Ask in the words you use at work.
Natural-language questions
Ask for a report, an explanation or a draft. Cobalt resolves the request, retrieves the permitted records and uses checked results to prepare the response.
Why did gross margin change this month?
Translate the question into the relevant period and measures, calculate the movement and explain the result with links to the supporting records.
Can you explain the gross profit movement?
Cobalt calculates the figures used in the explanation.
Natural-language reportingIllustrative records| Measure | Actual | Budget | Variance |
|---|---|---|---|
| Revenue | 1,240 | 1,200 | +40 |
| Cost of sales | 742 | 720 | +22 |
| Gross profit | 498 | 480 | +18 |
Gross profit is £18,000 ahead of budget.
Revenue is £40,000 higher and cost of sales is £22,000 higher. A language model can explain those checked results; it does not supply the ledger totals.
- Level 1, built-in understanding — recognises your own terms, fills in the details and remembers what you were talking about. Free and instant.
- Level 2, a simple sorter — a small language tool that runs inside Cobalt ERP and sends nothing outside. Free.
- Level 2b, remembered questions — a question your team has asked before, answered the same way and checked against your current permissions. Free.
- Level 3, a paid AI model — which must answer in a fixed format that has no way to carry out an action.
- Level 4, reading documents and images — for documents that need it.
Understand, ask, fetch, explain.
Understanding the question
Works out what you’re asking, about which records, on which page. If the question could really mean two things, it says so rather than guessing.
Asking back
Asks the one question that clears up the confusion, rather than confidently answering the wrong one.
Fetching the answer
Uses a fixed set of approved, permission-checked tools to read your records. It can’t write its own searches of your data.
Explaining it
Answers in your own terms, with the records it read attached and every figure labelled with how it was worked out.
It can’t change your data or make up figures.
Whatever the AI model says is treated like a message from a stranger. It only gets into the system in a fixed format that has no way to describe an action. Any record it names is looked up again in your data, so its pointer is only a hint. Permissions are checked live, never taken from the model. It can’t search your data directly, change it, invent a calculation or claim something it can’t back up.
Wherever you ask, it works the same way.
The assistant panel, the command search, the ask box on lists and error messages, the suggestions on each page and the insight boxes all send the same kind of request. A typed question and a clicked suggestion are handled in exactly the same way.
- “Open sales order SO-10407” — found and opened without using an AI model.
- “What needs attention today?” — your work list, put together by your rules and summed up in a sentence.
- “Show me overdue invoices for this customer” — understood, looked up with an approved tool, and answered with the records it read.
- “Why is this on hold?” — the reason from that record, not a help article about holds in general.
- “Draft a credit note for the short delivery” — a real draft, not posted, ready for you to confirm.
Frequently asked questions.
Which AI model does it use?
The AI provider can be swapped without changing how anything works. Whichever model answers, it has to reply in a fixed format, and its reply is checked before anything acts on it.
Can I ask it to change something?
You can ask it to prepare something. It creates a real draft the normal way, unposted, and waits for you to confirm. Posting, approving and paying are outside what it can do at all.
Does it make up numbers?
It isn’t allowed to do its own calculations. Figures come from the rest of Cobalt ERP, and each is labelled with how it was produced: calculated, set by a rule, read from a document or written by AI.
What if it doesn’t understand me?
It asks you a short follow-up question. It would rather check with you than answer a question you didn’t ask, or guess which record you meant.
Does it cost credits every time?
Not every question uses external-model credits. Built-in checks and on-device processing do not use them. External-model answers and explicitly enabled external vision may use the AI allowance; local document text recognition uses the separate document-page allowance. Check the allowances in your plan.
Try it on your own records.
Book a demo and ask it a real question about your business. Every answer shows how it was worked out and which records it used.
Book a demo