A large language model is a tool, not a solution. It can make a contact process faster to build and better at understanding messy speech, but it does not decide who to call, when to stop, what a bank is allowed to say or how the result is measured. We use LLMs where they help, and we keep them out of the places where a single wrong sentence costs a client money, trust or a licence.

That position is not caution for its own sake. We have built conversational systems since 2012 and run them in production since 2017, through the era of rigid IVRs, through our own intent-understanding engine, and into the arrival of LLMs. Each phase brought something genuinely useful. None of them removed the need to design the process.

The question is never "LLM or not"#

Most debates about generative AI in customer service are framed as a single choice for a whole system. That framing is wrong. A telephone process is a chain of stages, and each stage has a different risk and a different payoff. The right question is: at this stage, in this domain, what does a language model add, and what could it break?

Our engine, Brilliance, is built so that question can be answered per process rather than per company. It runs in several variants, from fully deterministic to generative, and we choose between them according to how sensitive the process is.

Brilliance variants and where they fit
VariantHow it worksWhere it fits
HDeterministic dialogue, no LLM and no neural networks; closed content libraries, explicit transition rules, full auditabilityClients with the strictest compliance requirements
1 and 2Our own small language model used only to classify intent; the dialogue outcome stays deterministicRegulated processes that need more flexible understanding
3An LLM classifies what the customer said; the reply still comes from the state machine, and the LLM generates no repliesFaster launches with less training data, without giving up control
4Generative mechanisms produce repliesLow-sensitivity processes for smaller businesses; technically still being built

Vee architecture description. The principle behind the table: deterministic where it matters, generative where it helps.

Where an LLM earns its place#

The clearest win is understanding. People on the phone do not answer in tidy sentences. They hesitate, correct themselves, answer a different question, or say three things at once. An LLM used as a classifier can map that speech onto the intents a process expects with far less training data than a purpose-built model needs. In practice that shortens the time to a working first version, especially for a new language or a new type of conversation.

It is worth remembering that understanding was not invented by LLMs. Our first dialogue prototype in 2013 simulated a motor-insurance premium process: it understood that a driver who said they "put in oil" meant diesel, worked out annual mileage, and noticed when the make and model did not match. What LLMs changed is the cost and speed of reaching that level of understanding across many domains, not the principle.

There are other places where generative models help without touching the customer. Our own team uses AI tools in its daily work, as most engineering and analytical teams now do. The rule is the same as everywhere else: a person reviews the output before it becomes part of anything a customer hears or anything a client relies on.

Where we deliberately keep it out#

In regulated processes, such as early collections for a bank or receivables monitoring for an insurer, everything the customer hears comes from a closed, approved and versioned library of utterances. No new customer-facing language is generated during the call. This is a design choice, made with the client, and it rests on specific risks rather than general unease.

  • Untrue or incomplete statements. A generated sentence about a debt amount, a deadline or a consequence can be fluent and wrong. In a regulated conversation the wording itself is part of the bank's obligations.
  • Decisions with material effect. Whether a customer is offered a different path, transferred, or told about consequences must follow explicit rules and the client's own decision boundaries, not a model's inference about their situation.
  • Identity before detail. Debt details are never disclosed before identity is verified. That gate is enforced by the platform, not written as an instruction a model might reinterpret.
  • Silent change. Nothing learns in production. Changes to the interaction layer are versioned, approved and dated, so the process the client signed off is the process that runs.
  • Reconstruction. Each call can be rebuilt step by step from the recording, logs, statuses and the script version in force. That is much harder to ensure for free-form generated replies.

These are exactly the questions an auditor, a compliance officer or a regulator asks. Under the AI Act, the Article 50 transparency obligations have applied since 2 August 2026, and every Vee conversation opens by making clear that the caller is speaking with AI. Transparency, however, is only the beginning. A client must also be able to show what was said, why, and under which approved version.

We are not a software house. We don't build a process, switch it on and say: now it will fly by itself. – Dawid Wójcicki, Founder & CEO of Vee

Deciding stage by stage#

In practice we walk through a process and decide, for each stage, what kind of component belongs there. A simplified version for an early-collections call looks like this.

One process, different components
StageWhat decidesRole of a language model
Who is called and whenThe client's analysts and agreed contact rulesNone
Opening and AI disclosureFixed, approved wordingNone
Identity verificationPlatform ruleMay help recognise the answer; never opens the gate on its own
Understanding the customer's replyIntent classificationUseful, depending on the variant
What the customer hears nextState machine and closed libraryNone in regulated set-ups
Transfer, stop or referralExplicit exception rulesNone
After the callStatuses, QA listening, analysisCan support analysis, with human review

Illustrative mapping for a regulated outbound process; the exact configuration is agreed with each client.

The domain changes the answer. An appointment reminder for a small business carries far less risk than a conversation about an overdue loan, and a generative variant can be reasonable there. The same company may need both: a deterministic configuration for collections and a more flexible one elsewhere. Choosing one architecture for everything is how organisations end up either over-engineering trivial calls or under-protecting sensitive ones.

People stay in the loop, every day#

Keeping a model in its place is not only an architectural decision made once at launch. It is a daily practice. In every project, our people listen to selected calls, especially the unusual ones: the conversation where recognition may have been entitled to make a mistake, the customer who sounded upset, the call that ended in an unexpected status. Automated quality measures tell us where to look; human ears tell us what actually happened.

That listening is also how we decide whether a more flexible component is worth introducing. If we hear that customers in a particular process answer in ways a narrow classifier keeps missing, a stronger model at the understanding stage may pay for itself. If we hear that the problem is the wording of our question, no model will fix it, and the change belongs in the dialogue. The diagnosis comes first; the tool follows.

Because we employ no call agents, a change agreed with the client applies to every live conversation at once. That is a strength, and it is also why every change, including a change of model, goes through the same versioned approval as a change of wording.

Why "just use the LLM" is not a strategy#

In 2026 the cost of building conversational software fell sharply. Language models, AI-assisted programming tools and automation frameworks make it possible to produce, in weeks, what used to take a large team months. That is good news. It also means the model is no longer what distinguishes one service from another.

What still decides the result is everything around the model: reaching the customer at all, telephony and number strategy, retry timing, voice and wording, exception paths, integration with the client's systems, daily human listening, measurement against a comparator, and the discipline to change one thing at a time. A strong LLM inside a badly designed process produces a confident, articulate failure.

There is also a cost argument. Generative calls are not free to run, and the industry is only beginning to price that honestly. Using the heaviest tool at every step of every call is rarely the efficient choice; using it at the steps where it changes the outcome usually is.

What to ask a supplier#

If a supplier tells you its service is "powered by" a particular model, ask what that means at each stage of your process.

  • At which stages does a language model act, and at which does it not?
  • In our regulated processes, does any model generate words the customer hears? If so, how is that content approved?
  • Which rules are enforced by the platform rather than by a prompt?
  • Does anything learn or change in production without a versioned, approved release?
  • Can you reconstruct a single call, with recording, logs, statuses and the script version in force?
  • Can the architecture differ between our sensitive and our low-risk processes?

Good answers will be specific and a little boring. That is a healthy sign. We make these choices with each client, per process, and we revisit them as models improve, because the right place for a tool changes when the tool does.

Follow The First Second

Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.

Follow via RSS