Spreadsheet tangle → governed system
The problem
In many mid-market firms the system is a workbook and a macro: statuses that are only colours, lookups into a hidden sheet, and one person who understands them. Moving onto real software usually starts with weeks of interviews about what the workbook means.
How it's resolved
Tarnbrook Facility Services, a fictional facilities contractor in Harrisburg with four subcontractor crews. Its operations manager’s workbook holds 40 jobs whose status is only a fill colour, lookups into a hidden sheet of sites, and a Recalc macro that works out what each crew is paid. The AI reads every sheet, formula and fill, and the macro’s source, and proposes a model: jobs become orders from a crew to a site, the rate card a catalog, crews sellers with day capacity, the macro the crews’ terms. Code holds every claim against the cells: 33 of 33 columns accounted for, 120 of 120 lookups, and 24 of 24 payouts reproduced to the cent. The owner approves before anything is written. The load runs through the API: the rate card through the platform’s own import, crew by crew; the 40 jobs as orders dated today, each keeping its original date and source cell; and upcoming jobs booked into their crew’s day. Each crew’s payable equals the macro’s payouts, $3,869.26 in all. Click any record, and a thread runs back to its cell. An agent then asks which of crew B’s jobs are overdue, through the MCP server, and code checks its answer against the system.
What you'll see
- 01Two files and an empty system: a 40-job workbook whose statuses are only colours, and its Recalc macro
- 02The AI reads every sheet, the hidden one too, and the macro, and draws the model
- 03Code checks it against the cells: 33 of 33 columns, 120 of 120 lookups, 24 of 24 payouts to the cent
- 04The macro’s rule becomes the crews’ terms: 18% of each job and a $35 call-out fee
- 05The owner approves; nothing was written before
- 06The rate card loads through the platform’s import, and a rate line threads back through its staged row to its cell
- 0740 jobs replay as orders dated today, and a request threads back to its row in the workbook
- 08Every crew’s payable equals the macro’s payouts: $3,869.26, to the cent
- 09An agent names crew B’s two overdue jobs through the MCP server, and code checks the answer
Under the hood
Each business term in the video, and the Traide building block it runs on.
| In the business | On Traide |
|---|---|
| The rate card | CSV import, one file per crew: each staged ImportRow links to the product it became |
| Leftover columns | Metadata from the import itself: public, or private for internal notes |
| The Recalc macro | Agreement: an 18% commission and a FIXED $35 fee on each order |
| A crew | Seller with a warehouse; its days are capacity pools the dispatcher holds |
| A past job | Order dated today, with its original date and source cell in metadata |
| The agent’s question | Traide MCP server: search-orders, reading public order metadata |
Guardrail. Nothing is written before the owner approves. Code checks every column, colour and rule the AI names against the workbook itself, and refuses a model with the cells it got wrong. Past jobs are dated today, never back-dated, with their original dates in metadata. Internal notes land in private metadata. Test money.
- Schema inference
- Verification against the source
- Tool use
More demos
- Workflow orchestrationDiscovery call → priced SOW → accept and pay
- Workflow orchestrationDay-1 ready: one approval, four teams
- Product developmentA catering order by text, in the customer’s language
- Workflow orchestrationThree-way match on autopilot
- Commerce and marketplace buildsAgent-ready tours and stays
- Product developmentAn activity guide → accessible, translated registration
- Product developmentA 13-week cash forecast that explains itself
- AI adoption and enablementTickets that fulfil themselves
Have a problem like this one?
You'll talk to a senior engineer, not a sales rep, and leave with a clear first move — whether or not we work together.