In a regulated process, nothing the customer hears is improvised. Every sentence our AI says in a bank's collections call comes from a closed catalogue of approved, versioned utterances, and the language models in the stack are used to understand the customer, not to compose the reply.
That is a deliberate architectural choice, not a lag in adopting new technology. We use large language models where they earn their place. But in a call where a wrong sentence can disclose a debt to the wrong person, misstate a consequence or promise something the bank never approved, we want the risk to sit in something a compliance officer can read in full before the first call is made. This article explains what a closed catalogue is, how the split between recognising and speaking works inside a live call, what it costs, and where we think generation genuinely helps.
What a closed catalogue actually is#
A closed catalogue is the complete, finite set of things the system can say in a given process. Each utterance has an identifier, a wording, a voice recording or a synthesis definition, an approval owner, an effective date and a version history. The dialogue moves between states according to approved scenarios, transition rules and logical conditions. At each state, the system picks the next utterance from the catalogue. It does not write one.
In practice a catalogue is not a short list. A collections process has separate dialogue paths for different customer groups running in parallel within one deployment, and each path has its own openings, verification prompts, negotiation steps, consequence language, confirmations and closing lines. A customer's history with the bank shapes which conversation they hear. All of those variants are written, reviewed and approved in advance.
Variables are part of the catalogue too. An amount, a due date, a payment method or a reference number is slotted into an approved sentence from the client's data, never paraphrased. The wording around the variable is fixed; only the value changes, and only from a defined source field.
Recognise, don't write: the split inside a call#
A phone conversation has two directions. The customer speaks, and the system has to work out what they meant. Then the system speaks, and the customer hears something. These are very different problems with very different risk profiles, and we treat them separately.
Understanding the customer is where statistical models are genuinely useful. People do not answer the way a script expects. They say "I paid it on Friday", "my wife deals with that", "can you call after five" or "what is this about?" with background noise, regional accents and half-finished sentences. Classifying that into an intent the process recognises is a recognition task, and recognition can be made robust with models. If a classifier is uncertain, the dialogue asks a clarifying question from the catalogue. A recognition error leads to a polite re-ask, not to an invented promise.
Speaking back is where the risk sits. What the system says can create obligations, disclose personal data, misstate a legal consequence or mislead a vulnerable customer. That is why, in regulated configurations, the output side is closed. A model may decide which branch the conversation should take within the approved rules. It does not decide the words.
This is not a new idea for us. Our first working prototype in August 2013 simulated a motor-insurance premium process and understood synonyms and colloquial speech, recognising that a driver who says he puts in oil means diesel. It did simple arithmetic on annual mileage and corrected inconsistencies between make and model. Even then, the intelligence went into understanding. What the system said back was designed by people.
The variants we run#
Our dialogue technology, Brilliance StateGuard, comes in variants that differ precisely in how much statistical or generative machinery sits in the loop. The choice is made per process, according to how sensitive the conversation is.
| Variant | What understands the customer | What produces the reply | Typical use |
|---|---|---|---|
| H | Deterministic rules; no LLM, no neural networks | Closed, approved library | Clients with the highest compliance requirements |
| 1 and 2 | A small proprietary model used only to classify intent | Closed, approved library; dialogue outcome is deterministic | Regulated processes that need more flexible understanding |
| 3 | An LLM at the classification stage, which needs less training data and allows faster launch | The state machine; the LLM does not generate replies | Faster deployments that keep control of content |
| 4 | Generative mechanisms | Generated within limits | Low-sensitivity processes for smaller businesses; technically still in build |
Vee's architecture description; the variant is chosen per process, not per client.
The line that summarises it internally is short: deterministic where it matters, generative where it helps. Note what changes between variants 1 and 3: the recognition engine becomes more capable, but the reply side does not open up. That is the point. We can adopt better understanding without accepting new content risk.
Why improvisation is the wrong risk to take in banking#
The case against free generation in a bank's customer calls is not that language models are bad. They are remarkable. The case is that a regulated institution carries obligations that free generation makes hard to meet, and the cost of meeting them falls on the bank, not on whoever supplied the model. When we analyse a regulated deployment, six design challenges come up every time.
- Autonomous language generation. The institution loses advance control over every possible message the customer might hear, and takes on the burden of content governance, transparency and evidence for each one.
- Uncontrolled content quality. Generated text carries a structural risk around truthfulness, completeness and consistency. Even a small error rate becomes a large number of wrong sentences at the volumes collections runs at.
- Decisions with material impact. If a model infers a customer's economic situation and routes them into a different consequence path, the process moves closer to the kinds of uses the AI Act treats as high-risk.
- Oversight that is only nominal. Human oversight has to be a real ability to interpret, intervene, override and stop, not a named person on an organisation chart.
- Drift after deployment. A system that changes its behaviour on its own after go-live changes its risk profile with it, and creates a continuing reassessment burden.
- No reconstruction. An auditor needs to know what input led to what output under which configuration, and what any human did. A generic call log does not answer that.
A closed catalogue answers most of these at once. Content is approved before use. Quality is checked once per utterance, not once per call. Consequential decisions stay with the bank's own rules and staff. Oversight has explicit transfer, stop and approval points. Nothing changes in production without a versioned release. And reconstruction is straightforward, because every sentence played has an identifier and a version.
There is also a plainer argument. In early collections, the conversation itself is short and has a narrow job: reach the right person, verify them, establish what has happened with the payment, agree a date within the bank's acceptable window, and confirm how to pay. Our historical data from banking shows an average handling time of 116.5 seconds in cases that end with a promise to pay. A conversation that tight does not need creative language. It needs exactly the right sentence at exactly the right moment.
The same logic holds outside banking. When an insurer's receivables process reminds a policyholder about an unpaid premium, the sentence that states the amount and what happens if it stays unpaid has been checked by the insurer's legal team. When a clinic's appointment process confirms a visit, the date, time and place it reads out must match the booking system exactly, and the cancellation option must be offered in the words the facility approved. When an energy supplier confirms a meter replacement, the technician's time window is a commitment someone will be held to. In each case the value of the conversation lies in precision, and precision is what a catalogue delivers.
Generation does have one clear advantage in these settings: it copes gracefully with the unexpected. But in a regulated call the unexpected is exactly where we least want a system to be inventive. A customer who mentions a bereavement, a legal dispute or a request to erase their data should not receive a creative reply. They should be routed into a defined exception path: a transfer to the client's staff, a structured stop or a referral, configured per market and recorded as a status the client can act on.
What it costs, and what it buys#
Closed catalogues are not free. We would rather name the trade-offs than pretend they do not exist.
- More design work up front. Someone has to anticipate the conversation: the questions customers ask, the objections they raise, the side-tracks they take. Our dialogue designers write and test those paths before launch, and add to them as listening reveals new ones.
- The long tail needs a route. A customer will eventually say something no designer anticipated. In a closed system the answer is a clarifying question, an honest "I can't help with that, I can pass you to a colleague", or a transfer to the client's own staff. Not a plausible improvisation.
- Less natural variety. A generative system can phrase the same thing a dozen ways. A catalogue has the phrasings someone approved. For most regulated tasks we think customers value clarity over variety, and we choose voices, pace and wording carefully so the call does not sound mechanical.
- Slower content change. A new sentence goes through approval before it is used. In a bank, that is a feature, but it does mean a change takes a release, not a prompt edit.
What it buys is predictability that you can prove. The bank's compliance team can read the full catalogue. Its legal team can sign off consequence language once. Its auditors can be shown the exact version that was live on any given day. And when a regulator, an ombudsman or an angry customer asks what was said on a specific call, the answer is a recording, a sequence of utterance identifiers and a script version, not a guess about what a model might have produced.
How a catalogue changes over time#
A catalogue is not frozen. It is versioned. Every day we listen to selected calls; every week we look at where conversations break off, compare dialogue variants and propose changes. In a regulated configuration each content change goes through the same path: the proposal, the client's content approval, a new version with an effective date, testing and release. Nothing is changed silently, and nothing learns in production on its own.
This is where running the service ourselves matters. We have no call agents to retrain. When a wording change is approved, it does not depend on a hundred people remembering a briefing. It applies to every live conversation from the moment it is released, and the version history records exactly when.
We don't manage people, so we control the service far better: one change and every live conversation already reflects it. – Dawid Wójcicki, Founder & CEO of Vee
The rules for a change are equally plain. If something on our side needs correcting, we correct it straight away. If the client's situation changes, for example a new product, a new verification requirement or a system that behaves differently than described, the change is approved on the client's side before it goes live. Either way it is versioned and approved, and either way the record shows what changed and why. How changes are scoped and settled is set out in the agreement.
Optimisation fits inside the same discipline. When we test two openings, two consequence phrasings or two voices against each other, both variants are approved catalogue content. The experiment chooses between approved options; it never invents a third one. With the client's agreement, those tests run on a large enough base to defend the result, and the winner becomes the new version.
Where generation genuinely helps#
None of this makes us sceptical about language models. We treat an LLM as a tool, not a solution, and we use it on some stretches of a process, in some domains, where it is justified. The question is never "LLM or not" for a whole deployment. It is which stage, which domain and what is at stake if the output is wrong.
- Understanding the customer. Using an LLM to classify what someone means reduces the training data needed and shortens launch, while the reply stays deterministic. That is our variant 3.
- Low-sensitivity processes. Where no financial, medical or legal consequence rides on the exact wording, and where a smaller business values flexibility over a fully specified catalogue, generative replies within limits make sense. That variant is technically still in build, and we say so.
- Behind the conversation. Analysis, design and quality work are not customer-facing. Tools that help our people review calls and spot patterns change nothing the customer hears without going through the same approval path.
The rule we apply is simple to state: the closer a sentence sits to money, health, identity or a legal consequence, the more firmly it belongs in an approved catalogue. The further away it sits, the more room there is for generation. Most enterprise processes contain both kinds of stretch, so the design question is about boundaries, not about choosing one technology for everything.
What the AI Act changes, and what it doesn't#
Since 2 August 2026, Article 50 of the AI Act has required that people are told when they are interacting with an AI system, and the European Commission published guidelines on these transparency obligations on 20 July 2026. Our AI always introduces itself as AI. That obligation applies whether replies are generated or drawn from a catalogue.
What the Act does not do is classify a customer-service process by its date or its label. The burden follows the architecture: whether the system generates content on its own, whether it makes or recommends materially consequential decisions, whether its behaviour changes after deployment without controlled change, and whether anyone can reconstruct what happened. A closed catalogue with separated statistical components gives a bank a much clearer starting point for that assessment. It does not replace the assessment. We help clients look at their own configuration component by component; we do not hand out blanket conclusions about it.
There is a market pattern worth noting here. Deloitte's research on AI-assisted customer service in US banks, published on 30 June 2026, found that 67% of the 30 US banking executives surveyed cite data security and compliance as a modernisation challenge. That is a small survey of US institutions, not Europe, but the concern is familiar from every bank we work with. Architecture that makes content governance tractable is one of the most direct ways to lower that barrier.
What to ask a supplier#
If you are evaluating voice automation for a regulated process, the demo will tell you little about any of this. A fluent, improvising bot can sound impressive for ten minutes and still be the wrong architecture for a bank. Ask instead:
- Can you show me the complete set of sentences your system can say in this process, and who approved each one?
- Which components in the call are deterministic, which are statistical and which are generative, and where exactly are the boundaries between them?
- Does any model generate customer-facing wording in this configuration? If so, how is that content governed and evidenced?
- Does the system learn or change in production without a versioned release? How would I know?
- When a customer says something nobody anticipated, what does the system do?
- For any given call, can you show me which script version was live and which utterances were played, in order?
Those answers matter more than the voice, the latency or the model name on the slide. They decide whether the process can be run, audited and defended for years, which is what a regulated organisation is actually buying. It is also how we build the processes we run for banks, insurers and healthcare providers: models to understand people, approved words to answer them.
Follow The First Second
Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.
Follow via RSS