In early collections, the variable that decides most outcomes is time, not the price of a call. Every day an assigned account goes untouched, the debtor drifts towards the next bucket, and the bank drifts towards building reserves. We have run early-stage collections for banks and consumer-finance lenders in Poland, Austria and Germany, and the lesson repeats itself in every portfolio: the supplier that reaches the whole assignment quickly beats the supplier that handles a fraction of it beautifully. This piece explains why, and what speed actually means in practice.
What a bucket really measures#
A bucket is a band of days past due (DPD). Banks slice their delinquent book into these bands because the behaviour of a customer who is a few days late is very different from the behaviour of one who has missed several instalments. Early in delinquency, many customers have simply forgotten, changed accounts, or been travelling. Later, the reasons tend to be harder: income loss, competing debts, disputes.
The bank’s own analysts decide which accounts go to which channel. Their allocation strategy works across three dimensions: the portfolio (credit card, cash loan, instalment loan, car finance, overdraft), the bucket, and the debtor category derived from history and risk. A contract that is not paid after contact can come back in the next allocation as a new case. One loan, several cases in a month, each one a fresh chance to prevent the slide.
That is why we keep three units apart. An account is the customer relationship, such as a loan contract. A case is each assignment of that account to us for handling. An attempt is a single dial. Confusing them produces arguments about performance that nobody can win.
Why the delay hurts more than the call#
Procurement teams tend to compare collection suppliers on the cost of handling. That is a reasonable number to watch, but it is not the one that keeps a head of collections awake. For a bank, the most painful part of delinquency is not the cost of the call. It is the delay that forces it to build reserves.
The mechanism is simple and unforgiving. A neglected debtor accumulates delay. Once the delay crosses the thresholds set by the bank’s risk policy, the bank must set capital aside against a possible full loss on that product. Frozen capital is capital that cannot fund new loans. If the account deteriorates further, the bank faces selling the debt, harder recovery methods or legal costs, and the margin on the product shrinks or turns into a loss.
Put differently: a contact that happens on day three and a contact that happens two weeks later may cost exactly the same, but they are not worth the same. The earlier one reaches a customer who is still in the habit of paying. The later one reaches a customer who has already moved one band down, with a lower chance of curing and a higher cost to the balance sheet.
Touching the whole assignment#
Here is the structural problem with handling early collections through people alone. A human team has a fixed number of seats and hours. When the allocation arrives, it must prioritise. Sensible teams start with the highest balances or the riskiest categories, and the tail of the assignment waits. Some of it waits long enough to roll into the next bucket before anyone dials.
Our position, earned over years of outbound campaigns at very high volumes, is that the first job in early collections is to touch every assigned case. Not to call the best-looking half well, but to make sure each account the bank has allocated is attempted, reached where possible, treated correctly and reported back. Prioritisation still matters, but it should decide the order within a day, not which accounts get no contact this week.
This is where executing conversations with AI changes the arithmetic. Capacity is a question of infrastructure, not headcount. We saw this first in March 2020, when demand from large clients rose roughly fivefold within weeks and channel capacity at one mobile-operator client went from 250 to 900. Nobody had to be hired, trained or scheduled. The same logic applies every month in collections: when an allocation lands, the whole of it can be in progress the same day, within the call windows the market allows.
Touching the whole assignment also changes what happens when a contract comes back. Under most allocation strategies, an account that does not cure is reassigned as a new case, sometimes more than once in a month. If the first case was never reached, the second one starts from nothing, a bucket later. If the first case was reached, the second one starts with knowledge: which number answered, at what time, what the customer said, whether a promise was made and broken. Speed on the first case makes every later case cheaper to work and more likely to succeed.
Without reaching the customer, there is nothing else. – Dawid Wójcicki, Founder & CEO of Vee
| Question | Seat-limited team | AI-executed, managed service |
|---|---|---|
| How much of the allocation is attempted on day one? | What the available seats and hours allow; the rest queues | Potentially all of it, within the market’s permitted windows |
| What decides who waits? | Capacity, then priority rules | Priority rules only; capacity is planned in channels |
| How fast does a change in wording reach every call? | After briefing and coaching | One approved change applies to every live conversation |
| What does the bank get back each day? | Often aggregated activity reports | Case-level statuses the bank’s systems can act on |
| Who keeps the allocation strategy? | The bank | The bank |
Qualitative comparison of operating models; not a measured result.
Speed is not the same as dialling harder#
Speed in collections is easy to misread as aggression: more attempts, earlier in the morning, later in the evening. That approach damages reach over time and invites complaints. Speed that works is a matter of tuning a set of contact parameters for each base, which together form what we call a reachability strategy.
- Outbound number. Which number the call comes from, managed so that the line stays recognisable.
- Destination number. Which of the customer’s numbers is used.
- Time window. When the call is placed, inside call windows that the system enforces per market rather than leaving to good intentions.
- Channel and content. How text messages support the call.
- Intensity and intervals. How many attempts, and how far apart, tuned per base.
- Persona, voice and style. Who the customer hears, at what pace, with what tone.
None of these touches the bank’s allocation strategy. The bank decides which accounts are called and when they are assigned; we decide how to reach them. Keeping that line clean matters in regulated lending, where the bank must be able to explain why a given customer was treated a given way.
Retry timing deserves particular attention because it is where speed is most often wasted. The next attempt should depend on why the last one failed and on what the customer asked for. Treating every non-answer the same is how a fast first attempt turns into a slow second one.
One allocation, from file to status#
It helps to follow a single allocation through a working cycle, because the places where days are lost become obvious. The bank’s analysts select the accounts and send the file, usually as a flat file over SFTP or through an API. That file carries what we need to reach and treat each customer correctly: the product, the bucket, the amount, the contact numbers and whatever the bank shares about history and preferences.
The first loss of time often happens here, before a single call. If the file arrives late in the day, or fails validation, or needs a manual step on either side, the allocation sits. We treat data exchange as part of the service rather than as the client’s homework, which is why integration is specified and tested before launch alongside the dialogue and the telecoms set-up.
Once the file is loaded, the cases enter the contact plan. Where the bank has agreed it, a text can go before the first call so that the customer recognises what the call is about. The first attempts run inside the permitted window. Every attempt produces a technical status: answered, no answer, voicemail, busy, operator error. Every conversation produces a dialogue status: verified or not, already paid, promise made, dispute, request for a person. The retry plan reads those statuses and schedules the next attempt accordingly.
By the end of the day, the bank receives the business statuses for each case, in its own language and categories. A customer who has paid or made a valid commitment can be closed or excluded from tomorrow’s attempts. A customer who disputes the debt goes to the right team. A customer who could not be reached stays in the plan with a different time window or number. The next morning the cycle continues, and when the bank’s strategy sends the same contract back as a new case, the history from the previous one is already there.
None of this is glamorous. It is the operational layer that voice demos never show, and it is where most of the difference between a fast service and a slow one sits. A conversation that sounds excellent but starts four days after the allocation has already lost much of its value.
What a fast first contact has to get right#
Reaching the debtor early is only valuable if the conversation that follows is correct. In banking, that means the order of the conversation is not negotiable. Our early-collections flow, configured per market and product, follows the same backbone.
- Disclosure and introduction. The customer hears that they are speaking with an AI calling on behalf of their bank. We have never hidden this, and since 2 August 2026 Article 50 of the AI Act makes it an explicit transparency obligation.
- Identity verification. No detail of the debt is disclosed before verification succeeds. This is enforced by the platform, not written as an instruction in a script that someone could skip.
- Payment status. Has the customer already paid? Many have, or are about to. Recording that correctly saves the bank a pointless follow-up.
- Commitment. A date and amount within the horizon the bank allows for a promise to pay, with the payment method explained.
- Business status. A result the bank’s systems can act on the same day: close, exclude from tomorrow’s calls, or trigger a specific next action.
Two refinements make the backbone fit a real portfolio. First, DPD shapes the base conversation: urgency, the language of consequences and the negotiation logic all change with the bucket. Second, behaviour history adjusts it further, for example when a customer has broken a previous promise. Separate dialogue paths for different customer groups run in parallel within one deployment.
Then there are the exits, which matter as much as the main path. A customer who asks for a person, fails verification, withdraws consent to recording, reports a death or a legal dispute, requests erasure, or describes financial hardship is routed according to rules agreed with the bank: transfer, hard stop or referral. Speed never justifies pushing through those moments.
Measuring speed honestly#
If speed is the argument, it has to show up in the metric the bank cares about, which is whether customers paid. Not minutes, not attempts, and not even promises. We report against a comparator on the same buckets and the same portfolios, using the bank’s own systems as the source of truth, at the level of contracts rather than calls.
We do not claim that every payment after our contact is ours. Some customers would have paid anyway, and the honest measure is the incremental effect over a comparator, not the gross recovery rate. That is also why we recommend starting on part of the volume: the bank can run Vee and its existing channel side by side on comparable buckets, see the payment result in its own data, and decide on scale with evidence rather than a demo.
There is one more effect that minute-based reporting never sees. Contact has value in itself. Some debtors pay because they saw that their bank called, without a conversation ever taking place. Others say they know and have no time, end the call without a promise, and pay the same afternoon. Measuring at the level of accounts in the bank’s systems is the only way to capture that, and it is one of the strongest arguments for touching the whole assignment early.
Daily reporting closes the loop. The bank does not need another dashboard to interpret; it needs case-level statuses it can act on before the next allocation. Which accounts are closed, which are excluded from tomorrow’s attempts, which need a different action. In our experience, the less the client has to aggregate and interpret, the faster the whole cycle runs.
Where speed has limits#
Speed is not a universal cure, and we say so in the first conversation. A base with poor phone data will not become reachable because calls start earlier. A portfolio where most customers are in genuine hardship needs a careful route to the bank’s specialist team, not a faster script. Calling windows and attempt limits differ by market, and in Austria or Germany the rules on authentication and the tone of consequence language are not the same as in Poland.
There is also a ceiling on what any contact supplier controls. We influence payment indirectly. Reach, continuation of the conversation and the quality of the commitment are ours to improve; the customer’s financial position is not. That is why we put our name to the incremental effect against a comparator and not to a promise about the absolute recovery rate.
Finally, speed does not make the bank’s own people redundant in collections. It changes what they spend their day on. When the repetitive first contact across the whole assignment is executed by AI, the bank’s specialists receive the cases that genuinely need judgement: hardship, disputes, customers who asked for a person. That division of work is designed on day one, with agreed transfer rules, rather than discovered after launch.
What to ask a collections supplier#
- What share of a typical allocation do you attempt on the first day, and what share do you reach within the first few days?
- How do your retry intervals change after no answer, voicemail, busy, a dropped call and a callback request?
- Who enforces call windows and attempt limits per market: the system or the operator?
- Is identity verification enforced by the platform before any debt detail is spoken?
- Will you report payment results against a comparator on the same buckets and portfolios, in our systems?
- Which statuses do we receive daily, and can our systems act on them without manual interpretation?
- Does our allocation strategy stay with our analysts?
Those questions move the discussion from the price of a call to the cost of a day. That is the comparison that matters in early collections. It is how we have run the service since we began handling collections at scale, and it is the conversation we are happy to start with any bank that wants to test it on part of its own volume.
Follow The First Second
Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.
Follow via RSS