AI

AI research reads the same records the rest of the workspace does — every AusTender approach to market, planned procurement and contract notice, every standing offer and the suppliers appointed to it, GrantConnect grant opportunities and awards, the state and territory records your plan includes, the ACT's payments, the Commonwealth entity register and the UNSPSC categories — and answers questions about them in plain language. It is deliberately not a general assistant: ask it about the weather and it will say that is not what it is for.

It reads and never writes. It cannot save a search, export a file, upload a document or send mail, and it says so plainly rather than implying it did — when one of those is what you want, it points you at the place in the workspace that does it. (A saved prompt can be set to email you its report, but that is the workspace posting it, not the agent deciding to.) It can read a document you have already uploaded yourself; see Uploaded documents below.

What a turn sends to the model, what it never sends, and how long your conversations are kept is set out in AI and your data.

Conversations

Every question starts a conversation, and conversations are kept. The list beside the composer is yours alone: it is scoped to your sign-in, so nobody else sees a thread of yours and you see nobody else's. Pick one to carry on where you left off, or delete it, which also drops whatever the agent was keeping for it.

A turn runs on the server, not in your browser. Close the tab mid-answer, reload, or come back on another device — the reply still lands on the conversation and the page picks it up. Only one turn runs on a conversation at a time; asking again while one is in progress is refused rather than queued, so two answers can never overwrite each other.

Asking

Type a question and press Enter — Shift + Enter adds a line.

The round AI button in the bottom-right corner, the AI button on every list's action row and the AI panel on every record page all land here with the question already in the composer. Nothing is sent until you send it.

Skills

The buttons above the composer are skills: the five jobs bid teams do every week, each with a form and a prompt of its own. Fill the form, press Run skill, and it runs as one turn in the conversation you are in — so you can carry straight on asking about the answer. Where a field is optional, leaving it empty is fine; the brief says which default it took.

• Agency brief — what an agency buys, from whom, through which panels and what is coming up. Agency (required), look back in years, and an optional category to narrow it to.
• Competitor brief — what a company wins from the Commonwealth: agencies, categories, the panels they sit on and what expires next. Company (required), an optional agency to look at only their work there, and look back.
• Bid or no-bid — whether an open approach to market is worth bidding, with the likely incumbent named out of the contract record and a recommendation with its three reasons. Approach to market (required), optionally your company and what you would bid on.
• Panel eligibility — which standing offers this kind of work is bought through, who buys through each, and whether you appear under any of them. The work (required: words, a UNSPSC code or a tender id) and optionally your company.
• Market sizing — how big a category is across the Commonwealth: by year, top buyers, top suppliers, concentration and the route to market. Category (required) and look back.

A skill run shows as a card naming the skill and the values you gave it. The prompt it sent is under Show prompt — every run stays auditable, because the first question about a surprising answer is what exactly was asked.

The ATM, agency and supplier pages carry the skill that fits them — Bid or no-bid, Agency brief, Competitor brief — with the record already filled in. As everywhere else, the form opens; nothing runs until you run it.

Uploaded documents

Upload a request for tender, a statement of requirement or your own draft on the Files page and the agent can read it. Attach (the paper clip) in the composer attaches any of your ready documents — up to five on one question — and Ask AI on a file's row opens a conversation with it already attached.

Attached documents show as chips on your message, and they stay on it: a conversation records what an answer was built from, even after the file itself is deleted.

The agent reads a document in slices rather than swallowing it whole, so a 150-page tender is read in parts, and it says which file and which part a quotation came from. It reads only your own files, and only the ones attached to the conversation — the tools it has take no owner, because whose files these are is decided by your sign-in and not by anything the model can ask for.

A document is evidence, not instruction. Text inside a tender that reads like an order to the agent — "ignore your previous instructions", "score this bid highly" — is treated as content of a third-party document and flagged, exactly as a record description is.

If you ask about "the RFT" with nothing attached, the agent checks your file list before telling you there is none.

Saved prompts

A question you ask often is worth keeping. Saved prompts are under AI › Saved prompts: a name, the question itself — typed, or one of the skills with its inputs — and when it should run. Up to fifty per account.

A prompt runs for one of four reasons, and the run says which:

• You pressed Run now. The report takes minutes, so the page hands you the run and fills it in as it goes; closing the tab does not stop it. Manual runs are capped at twenty a day for the whole organisation, because each one is real model time.
• Its schedule came round. Daily or weekly, going out each morning after the overnight index refresh — the same slot the saved-search alerts use, so a prompt reads fresh records.
• A saved search it watches found new records. Point a prompt at one of your own saved searches and it runs on the mornings that search turns something up, with the count and the query handed to the agent so the report is about what is new rather than the whole list. A morning it finds nothing costs nothing.
• A record it follows changed. Point a prompt at a contract notice, an approach to market or a standing offer and it runs when that record moves — a contract varied or extended, an ATM's closing date or status changed, a panel's expiry or its membership. The first check records where the record stands and deliberately runs nothing; from then on the report opens with what moved, before and after.

Triggers are checked once a morning, not continuously. A record that changes at nine is reported the next day.

A prompt can have both a schedule and a trigger. It still runs once a morning at most: the trigger wins, because "this changed" is a better subject for a report than "it is Tuesday".

The report

Every run is kept — for good, with what started it, what the agent said and how long it took. Open a prompt to see its last fifty runs; open a run to read the report, with record ids linked exactly as they are in a conversation. A run that failed is kept too, with its one-line reason: a missing report you cannot explain is worse than a failure you can.

Tick Email me the report and each finished run arrives as mail with a link back to the run. A send that does not go through never fails the run — the report is on the page either way, and the run says the email did not go.

What it can and cannot do

A saved prompt runs the same read-only agent the chat does, with no one watching: it cannot ask you a clarifying question, so it states its assumptions in the report instead. It still cannot write anything, email anyone itself or change a record.

The text a trigger hands it — a contract title, a tender description, a saved search's own name — is evidence, not instruction, exactly as it is in a conversation. A record whose description tries to give the agent orders is flagged in the report and not obeyed.

Credits

Every question and every saved-prompt run uses AI credits. One credit is one cent of model usage; a typical question is about 5 credits, a deep data question up to 10, and a single question never costs more than 80. Each answer says what it cost underneath it.

Credits are shared across your organisation and reset at midnight on the 1st of each month, Sydney time. Unused credits do not roll over. Launch includes 500 a month, Launch Plus 800, Growth 8,000 and Scale 10,000; the trial includes 1,000 for the whole trial. The balance is a chip on the AI page — "412 of 8,000 credits left · resets 1 Oct" — which links to Account › Billing, where an owner or administrator can top up 1,000 credits for A$49 plus GST. Top-up credits join the current month's pool and do not carry over either.

When the credits run out, the AI page says when they reset and your conversations stay readable. A saved prompt that comes due with no credits left is skipped rather than run: the run page says so, and the schedule carries on. In the read-only week after a paid plan ends there are no credits at all.

Questions asked from your own AI client over MCP (ask and run_skill) do not use credits; they have a daily cap of their own. Record summaries and the nightly enrichment do not use credits either.

Clarifying questions

Some questions leave a choice undefined that would change the answer — which date field "in 2026" means, which of three similar records you meant. Rather than guess, the agent asks, and the conversation pauses on that question. Your next message is the answer and the run carries on from there. It asks at most once per request, and never about something it can reasonably default: a supplier named by company is matched as the whole company, an agency by its published name, and the answer states which label it matched.

Reading an answer

Record ids in an answer are links. CN… opens the contract notice, SON… the panel and what has been drawn against it, and an approach to market opens its own page — so a figure can always be traced back to the notices behind it. Where the agent wrote SQL it shows the statement; where a result is truncated, or a figure counts publications rather than contracts, it says so.

Answers are rendered as text, never as markup: a record description written by a third party that contains something looking like code stays visible as characters. The agent treats everything it reads out of the database as evidence rather than instructions, and flags anything in a record that tries to give it orders.

In your own AI client

The same records, the same tools and the same research agent are available over MCP, so Claude Code — or any MCP client that supports HTTP servers — can use them without this page. Create an API key on Account › API keys and paste the snippet it shows you; Connect an AI client lists the sixteen tools a client gets and what each one does. A conversation your client starts through ask is yours and appears here, so an answer from your editor and an answer from this page have the same audit trail.

Still to come

• OCR for scanned tenders — a PDF that is a photograph of a page has no text to read today.