only, in an afternoon.

Sign in with Google or Microsoft, choose the drives and channels in scope, and the first index finishes while you get on with the day. No write access is requested, and disconnecting removes the index. Anything outside that scope is never read.

How indexing works

Google WorkspaceDrive, shared drives, Gmail

Microsoft 365SharePoint, OneDrive, Outlook

SlackChannels and threads in scope

NotionPages and databases

Ask it something your wiki gets wrong.

Connect one source, ask about a procedure you know has drifted, and see whether it catches the stale page. Seven days, no card.

Start 7-day free trial Contact sales