עברית
Natural-language BI for Priority ERP

Ask your Priority a question. Get a live dashboard back.

In plain Hebrew or English. No data warehouse, no ETL, no report developer. The query runs inside your Priority, so the number you read is today's.

Live at bi.flow-chain-ai.com Hebrew and English, both native 481/481 tests passing

▸ You How much did we sell last month?

Topic: Sales Query validated Ran on your Priority

Sales — last month

live
Revenue
₪4.28M
+12% vs. prior month
Orders
1,214
+87 orders
Avg. order
₪3,526
+4%
Revenue by week
W1892K W21,140K W31,015K W41,233K

Two languages, one engine. Figures shown are illustrative; in your workspace they come from your Priority.

Why a simple number takes three weeks

Priority holds the answer. Getting it out is the project.

The report-developer queue

Every new number is a ticket behind other tickets. By the time the report lands, the meeting it was for has already happened.

Exports age on contact

A spreadsheet is a photograph of last Tuesday. It gets mailed around, edited, and argued over long after the ERP has moved on.

A general AI tool guesses

Point generic text-to-SQL at Priority and it sees thousands of cryptic table names with no map of which one means what. It answers confidently, and it answers wrong.

The fix is not a better guesser. It is a system that has read your install, and asks when it is unsure.

From question to dashboard

Four steps, and you watch all four happen.

1

Understand

The question is matched against the map of your install. The topic is detected automatically, or you name one: Sales, Finance, Inventory, CRM, HR, Manufacturing, Logistics, Projects, Procurement, Service.

2

Compose

One Priority query, written against the columns your install actually has — not the ones a general model expects to find.

3

Validate

Dialect and column inventory are checked before anything runs. A query that fails the check is corrected and retried, never shipped.

4

Draw

You approve the KPIs. The dashboard renders on live rows and stays clickable, down to the records behind each bar.

It asks instead of guessing

When a question could mean two things, you get a question back rather than a number. Ambiguity is never resolved silently — that is a design rule, not a setting.

It checks before it runs

Each generated query is validated against the real column inventory and Priority's own query dialect. Mistakes are caught on our side of the wire.

It says when a question is too heavy

A guard stops runaway queries before they load your server, and explains the fix in one sentence instead of an error code.

In the product “That query was too heavy to run. Add a filter…” · “Showing the first 5,000 rows of a large table. Add a filter or aggregate…”

It knows your Priority, not Priority in general

A one-time bootstrap of about thirty minutes reads your install and builds a graph database of it — including the tables, columns and forms your implementer added. Every question after that is answered against your map, not against a generic idea of Priority.

3,086Tables
4,641Forms
34,578Columns
11,390FK edges
576Concepts

Measured on a reference install. Your bootstrap reads your own Priority, so your map is your own — custom entities included.

Your customizations are on the map

Most Priority installs stopped being standard years ago. The bespoke table your implementer built for quality checks, the four columns somebody added to the order line, the form that exists nowhere else — the bootstrap reads them from your own metadata, so they sit in the graph beside the standard entities, with the same field meanings and the same join paths. A tool that only knows stock Priority cannot answer a question about them at all.

FORM AINVOICES resolves to TYPE = 'A' TABLE INVOICES foreign key · CUSTNAME header → lines TABLE CUSTOMERS TABLE INVOICEITEMS second hop TABLE PART Field and form help text read from your own install

Ask for A-type invoices and the map already knows which table that is, which filter makes it that kind, and how to reach the customer and the lines from there.

Forms, not just tables

4,641 forms mapped to the tables behind them, together with the filter that makes each one different. That is how an invoice question comes back with the right kind of invoice.

Join paths, already walked

11,390 foreign-key edges, header-to-line relationships, and second-hop routes. The joins are correct because the route was mapped in advance, not improvised at question time.

Your Priority's own words

The bootstrap reads the help text your install ships — 80.4% of forms carry it — and classifies every column as a code, a description, an amount, a quantity, a date, a status, or a flag.

While the map is being built “Building your schema — We are analyzing your Priority data. This can take up to 30 minutes — you can safely close this tab and come back later.”

It proposes the KPIs. You pick.

A dashboard is a short conversation before it is a picture. The assistant suggests four to six measures and says why each one belongs. Nothing renders until you choose.

In the product “Pick your KPIs in the chat — your dashboard will appear here once you approve.”
Compose dashboard

Sales, margin and stock cover for our top ten customers

Sales Margin Stock cover Top customers

One compound question, planned into four sub-questions, returned as four tabs. Six is the cap.

Illustrative.

Clickable, not a picture

Click a bar or a slice and the rows behind it open in place. The dashboard is the summary and the drill-down at once.

One compound question, one dashboard

A question with several parts is planned into separate sub-questions and returned as a multi-tab dashboard, up to six tabs.

Yours to take away

KPI cards, bar, line and pie charts, and tables. Export the rows as CSV, the whole dashboard as one self-contained HTML file, or the dashboard definition itself.

Teach it your words once

Every company has its own nouns. Define them once and every query after that respects them.

Business vocabulary
Dead stock

No movement for twelve months or more.

Applied to every question that mentions dead stock.
Revenue this month Open orders Dead stock value Gross margin %

Approved KPIs become one-click tiles at the top of the chat.

Illustrative.

A business profile in one step

Give it your website address or a paragraph about the company. It writes a summary and names the industry it detected. An administrator corrects anything it got wrong.

Terms and KPIs drafted from real tables

One click drafts up to 25 business terms and 15 KPIs. Every draft is grounded in tables that exist in your install; invented names are discarded before you ever see them.

Synonyms that route

Whatever your team calls a thing gets mapped to the form that actually holds it. Ask in house language and the query still lands on the right data.

It gets better with use, and you sign off on every change

A distill session replays a conversation together with the retrieval trace behind it, then proposes corrections to the map. Nothing is applied on its own.

Distill review — 3 of 7 proposals
Fix relationship High confidence

Link order lines to the part catalogue

Questions about part descriptions on order lines needed a second query to answer. One mapped edge removes that hop.

From the chat: “why is the part description empty here?”
Approve Skip

Every applied edit is audited and reversible.

Illustrative.

Every proposal cites its evidence

Each one arrives with a confidence, a rationale, and the exact line from the chat that prompted it. You are approving a change you can trace, not a black box.

Four kinds of correction

Fix a relationship, route a concept to the right form, bring back a form that was hidden, or stage a new business term or KPI. You approve them one at a time.

One click undoes it

Every distillation is audited as a unit, so a change that turned out wrong comes back out whole.

In the product “Undo this distillation (3 audited edits)”

One question, every company

Switch between the companies you are allowed to see, or put them in a group and ask all of them at once. The group runs in parallel and comes back as one table with a company column.

Company group — Europe & Israel
Ran on 4 of 6 companies — failed: FIN, UK. Retry failed
CompanyOrdersRevenue
IL6122,140,800
US3411,286,200
DE2881,004,500
NL173640,900

Illustrative.

Consolidation without a consolidation project

Six subsidiaries, one question, one table. The company column travels with the rows, so the split is visible in every chart built on them.

Partial failure stays visible

When one company does not answer, the result says which one and offers a retry for that company alone. A missing subsidiary never hides inside a total.

Send the answer, not the query

A published dashboard is a link. What sits behind it stays on the server.

Publish and copy a link

Anyone signed in can open it, with no workspace of their own to set up. Revoke the link and it stops working.

Put it on the wall

Fullscreen mode refreshes on a timer and shows the countdown, so a floor or lobby screen stays current with nobody touching it.

The viewer never sees the query

A published dashboard renders in a sandboxed frame. The query, the connection and the credentials behind it never leave the server.

In the product “Share link copied — anyone signed in can open it.”

Your data stays where it is

The shortest security review is the one where nothing was copied.

No copy, no sync, no lag

There is no warehouse and no nightly extract. Each question becomes one query that runs inside your Priority and returns rows. Nothing is staged, so nothing can be stale.

Priority credentials never touch the browser

Connecting runs through a gateway-hosted popup and a server-to-server token exchange. The page you are looking at never handles the password.

Your own model key

Each workspace supplies its own AI key, OpenRouter or Anthropic. There is no shared pool to fall back on and no cross-tenant sharing.

The workspace itself runs on Amazon Web Services, in Amazon's Tel Aviv region, with Cloudflare in front of every request: TLS, DDoS protection, a web application firewall, rate limiting and bot filtering. The servers accept no inbound connections — they dial out to Cloudflare and traffic comes back down that tunnel, so there is no public port to scan, and admin surfaces sit behind a second gate. Your AI key and the rest of the workspace secrets are held encrypted in AWS Parameter Store, never in a database or a config file, and deploys run from a versioned pipeline.

In the product “Used for AI query + dashboard generation. Stored securely; never shared.”
481 / 481Tests passing
4 tiersEval harness gating every release
1,235Commits through 7 August 2026
Livebi.flow-chain-ai.com

Getting connected

BI Generator is invite-only while the first workspaces come on board. Access is requested through the licensing portal, and registration takes about five minutes.

One workspace per Priority install

The first person to connect an install runs the bootstrap. Everyone after that joins the workspace that already exists and starts asking questions immediately.

Bring a file instead

Drop a JSON, CSV or Excel file and get the same dashboard treatment with no Priority connection at all. Useful for a one-off, and for trying the product on data you already have.

On-premises edition

A single-tenant Windows edition installs as a service for customers who keep everything inside. Its schema graph is built centrally and ships with the install.

Joining a workspace that already exists “Your organization already has a workspace here. Pick it and verify your Priority login — no other setup needed.”

Part of the Priority AI-Tools suite

One registration connects your team to Priority through AI. BI Generator reaches Priority through the same Priority Gateway every product in the suite uses, and the schema-graph engine it is built on ships into that gateway too.

Priority Console

Describe an app or a change to your Priority and get working Priority code back. Same sign-in as BI Generator.

chat.flow-chain-ai.com

Priority Gateway and the n8n node

The REST and MCP layer BI Generator queries through, plus a drag-and-drop n8n node on the same foundation for automation work.

Licensing portal

One registration provisions the whole suite. Tenants, users and access tokens are managed there in self-service.

licensing.flow-chain-ai.com