Give the agent a written list of difficulty signals, agreed with your support and data protection teams, and let it watch for them in account data and in conversations. When signals appear, it adds a support flag, which pauses collections pressure, and drafts a note for a support specialist. The specialist decides whether and how to contact the customer.
What is the process, and who owns it?
Customers rarely announce financial difficulty. It shows first in the account: income falling or arriving late, essential bills bouncing, the overdraft used to the limit every month, cash withdrawn on a credit card, new short-term lenders appearing among the credits. Later it shows in conversation, often obliquely: "things are a bit tight", "I've had to cut back", "can I pay next week?".
Spotting it early lets a bank offer help before arrears, fees and collections make things worse. The argument for doing it before any request for money is made in collections agents must spot difficulty before they ask for money. The head of financial support owns the process and the signal list. Data protection owns the assessment of the profiling involved. Support specialists own every contact. For the wider frame, see our guide to AI agents in banking.
What are the steps, and who acts at each one?
- Person, head of financial support: writes the signal list. Each signal defined precisely, with the data it uses. For example: "two or more essential payments returned unpaid within a month". Evidence: versioned signal list.
- Person, data protection lead: assesses the processing. Inferring difficulty from transaction data is profiling. Under GDPR, the Article 5 principles and an Article 6 lawful basis apply, and a data protection impact assessment under Article 35 is likely to be needed. Signals must not infer health or other special-category information. Evidence: the assessment.
- Agent, depth 1: monitors account data for signals. On the accounts and data the assessment allows. Evidence: signals found per customer, with the data behind them.
- Agent, depth 1: listens for signals in conversations it handles, in any service context. Evidence: the customer's words, recorded.
- Agent, depth 3: adds a support flag. The flag pauses collections contact and fee escalation and routes future contacts towards support. Adding is protective; only a specialist removes it. Evidence: flag with the triggering signals.
- Agent, depth 2: drafts a case note for the specialist: signals, account position, any recent contact. Evidence: case note.
- Person, support specialist: decides on outreach. Whether to contact, by which channel, with what offer. Evidence: decision.
- Agent, depth 2: drafts the outreach in approved wording for the specialist to send, including free independent debt advice where available. Evidence: message and sender.
- Person, support specialist: agrees any treatment. Plans, holidays, fee reviews, referrals. Evidence: outcome recorded.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Signal list | Nothing | none | Head of financial support |
| Data assessment | Nothing | none | Data protection lead |
| Monitor data | Finds signals in permitted data | 1 | Nothing |
| Listen | Records signals in conversation | 1 | Nothing |
| Flag | Adds a support flag | 3 | Removing the flag |
| Case note | Drafts the summary | 2 | Nothing |
| Outreach | Drafts the message | 2 | Specialist decides and sends |
| Treatment | Nothing | none | Specialist agrees options |
Which controls matter, and what evidence do they leave?
Signals written down. The agent finds what the list defines, not what it guesses. Evidence: signal list version on each flag.
One-way flag. The agent adds flags freely and never removes them. Evidence: refused verdicts on removal attempts.
No sensitive inference. Signals may not stand in for health, disability or other special-category data. A customer who tells you about an illness has shared it; the agent records it as said and routes the case. Evidence: signal definitions reviewed by data protection.
Contact by a person's decision. Unsolicited contact about money worries is sensitive. A specialist chooses whether and how. Evidence: decision before any outreach.
No advice from the agent. Pointing to free debt advice is support; recommending a course of action is advice. The line is drawn in whether an AI agent can give financial advice. Evidence: approved outreach wording.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What should you measure?
- Customers flagged, by signal.
- Flags confirmed by specialists, against flags the specialist judged unnecessary.
- Time from first signal to specialist review.
- Customers who later reached collections without a flag. Each is a missed signal worth studying.
- Outcomes after support: arrears avoided, plans kept, complaints.
- Flags removed within a month, a sign the signal list may be too sensitive.
Where does it usually go wrong?
Signals are vague. "Signs of stress" lets the agent flag almost anyone. Define each signal in data.
Flags do nothing. The customer is flagged but collections continues. Wire the flag into every contact system.
Outreach feels intrusive. A message saying "we noticed you're struggling" can alarm or offend. Specialists choose words and channels with care.
The agent tells the customer what to do. Well-meant suggestions about cutting spending become advice.
Signals fire on one-off events. A single late salary or a large one-off purchase is not difficulty. Require signals to persist or combine before flagging.
Signals spread beyond the purpose. Difficulty flags start informing marketing or credit decisions. Keep the purpose narrow, as the assessment requires. The capability itself is examined in whether an AI agent can spot a customer in financial difficulty.
Questions people ask
Should customers be told they were identified by an agent?
Be open about it if asked, and describe it in your privacy information. The outreach itself should focus on the help available, not the detection.
Can the agent use data from other banks?
Only what the customer has shared with consent, such as through open banking, and only if the data protection assessment covers that use.
What if a flagged customer says they are fine?
The specialist records it and decides whether to remove the flag. Some customers are fine; others are not ready to talk. Leave the door open.