We never launch a new telephone process on a client's whole database. We start on a sample, watch what real people do with it, correct what needs correcting and only then decide, together with the client, how much of the base comes next. This is not caution for its own sake. It is the cheapest insurance in customer contact, and in our experience the fastest route to a process that runs at full scale.

The recommendation surprises some clients. Everyone wants to see the process in production as quickly as possible, and so do we. But a database is a finite asset. Each person in it can be called for the first time only once, and the first impression a debtor, a policyholder or a patient forms of the call is very hard to undo.

Why the whole base is the wrong first audience#

A process that has passed every test still meets reality for the first time on launch day. Tests tell you that the dialogue works, that integrations exchange data correctly and that transfers land where they should. They cannot tell you how a particular population will react to a particular sentence at a particular hour. That only appears in production.

There is a second reason, specific to how we work. Vee employs no call agents; the conversations are executed by our AI. That gives us a property a traditional call centre does not have: one change applies to every live conversation at once. It is a great advantage when you are improving a process. It also means that an unnoticed weakness is repeated in every conversation until someone notices it. Starting on a sample keeps that exposure small while the process is still meeting its audience.

Some things are simply unknowable before launch. The input file may contain phone numbers in formats nobody checked. The client's sales or service team may receive transfers more slowly than planned. A call window that looked sensible on paper may coincide with the hour this population never picks up. None of these is a failure of preparation. They are facts about the real world, and a sample is how you learn them cheaply.

What has to be true before the first call#

A controlled pilot is not a substitute for preparation. It is the step after it. By the time we agree a start date, the process has been through atomic tests of each element and end-to-end tests of the whole path. Quite often the client also runs a “friends and family” stage, in which its own employees call or are called, try to break the dialogue and send their remarks. There are usually several rounds of corrections.

Before launch we also agree the operating rules in writing, because these decide how the sample behaves:

  • Schedule and call windows: which days and hours, set per market and enforced by the system rather than by good intentions.
  • Attempts and intervals: how many tries per record and how they are spaced.
  • Call-backs: whether we accept them, and until what time.
  • Transfers: in which situations the conversation goes to the client's own staff, to which number, queue or skill, and during which hours.
  • Exception paths: what happens when someone asks for a person, fails identity verification, withdraws consent to recording or reports hardship.

The criterion for starting is simple: the process behaves as it should in internal and external tests, or in the friends-and-family stage. When it does, we decide jointly to start. Sometimes there is a little more caution in the room than the evidence warrants; that is a conversation worth having openly rather than a reason to skip the sample.

What we watch in the first days#

The first days after launch are the hottest moment of any project. Both sides watch in real time. The client gets access to Vee Analytics, where it can load bases, follow progress and see how many records have already been “touched”, meaning the first contact attempts have gone out. It can see how many attempts were made, what each one produced, which conversations took place and whether a case has closed with a business status or is still in progress.

We read those signals in the order of our four-rung KPI ladder, because each rung depends on the one below it.

Early signals in a controlled pilot and what they usually point to
SignalWhat it tells usTypical correction
Reach (rung 1)Whether numbers, windows and retry spacing fit this populationContact timing, numbers and retry set-up
Continuation (rung 2)Whether people stay on the line after the opening and the company nameWording and order of the opening, voice, pace
Key information (rung 3)Whether the conversation obtains what it is for: a promise to pay, a confirmation, a full surveyQuestion order, phrasing, handling of objections
Technical statusesData quality in the input file, voicemail patterns, operator errorsData validation and technical set-up
Transfers and exceptionsWhether the human route works in practice and the receiving team keeps paceTransfer hours, pacing, routing rules

The fourth rung, the real business effect (payment, contract, attendance), arrives later and comes from the client's systems.

Numbers are only half of it. Our quality team listens every day, in every project, to selected recordings, especially unusual ones. In a pilot this matters more than ever. A dashboard shows that people hang up at a certain second; listening shows whether they hung up because the sentence was unclear, because recognition misheard a dialect word or because the line dropped. Only the last of these is not a design problem.

Every day, in every project, our people listen to selected calls, especially unusual ones, to hear with their own ears what technology still cannot. – Dawid Wójcicki, Founder & CEO of Vee

Measure the pilot the way you will measure the process#

A pilot is only useful if its results can be trusted, and that has to be designed in from the start. We compare our results with a comparator on the same buckets and the same portfolios, measured the way the client already measures its own call centre or external partner. A bucket, in collections, is a group of accounts at the same stage of delinquency; comparing an early bucket with a late one tells you nothing about the process.

We also separate declarative success from real success. A promise to pay (PTP) recorded in a conversation is a declaration, not a payment. The payment appears in the client's systems, often days or weeks later. So the pilot decision usually rests on the early rungs, while the business effect is confirmed as outcome data comes back, whether as periodic files we can match by contract or lead number, as monthly statistics or, ideally, as an automatic return after each closed base.

The decision to scale is made together#

After the first days we meet, look at the results and decide on the next base. It may be the same size again, with corrections, or it may already be larger. Sometimes the right answer is to be bolder than planned: a sentence that tested as risky turns out to work, or the population is more reachable than expected and the schedule can be compressed. Sometimes something escaped one side or the other during preparation and has to be fixed first.

The decision belongs to both sides because both sides own part of the answer. The client knows its portfolio, its allocation strategy and its budget; its budget owner gives the final approval. We know how the process behaved conversation by conversation and what we would change. If we think the process is not yet ready to carry a larger base, we say so, even when everyone would prefer to move faster.

The trade-off, stated plainly#

A sample costs time, and time is not free. In early collections, each day without contact moves some debtors towards the next bucket. In lead qualification, a fresh lead cools by the hour. So the pilot should be short and decisive, not a ritual. Its size has to be large enough for the signals to mean something for that population and small enough that a weak first version does not burn the base.

In practice the pilot rarely slows a programme down. The corrections it surfaces in the first days are the same corrections that would otherwise be made in week three, on a much larger and more expensive population. And because one change reaches every live conversation at once, a process that has been corrected on a sample scales without retraining anyone.

What to ask before your own launch#

  1. Which part of the base goes first, and who chose it? It should represent the buckets or segments you will scale into.
  2. What are the operating rules for the sample: windows, attempts, retry intervals, call-backs, transfers and exceptions?
  3. Which signals will be read in the first days, and in which order?
  4. Who listens to recordings, how often, and how do their findings reach the process?
  5. What is the comparator, and on which buckets and portfolios is it measured?
  6. How and when will outcome data come back from your systems?
  7. Who takes the decision to scale, and what will each side bring to that meeting?

A supplier who is happy to dial your entire database on day one is telling you something about how it sees the process: as capacity to be used, not as a service to be run. We would rather spend a few days on a sample and then run the process with you for years.

Follow The First Second

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

Follow via RSS