Brief · Support agent exercise · For João
An explanation of the current process, the numbers from the 2025 dataset, and the agent this exercise builds. No engineering detail.
Every ticket in the queue is handled by a person. In 2025 that meant 7,820 tickets read, researched and answered by twelve agents, with up to 20 new tickets arriving per minute during the peak after a tax filing deadline.
Answering one ticket requires two things that live in different places.
The first is the customer's account. It sits in the operational database, lake.db, which exposes companies, contacts, advisors, plans, subscriptions, add-on contracts, orders, invoices and their line items, payments, refunds, credits, tax deadlines and the full history of earlier tickets. Anyone on the team can query it.
The second is a price, and prices are not in lake.db. Each of the 100 partner advisors publishes their own price list as a PDF. Each list has its own layout, its own names for the same service and its own way of stating whether a price is net or gross. The folder holds 232 of these files.
After that the agent decides whether a credit or a discount applies, writes the reply, and either closes the ticket or sends it to a colleague. When the customer writes back, the ticket returns to the queue and the same steps start again.
The median first reply in 2025 was 38.8 hours, and the average was 45.2. On the Monday after a filing deadline the queue is longer than the team can clear that day.
The 232 PDFs were written by 100 different firms. The same service appears as Lohnabrechnung, Entgeltabrechnung and Gehaltsabrechnung pro MA. Some list net prices, some gross, some charge per hour, per employee or per filing. The folder contains duplicates and superseded versions, 38 files are scans without a text layer, and the longest document has 132 pages. Nothing marks which file is current.
Some services are quoted as "auf Anfrage". In 2025 the team sent 845 questions to advisors. The average answer took 6.4 days, one in ten took longer than 14 days, and 112 were never answered at all.
The team granted 145 credits in 2025: 83 at 10% for six months and 62 at 15% for one month. The same complaint could end with a different amount depending on who picked up the ticket and what they remembered.
196 of 800 companies cancelled during 2025. The cancellation rate tracks how quickly the customer's advisor answers. Customers of an advisor who answers in under two days cancelled at 12.9%. Between two and five days, 18.4%. Above five days, 42.7%. The plan, the software and the price of the plan are the same in all three groups.
811 tickets in 2025 concerned another company's data, someone impersonating an owner, a phishing email, a racist message or a report of discrimination. Each one has to reach a person, and the route it takes today is also undocumented.
"I have asked three times what payroll costs. The first answer said 28 €, the second said 40 €, and the invoice came to 46.50 €. Which one is right?" Owner of an electrical installation company with 12 employees, March 2025. She cancelled in May.
A new agent learns the plans in about three weeks. Learning the folder takes longer, and the exceptions are learned by asking a colleague. For most tickets the agent already knows what to do. The time goes into locating the document, checking that it is the current version, and waiting for an advisor to answer a question that only the advisor can answer.
Adding people shortens the queue. It does not remove the wait for a document or for an advisor, and it does not make the credit decisions consistent.
A program that reads the account from lake.db, selects the advisor's current price list, applies the written policy and writes the reply. It cannot quote a price without the source document, so the number in the reply can be traced to the page it came from.
The 232 PDFs become 100 Markdown files. Each file keeps the current list, removes duplicates, newsletters and old versions, and states the unit and whether the price is net or gross. Every line keeps a reference to the PDF and page it came from.
The rules the team applies today, with limits and numbers: who receives a credit, how much, how often, and who receives nothing. I will test each rule against the 2025 history before it goes into the agent.
Requests for another company's data, impersonation, phishing reports, racist messages and reports of discrimination go to a human, with no reply sent to the customer in the meantime. The classifier runs before the model sees the ticket.
A ticket counts as resolved only when the customer says they need nothing more. The customer's next message is what closes it, so closing a ticket to improve a number does not work.
Every reply is compared with the facts behind the ticket. Categories are released one at a time, starting with the ones where the agent's accuracy passes a threshold we agree on. Until then the agent drafts and a person approves.
| Today | With the agent | What it changes |
|---|---|---|
| First reply after 38.8 hours (median) | Minutes for the categories the agent handles | Fewer cancellations caused by waiting |
| Price read from a PDF by hand | Price taken from the current list, with the source page | Fewer wrong invoices and fewer refunds |
| Credit decisions from memory | Written rules applied the same way for every customer | Predictable cost and fewer exceptions |
| Advisor quality invisible | Report per advisor: complaints, late answers, cancellations | Data for renegotiating or replacing a partner |
| Security cases handled by whoever picks them up | Fixed route to a human with a written procedure | No automated reply to a fraud or data request |
Most of what the agent needs is not written down. It lives with the twelve agents, in the correspondence with advisors and in the 232 PDFs. Extracting it is a joint job, and it has a fixed shape.
| What we need | How we get it | What your team approves |
|---|---|---|
| Credit and discount rules | Two sessions with the agents who grant them, then each rule tested against the 2025 history | The final thresholds: who gets a credit, how much, how often |
| The churn signals | The team lists what makes them think a customer is about to leave; we check which of those signals actually came before cancellations | The list of signals and the action for each one |
| Which price list is current | One named owner per advisor confirms the valid document and its start date | The current list per advisor, including the advisors with no valid list |
| Consolidation rules | We propose what to drop (duplicates, newsletters, superseded versions) and how to normalise net, gross and units | The rule for each ambiguous case |
| Sample check of extracted prices | Your team reviews 20% of the extracted prices, weighted toward scans and "auf Anfrage" | Sign-off before the agent quotes a price |
| The pilot group | We rank advisors by complaints, late answers and cancellations | The advisors and customer segments in the pilot |
The rules are commercial, so they cannot come from engineering alone. The agent will be as good as the rules we write down together. The harness, the consolidation pipeline, the security layer and the evaluation we build and bring back with numbers attached.
| Week | Deliverable | Your part |
|---|---|---|
| 1 to 2 | Database access and the consolidated price lists, with the PDFs reviewed | Review 20% of the prices |
| 3 to 4 | Harness and the security tests | None |
| 4 | Agent against the 2025 history, on the same tickets a person handled | Read the result with me |
| 5 to 6 | Shadow mode: the agent drafts, a person approves | Weekly review of the edits |
| 7 to 8 | Autonomy for the categories with the best record | Choose the categories |
| 9 | Final comparison and the first advisor report | Conversation with the partners |