A chat window onto your own Priority. Ask a question and get the number from live data. Open a service call. Print an invoice. In Hebrew or English, without knowing which screen it lives on.
It is not a chatbot pointed at your ERP. It runs on our servers under your account, it acts as the person who is signed in, every change stops for a yes, and it cannot alter how your Priority is built.
One question, one record created, one document produced — the three things people actually do all day. Nothing below is a mock-up of a different product: this is the shape of a real turn.
How many open orders do we have this month, per customer?
Twelve open orders this month, across four customers.
| Customer | Orders | Value |
|---|---|---|
| Acme Ltd | 5 | 142,300 |
| Bartel Systems | 4 | 88,150 |
| Cohen Metals | 2 | 31,900 |
| Delta Foods | 1 | 9,400 |
Filter applied: order date 01/08/26 – today · status not Closed or Cancelled
It says which filter it used, every time, so you can correct it in one line instead of wondering what it counted.
Open a service call for Acme: printer jams on tray 2.
It finds the customer, fills the form the way you would, and then stops. Nothing is written until you press the button.
Allow once — the only two buttons on the card. Approving this one does not approve the next one.
After you approve, it reads the record back: “Created service call SC24-0912 for Acme — status Open.” You get the number, not a promise.
Print invoice T2400123 as a PDF.
It finds the print report bound to that document, shows you the parameters it will pass, runs it once you say go, and hands back the file.
T2400123.pdfSame for the reports your team already has. Excel for the tabular ones, PDF for documents — it asks which if you have not said.
Ask for a trend or a chart and it will answer the number, then point you at the BI workspace — charts are that product’s job, and it says so in one line rather than drawing you something wrong.
The difference between an assistant that answers and one that guesses is whether it has a map. This one does.
Before it answers anything with a number in it, it looks the question up in a map of your own installation — your tables, your fields, and the relationships between them, including everything your implementer built for you over the years. It comes back with the right tables and how they join.
That map is built once, in about half an hour, and it is the same one behind the BI product. Nothing is copied out of Priority to make it: it records where things are, not what they say.
This is the part generic AI tools cannot do. Pointed at a Priority database they see thousands of cryptic table names and guess — and a plausible wrong number is worse than no number.
The map also knows your vocabulary. It knows the customer invoice table is not the tax invoice table, and it knows which status codes mean “open” on your system rather than assuming. When a question is genuinely ambiguous it asks which figure you meant instead of picking one.
Figures from one reference installation, custom tables and custom fields included. Yours will differ — that is the point. Priority changes and the map is rebuilt; it is never a stale copy of last year’s schema.
Three jobs, and they cover most of what anyone does in Priority in a day.
Totals, counts, rankings, breakdowns, one record, or a date range. It reads live — there is no nightly copy to go stale.
Create, update and close records — always through the Priority screen, so your validation, defaults and business rules apply exactly as they would if a person typed it.
The procedures and reports your team already has. It finds them, shows you the parameters, and runs them after you confirm.
Not policy, not a prompt asking it nicely. Each of these is a limit in the software: the capability is absent, so there is nothing to talk it out of.
Each create, update, delete or procedure raises its own card naming exactly what will happen, and the card offers two buttons: allow once, or deny. There is no “approve all for this task”, no per-screen blanket, and no setting anywhere that turns approvals off — on this tab those options are not built. Deletes always ask on their own, never bundled with anything else.
It reads with a query language that is read-only by construction — the write commands are not merely blocked, the channel does not carry them. Every change it makes goes through the Priority screen instead, which means your mandatory fields, defaults, business rules and warnings all still apply. There is no path that bypasses them.
No new screens, fields, tables, business rules or code. The development tools are not merely forbidden — they are never loaded into the conversation at all, and the request is refused before it reaches them. Ask for one and it says so in a line and points you at the Dev tab, which is a separate, separately licensed product.
No shell, no reading or writing files on any machine, and it cannot spin up helpers of its own. It has one way to reach Priority — the gateway — and that path is the only one, for every read and every write. What it did is in the transcript.
The usual worry about AI and an ERP is people pasting company data into a public chat window. This is the other shape: your data stays on a path you control, and the person’s Priority password never leaves it.
Connecting opens Priority’s own sign-in window. What comes back is a token, held encrypted on the server and never handed to the page. There is no password box in the product, and nothing to leak from a browser.
Each user connects their own Priority identity, so everything it does in Priority is done as them, under their permissions, and shows up in Priority as theirs. One person can be cut off in seconds without touching anyone else. Reading can be held to each person’s own authorisations too — a setting per Priority system.
No warehouse, no nightly extract, no second copy of your data sitting on our side. Every answer is read live from your Priority at the moment you ask it.
The servers sit in Amazon’s Tel Aviv region. Passwords and keys are kept in Amazon’s encrypted store, never in a config file, and deployments come from a versioned pipeline.
Cloudflare sits in front of every request. The servers themselves accept no incoming connections — they dial out and traffic returns down that line — so there is no public door to find. Admin screens sit behind a second lock.
There is no server of ours to stand up, patch or monitor. Your Priority can sit in Priority’s cloud or on your own network — we reach it from here, over a connection you approve and can withdraw.
Every conversation is kept, so there is a record of what was asked and what was changed, per person. Worth saying plainly: the assistant reads your Priority data in order to answer, and a model provider processes that text to produce the reply — the point is not that no data moves, it is that it moves along one path you can see, under one identity you can revoke, with every change requiring a person to press a button.
Chat is the tab everyone lands on. The developer products next to it are the ones you buy per person.
One invite link, a five-minute form, and one approval on our side sets up the sign-in, the key and the workspace together. Welcome instructions go out in English and Hebrew.
The first time they open the console they press Connect, sign into Priority in its own window, and that is it. No password typed into our product, no token to copy anywhere.
Every person on the account has this tab, with no separate licence and nothing to switch on. Dev and App Generator sit beside it for the people you grant them to.