From database restore to a question for a human. Every night. Every step traceable.
No demo pipeline. This is the chain that runs every night against a day-fresh copy of a real back office, eight steps, four gates, one exit.
Most recent full run, 17-09-2026: 639/639 monitored client files, 30 phases, complete in 89 minutes.
Working with AI, the human way
- 01Agents search Eleven agents, each with one task. Hours, invoices, work orders, absence, margin and terminations.
- 02Checks verify Before you get a question, the data and the reason for it have been checked. By 22 automated checks.
- 03You decide One question to the right person, with the source attached. Never an instruction. Every action waits for your approval.
The night cycle
8 steps · runs dailyThe full back office is restored as a day-fresh copy inside our own network. The system analyses reality as of yesterday evening, not an export from last month.
Documents and records are read and classified per object and period, in bulk on our own GPU. Personal data never leaves the network.
Before any analysis, recorded domain knowledge for that specific client is loaded, validated by the domain expert. Hard rule: the system is forbidden to ask a question the organisation has already answered.
Multiple models analyse independently; a pattern only becomes a signal when they agree. Output is enforced as strict JSON: reasoning text cannot leak into report fields. Every amount comes from code on our own hardware inside our own network, zero tokens leave the building for a measurement. The question to the manager is built from a fixed template, without a language model. More on cross-validation →
Sanity gates on every parse path: micro amounts, margins outside the bandwidth and empty actions are suppressed and logged. Such an artefact becomes a data check, never a signal to a manager.
Deterministic, before anything goes out: every action must be a question with an offer to help, addressed to a named person. Imperatives are blocked and returned for reformulation. AI never gives orders.
There is exactly one path to the outside world. Every message carries a mandatory destination and a verification block: amount plus source. No destination? Refused and logged. Nothing disappears silently.
Validated patterns are written back to the knowledge base at the end of each run. The next night, every agent knows them, step 03 reads what step 08 learned. Agents learn from each other, cumulatively, every night.
Since Platform 7.2 a counterforce rule applies on top: no change without independent AI review, and in doubt the human decides.
One night, measured
Figures from the night run of 17-09-2026. Measured, not extrapolated.
21 stopped cards are 21 questions a manager never had to waste time on. A gate that stops things is the product working. See client cases →
We asked for a closing date. There is none.
In practice, periods are never closed. More than half of a period's revenue is booked after the period has ended. A system that waits for a closing email waits forever. A system that trusts the calendar asks questions about figures that are not finished. So GuardPilot decides per client and per period whether a figure is ripe.
A period only counts after the back office's ready day. That is a company parameter, every company closes differently, and it is one of the four things re-established per connection.
If a draft or a pending invoice is still open for that client and that period, the period is not ready, whatever the calendar says.
If a client materially booked after the ready day in earlier periods, that client gets extra weeks. The window is per client, not portfolio-wide.
Code calculates and chooses. A template asks.
A language model makes mistakes, a sum does not. So the model never touches an amount. Code fetches the amounts, dates and thresholds and decides which signal it becomes and for whom. The question to the manager is built from a fixed template, without a language model. Language models are only used for reading, searching and summarising, never for the decision.
The model did not win. Fixed rules in code calculate and choose. The question to the manager is built from a fixed template, without a language model. Language models are only used for reading, searching and summarising, never for the decision.
What the mailroom refuses, and what passes
The knowledge circle is closed
A pattern confirmed at one client today is known knowledge for every agent tomorrow. And the answer a manager gives once, is never asked again.
Why you can trust this
After every change to the chain, a fixed set of ten clients, validated by the domain expert, runs again. Fails the set? The change does not ship.
A targeted test run is automatically routed to a separate environment: no state is persisted, no knowledge is written back. Test output cannot contaminate the knowledge base.
A hard reporting rule: no amount leaves the system without agent, file, timestamp and category. No trace, no publication, that rule also applies to this website.
Human or agent: every significant action is logged with actor and scope. Accountability is not a promise but a log line.
No black box. Every decision traceable. Every action waits for your GO.
See the engine room for yourself
The live dashboard shows what ran last night. The waitlist gets you the monthly under-the-hood updates.
We take on a limited number of engagements per year. The first companies build along and notice it. Join the waitlist
Alongside the agents, GuardPilot works on an organisation's history and context. Fourteen days of production — counts, cases, working rules and incidents — are described openly and without names.
View context and history →