Yes, and it is one of the better uses for an agent in compliance. An AI agent can match customers through shared addresses, devices, directors, employers and payment flows that no single record shows. What a link means, whether fraud, a connected group, coercion or nothing at all, is a judgement a person makes with the evidence.
What kinds of hidden link can an agent find?
Customer records are built one customer at a time. Links between customers are not stored anywhere; they have to be found by comparing records. The useful signals fall into four groups:
- Shared identity details: the same address, phone number, email pattern, device or login location across apparently unrelated customers.
- Shared corporate roles: the same person as director, shareholder or signatory across several business customers.
- Money flows: funds moving in circles, fanning out to many accounts, or converging on one.
- Shared behaviour: accounts opened in the same week through the same channel, with similar profiles and similar first transactions.
No single signal proves much. An office building houses many companies, and families share addresses. The value comes from links that stack: three customers sharing a device, a director and a payment counterparty is a different picture from three customers in the same block of flats.
What evidence does the agent assemble?
The agent produces a link map with a note on every connection: what matches, which records it came from, and how reliable each record is. A device fingerprint from your own login system is strong. An address the customer typed in is weaker. A link inferred from a news article is weaker still. Each link carries that grading so the reviewer can weigh it.
It also records the searches that found nothing. If the agent checked for shared beneficiaries and found none, that matters to the reviewer as much as what it did find. Our view on this is set out in compliance agents must record what they did not find.
What does a person decide?
| Question | Why it needs a person |
|---|---|
| Is this link real or coincidental? | Context the records do not hold |
| Are these customers a connected group for exposure or risk purposes? | Changes limits and ratings |
| Does the network suggest fraud or laundering? | Leads to investigation and possibly a report |
| Does it suggest one person is controlling another's account? | May indicate abuse or coercion of a vulnerable customer |
| What action follows? | Restriction, exit and contact all affect customers |
The fourth row is easy to miss. A shared device between an elderly customer and a younger, unrelated one may be laundering. It may also be a carer, or financial abuse. Treating it only as a crime signal can harm the customer who most needs help. The FCA's expectations on this are covered in FCA Consumer Duty, vulnerable customers and AI agents.
What does a link investigation look like?
Take a digital lender. Fraud losses on small personal loans have risen. An investigator asks the agent to look for connections among the forty accounts that defaulted on their first payment.
The agent finds that eleven of the forty were opened from four devices. Seven list addresses on the same two streets. Five named the same employer, a company the registry shows was dissolved before any of the applications. The loan funds from nine accounts were sent within hours to the same three external accounts.
It also reports what it found against a ring: the other twenty-nine accounts show no shared devices or counterparties, and look like ordinary defaults. The investigator treats the eleven as a likely organised fraud and refers them on. She does not act against the twenty-nine. The lender adds the four devices to its application checks, a change a person approves.
What controls keep link analysis fair?
- Grading on every link. Reviewers see how strong each connection is, not just that it exists.
- No automatic action on a link. Finding a connection never restricts an account by itself.
- Purpose limits. The agent runs link analysis for defined purposes, such as fraud and AML. Using the same data to profile customers for marketing is a separate decision.
- Proportionate scope. It searches the records it needs, not every data source it can reach.
- Review of what the agent surfaced and what was decided. More findings mean more decisions, which is why surveillance agents can find more, so people must decide more.
This follows the approach in AI compliance agents: second line, never the last word.
Questions people ask
Can the agent link customers across different products?
Yes, if its role scope covers those products' records. Many useful links only appear across products, such as a savings customer funding several new current accounts.
Is link analysis a form of profiling under data protection law?
It can be, depending on what you do with the result. Your data protection assessment should cover the purpose, the data used and the safeguards.
How often should link analysis run?
Continuously for new applications, and on demand for investigations. A periodic sweep of the whole base also catches networks that formed slowly.