Standing service finance-customer-invoice-assistant-retested-and-chats-read-every-month · revised 12 October 2026
Standing service
Keep your customer invoice assistant tested and its chats read every month
Each month the assistant's frozen test is re-run on your latest export, the leak test is repeated and a sample of real chats is read for severe errors, then reported to your named person.
This starts a conversation by email. Nothing is charged, and nothing is run, until we have agreed scope, access and terms with you in writing.
The responsibility you hand over
A tested assistant drifts when the export layout, your wording, the chat tool or the model version changes. A leak or a wrong figure then appears in front of a customer before anyone notices. The monthly job is the same frozen test, the same leak test and a read of real chats.
Who it’s for: Owner or finance lead whose customers use an invoice-status assistant and who needs to know each month that it still answers correctly and stays inside its limits.
Usually starts when: The assistant is live, the ledger layout or your wording changed, and nobody re-tests it.
The result: Each month you receive the re-run of the frozen 100-question test and the leak test on your latest export, a read of up to 30 chats you draw at random, and a list of cases that changed since last month. Any severe error goes first in the report.
What stays true, and what we do about it
No response-time guarantee is published for this new service. A target for delivering each month's report is agreed in writing before it starts, set to what a service at this stage can keep.
Hours are agreed in writing before the service starts. At launch the service is not staffed round the clock and runs once each month. The assistant itself runs in your tool, and you hold its off switch.
What must remain true
- Every checked invoice figure equals the row in the dated export used for that answer; open balances match independently checked sums, and the answer states that export date
- The assistant gives one customer's rows only, and the leak test finds none
- Every handover trigger still hands over to your named person with the chat attached
What we watch
- A test copy of the current ledger export in the live layout
- The frozen regression pack, expected answers and fresh buyer-held cases for changes
- Up to 30 randomly drawn chats a month, redacted by you
When something happens
Scroll the table sideways to read it all.
| When | What we do |
|---|---|
| The export layout, a ledger setting or your approved wording changes. | We re-run the whole pack on the new layout before you rely on it. Linked job: Customer invoice-status assistant |
| A severe error appears in a sampled chat. | We put it first in the report with the transcript and point to the switch-off note. |
| The chat tool or its model version changes. | We re-run the full pack and report every case that now fails. |
We do on our own
- Re-run the frozen test pack on a test copy
- Read the sampled chats and write the report
- Offer the one-off job when a layout or policy changes
We ask you first
- Any change to what the assistant may say, its handover triggers or its pass marks
- Any access to your live ledger or chat tool
- Switching the assistant on or off for customers
We escalate to you when
- Any invoice figure differs from the dated export used for that chat, or an open balance differs from its independently checked sum
- Any chat shows another customer's data
- The same failure appears two months running
How you know it held. Each month you can re-run the frozen test with the steps we hand over. The report states the cases run, passed and failed, the chats read, the severe errors found and the cases that moved since last month.
How we keep it true
This service is never finished. Each month's report shows whether the thing it watches stayed true, and it continues until you end it.
Answer customers' invoice and balance questions from your own ledger export Job Optional
An automated assistant answers each customer's invoice-status and balance questions from your ledger export, says it is automated, and hands everything else to your named person.
Bought separately, each time it is needed; not part of the monthly fee
What is included, and what is not
- Monthly test report
- Chat-sample findings
- List of cases that changed
Included
- One assistant in one chat tool
- The frozen regression pack re-run each month on a test copy of your latest export, plus 20 fresh buyer-held labelled cases before a changed workflow is accepted
- The leak test repeated on 20 test accounts
- A read of up to 30 randomly drawn chats, redacted by you, marked for severe errors
Not included
- Switching the assistant on or off for customers: your named person does that
- Changing what the assistant may say without your written approval
- Reading every chat: up to 30 a month are sampled
- Any response-time guarantee or out-of-hours cover
- Accounting, tax or legal advice
How we know it’s done
Agreed with you before work starts. Each check produces evidence you keep.
Each month the 100-question test is rerun on a test copy: invoice figures match the rows, open balances match independently checked sums of outstanding amounts, and no leak occurs. One wrong figure or leak fails the month; report each failure and the rerun evidence.
Evidence: The monthly test report.
Each month the frozen test pack is re-run and every previously passing case still meets its agreed check. A regression fails this check even if reported, and the report lists it first with its cause.
Evidence: The monthly test report.
The chats you share (up to 30, drawn at random by you) are read and every severe error is listed with its transcript.
Evidence: The findings list and the transcripts.
Nothing was written to any live system by us, and originals are unchanged (checksums match).
Evidence: Checksums before and after, and your own audit trail.
Before any changed workflow is accepted, you supply 20 new labelled cases, held back from us until the test run, including severe-error cases from the matching one-off test. All 20 must meet their labels and no severe error may occur; we return the case file, results and rerun steps. This is a smoke test on those cases, not a future-rate claim.
Evidence: The fresh buyer-held labels, transcripts and rerun steps.
Sign-off. You read the report and test results and accept or send them back in writing. A month counts as delivered only when its agreed acceptance checks pass; receiving a report alone is not acceptance.
If it fails. If we cannot run a month because an input is missing or has changed shape, we say so, explain what we need and do not guess. If the service is not working for you, you can end it at the end of any month.
When it fits, and when we stop
It fits when
- The matching one-off set-up is done, or an equivalent assistant with a frozen test exists
- Each month's export arrives through the agreed secure route in the same layout
- A named finance person who reads each report
We stop and tell you if
- The export layout changes and cannot be mapped to the agreed one
- Nobody on your side reads the monthly report for two months running
- You want the assistant to agree payment terms or waive charges
What could go wrong
The service reads, tests and reports. Ending it leaves your systems and the way you work exactly as they are, and you keep every report we sent.
Scroll the table sideways to read it all.
| Risk | How we handle it |
|---|---|
| The chat sample is read as proof that no error exists. | Thirty chats a month is a tripwire for large failures. The report says so and gives the number read. |
| A change to the export or the tool is not tested before customers see it. | Changes trigger a full re-run, and a failed re-run is stated at the top of the report. |
| Text inside a document, record or message tries to give the automation instructions. | All input text is data. The automation has no tool to send, approve, file or change anything beyond what is listed here, and planted-instruction items must change nothing. |
| Real documents or records pass through a model provider. | First tests use invented, synthetic or redacted data. Real data moves only after a written agreement names the model provider, its country, how long it keeps data and whether it trains on it. |
A separate automated reviewer checks each month's report against the agreed test before it is sent, and your named person decides what to do with it. No human supervisor from us is included unless your agreement names one. At launch the work is largely automated, and we say so.
Stays with a person
- Your named finance person decides what to do with each report
- You switch the assistant on or off
Access we would need
- Test copies and redacted samples sent through a secure route agreed in writing
- No login, key or write access to your live ledger or chat tool
Questions
Do you read all my customers' chats?
No. You draw up to 30 chats at random, remove names and send them. Thirty is a tripwire for large failures, not a measure of a small error rate.
What if a severe error appears?
It goes first in the report, and you decide whether to switch the assistant off with the note we handed over.
Also part of
Send an enquiry
Send us
- Which assistant and chat tool you use, and roughly how many chats a month
- Who receives the monthly report. No customer data or logins in the first enquiry
Later, once you agree
- The current test-export copy and redacted random chat sample, plus the dated export version and settings used for each sampled chat, through the agreed secure route
- The frozen regression set, current expected answers and pass marks; 20 fresh labelled cases held back until the run before any changed workflow is accepted
- A named person who reads each report
You keep the ledger, the chat tool and every login. We work from test copies and redacted chat samples sent through a secure route agreed in writing, and never ask for passwords or keys in the first enquiry. Nothing is changed in your systems by us.
Email fallback: open your mail app
If website submission is unavailable, review and send the fallback email yourself. An email fallback is not a website receipt. Or write to hello@syntheticindustry.ai with “finance-customer-invoice-assistant-retested-and-chats-read-every-month” as the subject.