In every collections process we run, debt-sensitive content is never disclosed before the person on the line has been successfully verified. That rule is hard-coded at the platform level; it is not an instruction in a script that a designer could move, a branch could skip or a customer could talk their way around.
The distinction sounds technical. It is actually the difference between a guardrail you hope holds and one that cannot fail in the way that matters. This article explains why the rule exists, how it is enforced, what happens when verification does not succeed, and what the same principle means for other guardrails in regulated customer contact.
Why this is the first rule of a collections call#
An outbound collections call starts with uncertainty. The number in the bank's file may now belong to someone else. A spouse, a parent or a colleague may answer. A teenager may pick up a family phone. The customer may have changed numbers years ago and the line has been reassigned. Until the person has proved who they are, the system has to assume it might be talking to anyone.
Telling the wrong person that someone owes money is not a minor slip. It is a disclosure of personal data and, for a bank, of information protected by banking secrecy. It can damage a family relationship, embarrass a customer at work and turn a manageable delay into a complaint, a regulatory issue and a reputational problem. In early delinquency, where many customers are simply late and will pay once reminded, it also destroys the goodwill that makes the next conversation possible.
So the opening of the call is designed around one question: is this the right person? Our AI introduces itself as AI and says which organisation it is calling from, then moves to verification. Only after verification succeeds does the conversation reach the amount, the due date, the consequences or the payment options.
A script instruction versus a platform rule#
There are two ways to implement "verify first". The first is to write it into the dialogue: the script asks verification questions, then moves to the debt. The second is to make it a property of the system: debt-sensitive content and variables are locked until the conversation has reached a verified state, whatever path it took to get there.
The first approach works as long as nobody makes a mistake. But scripts change. A new branch is added for customers who say they have already paid. A callback path is introduced. An exception is created for a specific product. Each of those changes is a chance for a path to reach debt content without passing through verification. In a human contact centre, the equivalent risk is an agent under time pressure who skips a step because the customer sounds right.
The second approach removes that class of error. However the dialogue is restructured, the platform will not play a debt-sensitive utterance or insert a debt amount into a sentence unless verification has succeeded in that call. A designer cannot accidentally bypass it, because it does not live in the design. A customer cannot argue past it, because there is no sentence in the catalogue that answers "just tell me how much it is" before the gate opens.
What verification looks like in practice#
Verification is not one fixed question. Requirements differ by market, by bank and by product, largely because banking law and each institution's own policies differ between countries. We run early collections in Poland, Austria and Germany, and authentication is one of the areas where the processes genuinely diverge.
- Number of factors. One market or bank may require two pieces of information; another may require three.
- Type of data. Name, date of birth, postal code, an instalment amount the customer should know, or company data for business accounts.
- B2C and B2B paths. A business account needs a different route: the person answering may be authorised to speak for the company without being the named borrower.
- Product-specific paths. A credit card, a cash loan and a car loan can each carry their own verification rules.
- Imperfect answers. People give a date of birth in a different format, mishear a question or answer only partly. The flow handles partial or incorrectly formatted responses rather than failing the customer on a technicality.
All of this is configured per deployment. What is not configurable is the gate itself: whatever the verification route, debt content stays locked until it is passed.
In our collections processes, verification is also where conversations are most often lost. Ask too much, too abruptly, and a customer who was willing to talk hangs up; ask too little and the gate means nothing. So the way verification is phrased, the order of the questions and the voice that asks them are among the things we test, with the client's agreement and within the factors the bank requires. The rule stays fixed. How politely and clearly we get people through it is something we keep improving, because every customer lost at verification is a case that cannot reach a promise to pay.
When verification does not succeed#
A guardrail is only as good as what happens when it holds. If the call cannot reach a verified state, the process still has to end correctly, and the bank needs to know what happened so it can act.
| Situation | What the system does | What the bank receives |
|---|---|---|
| Someone else answers | Closes the call without disclosing anything about the debt and schedules another attempt | A status showing a third-party contact and the next attempt |
| Wrong number | Ends the call politely; the number is not called again for this account | A flag for data review |
| Verification fails | Escalates or transfers, depending on the market configuration | A status with the failure type |
| Customer refuses to verify | Does not proceed to debt content; ends or transfers as configured | A status the client can act on |
Exact routes are configured per market and per client.
The point of the table is that nothing in it requires judgement in the moment. Each outcome maps to a defined system action and a status in the daily results file. The bank's own systems do not need to interpret the call; they receive an outcome and act on it, for example by excluding a wrong number from tomorrow's calling list.
The same principle, applied to other guardrails#
Identity before disclosure is the clearest example, but it is one of several rules we keep out of scripts and in the platform. In regulated contact the question to ask of every rule is: what happens if someone changes the dialogue tomorrow?
- Calling windows. Market-specific hours are enforced by the system. A campaign cannot call outside them, whatever the schedule says.
- Maximum attempts and retry intervals. Limits are configured per market and enforced, so a busy period cannot quietly turn into over-contacting.
- Recording consent. If a customer withdraws consent to recording, that triggers a defined outcome, not an improvised response.
- Exception paths. A request for a human, a death or legal dispute, a data-erasure request or a declaration of financial hardship each lead to a configured transfer, structured stop or referral.
- Content. In regulated configurations, every sentence comes from a closed, approved and versioned catalogue, so there is no generated text that could reveal what the gate is protecting.
Together these make the difference between a process that is compliant when everything goes to plan and one that stays within the rules when things do not. And because every call is reconstructable, with a recording where agreed, connection logs, statuses and the script version, the bank can show that the guardrails held, not merely assert it.
Platform rules also make the process easier to change safely. Because the gate cannot be bypassed by a dialogue edit, we can test changes to the dialogue without re-proving each time that verification still comes first. The guardrail is not part of what is being tested.
Why this matters more with AI, not less#
Some assume automation makes this kind of rule easier: a machine follows the script. That is only true if the machine cannot be steered. A system that composes its replies freely can, in principle, be talked into revealing what it knows, or can volunteer it while trying to be helpful. That is one reason we do not let a language model write customer-facing sentences in collections. Models may help understand what the customer said; the sentences themselves and the order in which sensitive content becomes available are fixed.
There is a scale argument too. Collections runs at high volume. A rule that fails once in ten thousand calls fails many times a month. A rule enforced by the platform does not have a failure rate of that kind; it either exists in the release or it does not, and that can be tested and evidenced before launch.
What to ask a supplier#
- Where exactly is identity-before-disclosure enforced: in the dialogue design, or in the platform beneath it?
- If your team adds a new branch to the conversation tomorrow, what stops it reaching debt content without verification?
- Can a customer, by asking directly or persistently, get the system to state an amount before verification?
- How are verification factors configured for each market, product and B2B account?
- What happens, and what status do we receive, when someone else answers or verification fails?
- Can you show us, for a sample of calls, that no debt content was played before the verified state?
Good answers to these questions are short and concrete. If the answer to the first is "the script handles it", you are relying on every future script change being perfect. In the processes we run for banks, we would rather the rule held regardless.
Follow The First Second
Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.
Follow via RSS