We gave an AI agent two words of instruction.
Not a rulebook. Not a decision tree. Not a list of fraud indicators built up over years of compliance experience. Not even a schema of the database it was about to explore.
Just: "Find fraud."
What happened next is one of the most genuinely surprising things I've witnessed in my time working at the intersection of AI and enterprise technology. And it says something important - not just about what AI can do, but about the gap between what most organizations think they're buying and what's actually possible.
In traditional fraud detection, the logic is explicit. Humans define the rules: flag any transaction where the supplier bank account changed within 30 days of a payment. Flag any invoice where the amount exceeds the approved PO by more than 10%. Flag any supplier registered at the same address as an employee. I spent 10+ years partnering with some great technologies which used AI to detect fraud, but they were still rule-based with AI peripherally, not truly stochastic.
So yes, these rules work. They catch known patterns. But they have a fundamental limitation: they can only catch what someone already thought to look for.
The agent we deployed had none of those rules. It had access to a knowledge graph - a connected map of every entity in the procurement system: suppliers, invoices, purchase orders, approvers, requestors, contracts, bank accounts, addresses. And it had a goal.
Find fraud.
The agent began by exploring the database structure it had never seen before. It wrote its own queries. It ran them. It examined the results. It generated hypotheses. It tested them.
It found patterns.
Not the patterns we told it to look for. Patterns it discovered - through graph traversal, through relationship analysis, through the kind of connected reasoning that a relational database query or a keyword search simply cannot perform.
Relationships between suppliers and approvers that shouldn't exist. Unusual clustering of invoice amounts just below approval thresholds. Supplier entities with structural similarities to known fraudulent actors, connected through chains of relationships that would be invisible in a flat data view.
A human analyst, given the same database, might find some of these. Given months. The agent found them in minutes.
I want to be precise here, because the word "AI" has been so overloaded that it has almost lost meaning.
What made this possible wasn't the language model. Language models are increasingly commoditized - the same foundation models are available to almost every technology vendor.
What made this possible was the architecture underneath.
ORO's knowledge graph - built on Neo4j - models procurement data not as rows and columns but as a network of connected entities and relationships. Every supplier is connected to every invoice they've ever generated, every PO they're linked to, every approver who touched their transactions, every contract that governs their engagement. The graph holds context that a flat database cannot.
When you give a goal-based reasoning agent access to that graph - combined with proprietary algorithms sourced from academic research on community detection and anomaly identification - you get something qualitatively different from a chatbot answering questions.
You get a system that can discover what it doesn't know it's looking for.
Here's the thing that's stayed with me since we ran this experiment.
Most organizations are deploying AI to answer questions they already know to ask. Automate this process. Summarise this document. Classify this invoice. Where other Orchestration companies are struggling to even answer these questions, ORO’s 5 purpose-built data stores make easy work of them. The answers are valuable. But now we can also make them transformative.
You see, the larger transformative use cases are the ones where AI discovers something you didn't know to look for. Where the value comes not from faster execution of a known process, but from intelligence applied to an unknown problem.
That requires a different architecture. It requires data structured for relationship reasoning, not just retrieval. It requires agents designed for exploration, not just execution. And it requires the domain expertise to build the right pipeline - knowing which algorithms apply, how to structure the AI reasoning layer, how to validate the outputs in a way that's defensible in a regulated environment.
This is the gap between AI as a feature and AI as a capability.
If you are evaluating AI in procurement - or in any enterprise domain - here is the question I'd encourage you to ask (or you could even evaluate these kinds of questions in an RFP - we love it when we get asked such questions!):
Is this AI designed to answer the questions I already know to ask?
Or is it capable of discovering the questions I haven't thought of yet?
Most enterprise AI today is the former. It is fast, it is useful, and it has genuine value.
But the ceiling of "answer my known questions faster" is lower than most leaders realize. The real opportunity - the one that justifies the investment and the organizational change - is AI that expands the frontier of what's knowable.
"Find fraud" is two words.
But what they represent is a fundamentally different model of what AI can be in an enterprise - not a smarter search engine, but an investigator. Not a tool that executes instructions, but an intelligence that pursues outcomes.
Looking to the future
What’s even more interesting is what this new reality we have built here at ORO means for the future. Soon systems will be able to look at themselves in the mirror and propose answers to questions about themselves we aren’t even asking. Self-improving systems? Who knows, maybe…
The architecture to make all of this possible exists, though. Of that, I am now sure. I am also sure that no agent, no matter how “super,” can innovate a process better than you ever imagined unless the orchestration data model and underlying graph supports it.
If you want to learn more about our architecture and purpose-built 5 layer data strategy, or you want to understand the implications of how orchestration architecture is the key to unlocking your AI strategy, get in touch for a demo!
By Chris Vessey, VP of Innovation & Customer Value
Chris has spent his career helping large organizations untangle complex procurement systems, unify teams, and turn transformation into something people actually feel, not just a program plan. For 20+ years, he has led global Source-to-Pay, Third-Party Risk, and Spend Management transformations across FTSE 100 financial services and FMCG companies, managing $10Bn+ in spend and 100-strong global teams.