In the past eighteen months most fraud and AML platforms have shipped an AI feature, and the shapes are similar. The AI reads your past alerts and suggests a rule, or turns a sentence into a rule, inside the vendor's rule builder. A person approves it. The rule lives on in the vendor's database, in the vendor's format, on the vendor's infrastructure.
It looks like a category-creation moment. It is the autocomplete moment.
The Cursor parallel
For about a decade every IDE had autocomplete, and as language models improved it got good: tab through a signature, accept a plausible body. It was useful and measurable, and it changed nothing about what engineering was. You still wrote the code. The tool helped you type it.
Cursor changed the substrate. The model stopped suggesting and started writing. You describe the change, the agent writes it, you review it, keep what works and fix what does not. The activity moved from an engineer typing with help to an agent writing with an engineer reviewing, and engineers who made the switch describe going back to autocomplete as a loss.
Fraud detection is in the first half of that story. The AI rule features shipped so far are autocomplete on the rule builder. The substrate underneath, rule state in a vendor's UI, has stayed where it was.
The second half is an agent that reads your transaction, ledger, session and KYC data and writes the detection as code, in your repository, version-controlled and reviewed like any other code. The analyst's job shifts from authoring rules with help to reviewing the code the agent authored. That is what Detection-as-Code for fraud means.
Where the term comes from
Detection-as-Code came out of security operations. Around 2018, security engineering teams began treating their SIEM detections the way infrastructure teams had started treating servers: as code, in git, peer-reviewed, tested and deployed through CI. Panther productised the pattern. A detection is a Python file with metadata, in the customer's repository, deployed by pull request, with no UI between the detection and the customer.
Security adopted it because the alternative broke at scale. Past a hundred rules a team wants branches, reviews, rollbacks and tests. And once a detection is code it can be unit-tested, replayed over history, composed from statistical libraries and graph algorithms, and audited by reading the thing that fired.
Fraud has been slow to follow, for three reasons that are all eroding. Fraud teams were staffed by investigators and compliance officers who do not write code, so vendors built no-code authoring for them, and engineering-led fintechs now have a different kind of team. Model-risk documentation was easier to satisfy inside a vendor's UI, and it is now cheaper to generate from code, tests and replay results. And the real engineering happened one layer down, in feature pipelines, which are mature enough that detection authoring can move into the same workflow.
What it unlocks
Detection without labelled history. Rule suggestions learned from past dispositions need past dispositions. An agent that writes code from raw data can encode the universal patterns directly: velocity on heavy-tailed distributions, deviation from a peer baseline, correlation across sources, rings in a transfer graph. A company with no case history can have detection on day one, written from its own data rather than from someone else's templates. Our reference run did this. On one remittance operator's exports, with no labelled cases, the agent flagged five accounts and the operator confirmed four.
The audit artifact is the code. When an examiner asks why an account was flagged, a platform shows a narrative generated after the fact from the score that fired. With Detection-as-Code the artifact is the rule itself. A reviewer can read it, run it again, and replay it over any past month, and a backtest produces the same evidence as a live deployment.
Composition across sources. A rule builder is bounded by the features the vendor pre-built. Code is bounded by the language. A velocity three median absolute deviations above the peer baseline, on a device shared with an account flagged in the last thirty days, with a document uploaded from a different country than the transaction, is three sources, two statistics and a temporal join. Python expresses it in an afternoon.
A higher ceiling on autonomy. Suggest-only is a risk ceiling. A vendor cannot audit its own suggestions deeply enough to let them deploy on their own. When the output is code that is tested against fixtures, replayed over history and scored, a graduation system becomes possible, where rules below a threshold wait for review and rules above it move forward with a full audit trail. ATROSA stops short of that today, and every rule is reviewed by the customer's team before it runs on new data. The point is that the ceiling is set by the substrate, and this substrate's ceiling is higher.
Logic you own. Detection logic that lives in a vendor's database stays there when you leave. Code in your repository moves with you, and a vendor has to earn its place by writing better rules than you would on your own, which is where a software vendor should be.
Why now
Three things are converging. AI coding agents are arriving at fintechs, and the artifact they emit is code, so the substrate for detection follows them. The mid-market is underserved: a seed-stage company with two data engineers and no fraud team is priced out of the incumbents and either builds in-house or does without. And the language is open, because the adjacent vendors who used an as-code framing have moved on to calling themselves agentic.
The question worth asking
Engineers explain Cursor with a thought experiment. You have used it for a year, and tomorrow you have to go back to autocomplete. How does that feel? The answer is a loss, because the activity itself changed.
Run it for fraud. Your team has had an agent writing detections into your repository, replayable and auditable, from raw data. Tomorrow you go back to clicking through a rule builder. I think the answer will be the same once the second tool exists.
Most teams will not switch tomorrow. Procurement, compliance and habit all push the other way. But the substrate is moving, the way it moved in software when Cursor arrived, and the category that follows is a different one. The incumbents will keep being good at no-code rule building with AI assistance. It is no longer the only category.
The open-source detections are on GitHub under the MIT licence. The hosted version runs your first export at no charge.
