A supply chain analyst builds a text-to-SQL data specialist by conversation — every skill run live and confirmed before it answers anyone
An analyst builds a new data specialist by conversation in minutes, with no code.
See one case, screen by screen ↓Every new question — how many loaner kits are overdue at each hospital, why returns jumped last month — means filing a request and waiting in the engineering queue. The analyst knows which tables hold the answer, how they join and how the business counts. The engineer who builds it does not, so the analyst explains it all again, and checks the numbers when they come back.
Meanwhile the work goes on in spreadsheets someone rebuilds every Friday. And a query written without the analyst’s rules looks right and is wrong: count every return without a closed date as “open” and you include cancelled authorisations and goods that never arrived.
The 31 days in the queue is the median the working solution’s dashboard reports; the other rows are our estimate for a typical new request.
Builder agents interview the analyst and test every skill live; the analyst confirms each one.
Lena Ortiz, a supply chain analyst, builds a Returns specialist on the supply chain warehouse. Here is what happens, screen by screen, in the working solution — 10:31 to live at 10:45.
Nine live specialists answering in the supply chain assistant, 26 skills verified by people, 1,395 questions answered last week. Under “Needs you”: her own Returns draft, a Backorders skill whose result moved 14% overnight, Dana Okafor’s request for a loaner-kit specialist, and Shipments & carriers v1.4 ready to publish.
“I want a Returns specialist. It should answer questions about customer returns — how many, why, what they are worth, and how long they stay open.” The builder searches the supply chain warehouse, finds 7 tables that look related and highlights the likely ones: returns_header, return_lines, reason_codes, shipment_lines, item_master, customer_accounts.
“Which tables should it use, and how do they join? Anything about how you count a return helps too.”
Lena names the joins — rma_no, item_code, ship_to_id — and says to count a return by line and value it at standard cost. The builder checks all 5 joins against the warehouse, with no duplicate rows from return_lines to item_master, draws the six tables and writes her rules down. It adds one she did not say.
“reason_codes marks DEF — suspected product defect as a complaint, so I’ll report those as totals only, never case or customer detail.”
Returns by reason code: 1,126 lines and $1.84M for EMEA in September — the totals match the September returns report. Return rate by product family: four families, all with shipments in the window. Lena confirms both. The third, open returns past 30 days, came back with 212 rows — and a question.
“Check this: 41 of the 212 are cancelled authorisations (status CXL), and 63 were never received — the clock starts when the return is authorised. Is that what you mean by open?”
“Leave out cancelled RMAs, and start the 30-day clock when the goods are received.” The builder re-drafts the query and runs it again: 87 returns open past 30 days from receipt, the oldest RMA-04418 at St. Aldric Hospital, 71 days. Her fix becomes a rule of the specialist, and she confirms the skill.
Three of three skills confirmed, one after her fix. Going live saves Returns specialist v1.0, adds the skills to the skill library with their confirmation records and joins the supply chain assistant’s routing. Three probe questions route as expected: two to Returns, one that stays with Backorders.
“Live in the supply chain assistant’s routing · 3 verified skills · built in 14 min”
“Which reason codes drove EMEA returns value in September?” The assistant routes it to the Returns specialist at a 0.94 match, on the skill Returns by reason code: $1.84M at standard cost across 1,126 lines, short-dated or expired stock leading at 38% of value, 61% of those lines from distributor accounts. Suspected-defect returns appear as totals only.
Every answer carries “Verified skill · confirmed by Lena Ortiz” and “See the query”. It opens the exact query that ran, its parameters (EMEA, September), 0.9 seconds and 7 rows, with the lines that produced the cited number highlighted.
“Runs on the supply chain warehouse through a read-only connection; customer contact columns are hidden from every answer.”
Every live skill re-runs each night. Backorder age by customer moved 14%: 186 backorder lines were re-dated in the warehouse on Oct 5, so the oldest wait moved from 36 to 41 days. The skill waits for its owner, Sam Patel, to run it live and confirm. The re-run checks the re-dated lines: promise dates moved by the order desk — a real change, not a data fault.
The same specialists, seen by the four people who build, sign and ask for them — what their week looked like, and what it looks like now.
Builder agents interview the analyst and test every skill live; the analyst confirms each one.
Interviews the analyst — purpose, tables, join keys, rules, worked examples — reads the warehouse catalog, drafts 3–5 skills as question-and-query pairs and assembles the new specialist.
Runs each drafted query live and checks the result for cancelled records, duplicates, nulls and suspicious joins, explaining anything a domain expert would question in one sentence.
Receives every question and routes it to the specialist whose purpose and verified example questions match best, or asks the person which area they mean.
Re-runs every live skill each night and asks the owner for a re-check when a result moves beyond the set limit.
Confirms every skill before it goes live. The agents propose; a named person decides.
Builders and leads can ask about any specialist or skill — or tell it what to change. Changes are recorded as theirs.
Which specialists answered the most questions last week?
Inventory & stock positions answered 412 questions, then Backorders with 288 and Shipments & carriers with 231. Together they took 67% of the 1,395 questions answered from Sep 28 to Oct 4.
Show me the skills that need a re-check
One skill needs a re-check: Backorder age by customer in the Backorders specialist. Its result moved 14% overnight because 186 lines were re-dated in the warehouse.
How does the assistant decide which specialist answers?
It compares the question with each specialist’s purpose and verified example questions, and routes to the best match when it is at least 0.80. Below that it asks the person which area they mean. New specialists join the routing the moment you confirm them — there is no deployment step.
Add a rule: never show customer contact names in an answer
Added to Rules every specialist follows. It applies to every live specialist from the next question, and it is recorded in the audit trail as a change by you.
13 screens from the working solution, on its sample data. Pick one to see it large.
Who confirms, what specialists may read and how questions are routed are settings — changed in one place, recorded in the audit trail.
| Setting | Default | Choose from |
|---|---|---|
| Who confirms a new skill | The analyst who builds it | The analyst who builds it · The analyst and a data steward · A data steward only |
| Skills that read cost or price tables also need | Omar Haddad · Data steward | Omar Haddad · Data steward or Dana Okafor · Supply chain lead |
| Ask for a re-check when a result moves more than | 10% | 5% · 10% · 20% |
| Most rows a query may return | 50,000 | 10,000 · 50,000 · 250,000 |
| Run times out after | 30 seconds | 30 seconds · 60 seconds · 2 minutes |
| Hide customer contact details | On | On · Off |
| Route a question when the match is at least | 0.80 | 0.70 · 0.80 · 0.90 |
| Rules every specialist follows | Four rules, each with who added it | Add rule, in plain words |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample supply chain data · “target” = the design goal, measured in the live solution · “estimated” = our estimate. People, companies and products named on this page are fictional — characters and sample data in the working solution.
An assistant that answers one area of the supply chain — returns, backorders, shipments, supplier delivery — from verified skills. Each skill is a question-and-query pair a person confirmed on live data, and the supply chain assistant routes each question to the specialist that matches it best.
By conversation. The analyst says what it should answer, names the tables, join keys and counting rules, and the Specialist Builder drafts the skills. The Skill Validator runs each one live, and the analyst confirms it, fixes it in plain words or sets it aside. Going live adds it to routing with no deployment step.
No. Specialists read the warehouse through a read-only connection — SELECT statements only, with a row cap and a timeout you set. Customer contact details are hidden from every answer.
Every skill was run on live data and confirmed by the person who knows the data. Every answer shows the skill, who confirmed it, and the exact query that ran. Accuracy is measured on each specialist’s verified examples.
The supply chain assistant compares it with each specialist’s purpose and verified example questions and routes it when the match is at least 0.80. Below that it asks the person which area they mean.
Every live skill re-runs each night. When a result moves more than the set limit — 10% by default — the owner is asked to re-check it on live data and confirm the new result.
Yes. Nothing answers anyone until a person confirms it, skills that read cost or price tables also need a data steward’s sign-off, and every version keeps the conversation behind it and can be restored in one click.
The Agentic Solution Engine builds and deploys it from your requirements — your warehouse catalog, your counting rules and the questions your teams ask — and it goes live once every quality gate has passed. From there, your analysts build each new specialist themselves.
Ask any supply question in plain words and get a traced answer in seconds.
1-min filmSupply Early WarningEvery supply signal read as it lands, and a ranked, cited shortage alert raised weeks before stock runs out.
1-min filmTemperature ExcursionsEvery temperature alarm becomes a cited release-or-reject decision in hours, with every earlier leg counted.
1-min filmWe’ll build a specialist with your analysts, on one area of your own supply chain data.