
Finding the Automations Nobody Registered
Nobody deploys shadow automation deliberately. It arrives when a feature you already paid for gets switched on, and a process quietly starts depending on it.
- Most unregistered automation is not rogue. It is capability enabled inside tools the organisation already owns.
- Voluntary registration misses it, because the people using it do not think of it as an AI system.
- Discovery works better from four angles: tool releases, integrations, spend and process interviews.
- The purpose of discovery is registration and ownership, not enforcement. Framing it as an audit guarantees poor answers.
- ChatFin surfaces automated activity in the ledger itself, which is where undisclosed automation eventually shows up.
Ask a finance team whether they use AI and most will say a little, in a couple of places. Ask which of their tools have added AI features in the last year and the conversation changes.
This is the shape of the problem. Unregistered automation is rarely somebody deploying something they should not have. It is a feature appearing in a product you already use, a colleague finding it helpful, and a process coming to depend on it without any decision having been recorded.
Discovery therefore has to look for capability rather than for projects, because there was no project.
Why voluntary registration fails
Registers depend on people declaring things. That requires them to know the register exists, to believe it applies to them, and to recognise that what they are using counts.
The third condition is where it breaks. Somebody using a suggestion feature inside a spreadsheet tool or an assistant built into an ERP screen does not categorise that as deploying an AI system. They categorise it as using the software.
That is a reasonable view and it means the register will always be incomplete unless something actively goes looking.
The point is not that people are hiding things. It is that the framing of the question does not match how the capability arrived. Finance runs on a small number of large platforms, and every one of them has been shipping assistive features on a quarterly release cycle, so the capability arrives whether or not anyone asked for it.
The same effect appears in reverse when people over report. Asked whether they use AI, some colleagues will describe a search box or an autocomplete feature, which fills the register with entries that carry no risk and buries the ones that do. Both errors come from asking about technology rather than about what a tool is allowed to do.
Four angles that actually find things
Discovery works best from directions that do not depend on anyone volunteering information.
Each of these finds a different population:
Two ways to run it
The framing determines the quality of what you find, more than the method does.
One of these produces a complete picture and the other produces a defensive one.
Compliance audit
Presented as checking for unauthorised use. Produces careful answers, minimal disclosure and a register that looks clean while being wrong. Also makes the second round harder.
Support and registration
Presented as making sure useful things are supported, documented and not dependent on one person's account. Produces far better disclosure, and it is also true.

What to do with what you find
Discovery produces a list, and the response to each item determines whether people cooperate next time.
Three outcomes, and the middle one should be the most common.
| Finding | Response | Why |
|---|---|---|
| Useful and low risk | Register it, name an owner, leave it running | Removing something that works punishes disclosure and guarantees the next round finds nothing |
| Useful and higher risk | Register it and add the missing controls | Usually write access, evidence or approval. The capability stays, the gap closes |
| Dependent on one person's account | Move it to a proper account with an owner | The most common finding, and a continuity risk quite separate from any AI consideration |
Removing things found during discovery is the fastest way to ensure the next discovery round finds nothing.
The ledger tells on everything eventually
There is one place where undisclosed automation reliably becomes visible, and it is the transaction data.
Automated activity leaves patterns: timing clusters outside working hours, uniform descriptions, sequences that no person would produce at that speed, coding consistency that is too good.
Those patterns are worth looking for as a cross check on whatever discovery finds through interviews, because they do not depend on anyone telling you anything.
The check has a useful side effect. The same patterns that reveal undisclosed automation also reveal manual work that has quietly become mechanical, where a person is following a rule so consistently that the task is a candidate for proper automation with proper controls.
Nobody categorises using an assistant inside a screen they already use as deploying an AI system.
ChatFinRunning a discovery pass
Four steps, roughly a week of part time work, repeated quarterly.
Review release notes for your finance stack
Twelve months, major tools only. Produce a list of AI features shipped, then ask whether each is in use rather than asking a general question.
Pull the integration list
Everything connected to your ERP with the permissions each holds. Look particularly for connections whose original purpose nobody can explain.
Interview by process, not by technology
Walk through how three or four processes actually get done. The tooling emerges from the description without anyone having to categorise it.
Cross check against transaction patterns
Timing, uniformity and speed in the ledger. It catches what interviews miss and it requires nobody's cooperation.

Where ChatFin fits into discovery
ChatFin works in the transaction data continuously, which is the cross check that does not depend on disclosure.
Pattern detection across journals, invoices and payments surfaces activity that does not match human working patterns: entries clustered at improbable times, sequences produced faster than a person could produce them, descriptions with machine like uniformity.
That is ordinary anomaly detection rather than a governance feature, and it happens to be the most reliable route to finding automation nobody registered. It works across NetSuite, SAP Business One, JD Edwards, Acumatica, Dynamics 365 and Sage Intacct.
It also has a second use. The same detection distinguishes automation you know about from automation you do not, which is the distinction that actually matters for the register.
AI Cost Estimator for the CFO Office
Four inputs, a first year cost range, a timeline, and three delivery models compared. No login required.
Frequently Asked Questions
Is unregistered automation a compliance problem or a risk problem?
Usually a risk problem first. The common findings are continuity risks, such as a process depending on one person's account or an unsupported tool, rather than misuse. Framing discovery around support rather than enforcement produces a more complete picture.
How often should discovery run?
Quarterly is a reasonable cadence, because capability arrives through supplier release cycles rather than through procurement. Annual discovery leaves too long a gap for a process to become dependent on something undocumented.
What is the most common finding?
A useful automation running under an individual's account, undocumented, with no owner other than the person who built it. It is a continuity risk regardless of whether AI is involved.
How do transaction patterns reveal automation?
Through timing, uniformity and velocity. Entries clustered outside working hours, descriptions with machine like consistency and sequences produced faster than a person could work are all visible in the data without anyone reporting anything.
General information only. It is not legal, accounting or tax advice, and rules change. Check your own position with your advisers before acting on it.
Look for capability, not for projects.
Discovery framed as support finds considerably more than discovery framed as audit, and the transaction data provides a cross check that does not rely on anyone volunteering anything.
ChatFin surfaces non human activity patterns in the ledger continuously, which is where undisclosed automation eventually becomes visible.
Book a DemoSee the platform
See which activity in your ledger was not produced by a person.
Bring a period of journals and postings. We will show the timing, uniformity and velocity patterns that indicate automated activity and how they separate from normal work.