Nonprofit finance staff ask in Teams or Microsoft 365 Copilot: “what did we spend last quarter?”, “who holds cards?”, “write a donor follow-up for Q1”. The agent pulls their own card data, does the arithmetic in code, and writes a table or a three-paragraph donor email using only those figures.
“Estimate” numbers are engineering estimates to confirm with users. All others are counted from the code and the deployment artefacts.
Nonprofits that run program spending on prepaid cards owe their donors an account of where the money went. Producing one meant exporting transactions, mapping each to a category, adding up totals and then writing prose that a finance-literate donor would trust.
The card provider wanted that available where staff already work, in Teams and Copilot, without the provider holding a shared credential to every customer’s account, and without a language model “estimating” a single number.
The diagram is wider than the screen: swipe sideways to see all of it →
The model receives aggregates and a flattened transaction list, never raw API payloads to add up, and is told never to invent or estimate a figure. An empty date range is answered by code without calling the model at all.
Totals, category sums, top purchases and “untagged” counts are computed in TypeScript and handed to the model as data. A donor email can be wrong in tone, but not in its numbers. Trade-off: new kinds of analysis need code, not just a prompt change.
No shared service credential to customer accounts. The token is entered in a masked field and validated before use, so a bad token fails at connect time, not halfway through a question. Trade-off: for this pilot the token lives in server memory, and moving it to a secure store is the first production item.
The REST API was ready. The provider’s MCP server was the strategic direction. Both paths sit behind the same functions and one setting chooses between them. Trade-off: two response shapes to normalise.
“List” gets a markdown table sorted by date. “Summary” or “donor follow-up” gets exactly three paragraphs with no bullets, ready to paste into an email. Factual questions get one plain answer.
Most of the issues only appeared inside Teams and Copilot, not in local tests. A selection:
Copilot and Teams send invoke activities alongside messages. With no handler, the adapter answered 501 Not Implemented, and Copilot treated the agent as faulty.
Explicit 200 invokeResponse for invoke activities, silent acknowledgement for events, and an error handler that never throws on non-message turns.
The MCP server runs behind a load balancer. Every call after initialize landed on another instance and lost its session.
Capture every Set-Cookie (including the affinity cookie) and send them back on each call. Parse both SSE “data:” frames and plain JSON replies.
The transaction tool’s name was not settled on the server side, and a wrong name came back as a tool error, not an exception.
An ordered fallback over four candidate tool names that stops at the first success and returns an empty list otherwise, so a summary never crashes.
Some transactions had no tag, so category totals silently disagreed with the overall total.
Untagged spend falls back to the platform category, then to the merchant name, and the count is passed to the model as a warning that it must mention.
Left to the model, an unspecified period becomes a guess presented as fact.
A deterministic parser for common periods. If the message names no period, the agent asks a clarifying question instead of calling the API.
Follow-ups made no sense without the previous turn.
The classifier rewrites each message to stand alone, using the last 10 turns, before it is routed.
An illustrative replay with synthetic amounts and merchants. It contains no client data or code. Each scenario shows what the agent does before it answers.
| Measure | Before | After | Source |
|---|---|---|---|
| Donor spend update | Export, tag, sum and write, ≈60 min | One question, under 1 min | estimate |
| Round-trips per summary | 3 sequential | 1 parallel | derived |
| Copilot invoke handling | 501, agent marked broken | 200 acknowledged | code |
| Figures computed by the model | Possible | 0, all in code | code |
| Bad credentials detected | Mid-question | At connect, with a fix-it message | code |
| Deploy iterations to a working package | — | 7 in 3 days | artefacts |
Source: code = in the repository; artefacts = timestamps of the deployment packages; derived = follows from the code; estimate = to confirm with users.
Free scoping call · 30 minutes · No commitment
Write a few lines: the process or system, what hurts today (time, errors, cost or an AI pilot that stalled) and who uses it. After a short call you get a written plan and a fixed quote within ≈3 days.
Opens a pre-filled email · reply within one business day
The model reads. The code counts.
All projects