Most large organisations have a drawer full of good software. They bought it after a convincing demonstration, deployed it with real effort and then watched it fall quietly out of use, because nobody had designed and run the process the software was supposed to serve. Voice-agent platforms are now heading for the same drawer, and for the same reason: a platform that can hold a conversation is not a telephone process that delivers results.
We say this as a company that builds its own technology and has done so since 2012. We are not against platforms. We are against the assumption that buying one settles the question of who will make it work.
Companies have learned this lesson many times: they buy good software and it stays in a drawer, because nobody can design the process on their own. – Dawid Wójcicki, Founder & CEO of Vee
The tool is the easy part#
In 2026 it has never been easier to build a voicebot. Speech recognition is good in many languages, synthetic voices are pleasant, language models handle a wide range of phrasing, and several platforms will let a business user assemble a talking agent in an afternoon. That is real progress, and it changes the market. What it does not change is the distance between a talking agent and a customer-contact process that meets a bank's targets for months on end.
Building a voicebot is not the same as serving a customer. Between the two lies a very long road. – Dawid Wójcicki, Founder & CEO of Vee
That road is made of decisions nobody sees in a demonstration. Whom to call, from which number, at what hour, how many times and with what spacing. What to say in the first sentence so that people do not hang up after hearing the company name. How to verify identity before mentioning a debt. What to do when someone asks for a person, reports hardship or withdraws consent to recording. Which statuses to return to the client's systems, in which format, by when. Who listens to the recordings. Who notices that results dropped on Tuesday, and who is allowed to change the process on Wednesday.
A platform gives you a place to make those decisions. It does not make them for you, and it does not tell you when they were wrong.
Why the drawer happens even to good teams#
The organisations that end up with unused software are not careless. The pattern is structural. A project team is assembled to select and deploy the tool, and it does that job well. Then go-live arrives, the team returns to its other duties and ownership falls into the gap between IT, which runs the system, and the business, which owns the target. A telephone process needs attention every day, but nobody's job description includes listening to its calls, questioning its reach or rewriting its opening. Results drift, confidence fades and the next budget round asks why the business case was never met.
Meanwhile the vendor's own measures of success, such as licences sold, minutes consumed and features shipped, can look healthy throughout. That is not bad faith. It is simply that the vendor is paid for the tool, and the tool is working. The process around it is somebody else's problem, and that somebody was never appointed.
What a telephone process actually needs#
It helps to write out everything that must be true for an automated telephone process to run in a regulated enterprise. This is, roughly, the scope we take responsibility for in a managed service.
| Part of the process | With a bought platform | In a managed service |
|---|---|---|
| Telephony: numbers, SIP trunks, carrier routing, transfers, capacity | Usually the client or an integrator | The service provider |
| Contact strategy: call windows, attempts, retry rules, numbers and SMS | The client's team configures it | Designed and tuned by the provider within the client's rules |
| Dialogue design, voice, speech engine choice per language | The client's team, with platform templates | Dialogue designers, voice studio, speech specialists |
| Guardrails: identity verification before disclosure, exception and transfer paths | Configured by whoever builds the flow | Enforced at platform level and agreed with the client |
| Data exchange and statuses back to core systems | Integration project on the client's side | Provider integrates to the client's preferred method |
| Daily listening, analysis, A/B tests | Whoever has time | A dedicated quality and analytics team in every project |
| Change control, versioning, reconstruction of conversations | Depends on the client's discipline | Part of the service, documented and auditable |
| Incidents and SLA | Platform uptime only | Reaction, repair and communication for the whole process |
| Accountability for the business result | The client | Shared, with the provider's pay increasingly linked to it |
Simplified comparison. Allocation strategy, meaning which accounts are assigned and when, stays with the client in both models.
Read down the middle column and the problem is obvious. Each line is a job, and most require skills that a collections department, a customer-service function or a sales operation does not have in-house: telecoms engineering, dialogue design, speech technology, data integration, call analytics and compliance documentation. The platform licence is the smallest line in the real budget. The people to run it are the largest, and they are rarely in the business case.
Sector knowledge makes the gap wider still. In healthcare, a process that a platform would describe as appointment confirmation has a different real goal: learning early that a visit will not happen, so that the slot can be given to another patient. That changes everything about the design: when to call, how to follow up and how quickly a campaign must finish all follow from the goal of freeing the slot in time. In banking, the rule that no debt detail is disclosed before identity is verified has to be enforced by the platform itself, not left to a line in a script that someone might edit. None of this is in a template. It is knowledge of how customer service works in a particular sector, and it is the part of the service that comes before the technology.
Outbound is where the drawer fills fastest#
Clients often assume outbound is simpler than inbound. In our experience it is the other way round. Someone who calls you has a reason to stay on the line and will cope even with a clumsy IVR. Someone you call may not pick up at all, and if they do, may hang up within seconds. Every outbound process stands on the first rung of the KPI ladder: without reaching the customer, there is nothing else.
Reach is not a property of the voice agent. It depends on the outbound number and how carriers and handsets treat it, on the hour, on the spacing between attempts, on whether a text message went before the call, on the quality of the numbers in the file. A platform can dial. Knowing how to reach a particular population, in a particular market, at the right moment, is an operating skill built over millions of attempts. In early collections, where each day without contact moves debtors towards the next bucket, a few weeks spent learning that skill from scratch is expensive. In lead qualification, where a fresh lead cools by the hour, it can cost the campaign.
Judging by the demo#
We call the habit of judging automation by its voice, its model and its price per minute “AI theatre”. The market encourages it. In our view, presentations of voice automation tend to show how natural the voice sounds, how fast the model reasons, that it supports full duplex and barge-in. They say far less about the business effectiveness of a specific process: how many more customers paid on time, how many visits were saved, how many leads reached a salesperson, against which comparator.
This is not a new observation for us. In February 2021, at the HiperAutomatyzacja & AI conference, our talk was titled “Self-configuration or effective service: where the tool ends and results begin”. In a white paper published by a Polish speech-technology company at the end of 2020, our CEO argued that growth in voice automation would depend on the quality of the technology and on deployment experience. A year later, in a December 2021 review in BRIEF, he observed that many firms had rushed into cheap, ineffective bots and been disappointed. The tools have improved enormously since then. The gap between a tool and a result has not closed by itself.
If you must watch a demonstration, watch it differently. Ask to hear recordings from production rather than a scripted session: an unwilling debtor, a caller with a strong regional accent, someone who interrupts, someone who asks for a person. Ask what share of conversations reached the information the process exists to obtain, and what happened to the business result afterwards. Ask how long it took from the first workshop to stable results, and who did the work in between. A voice that sounds human is pleasant. A process that keeps people on the line, obtains a commitment and is honest about being AI, as the AI Act's transparency rules have required since 2 August 2026, is what you are actually buying.
What the research says about the gap#
Recent surveys of customer-service leaders describe the same pattern from the buyer's side. They are surveys of self-reported results, not audits, and they cover customer service broadly rather than telephone automation alone.
- Gartner's survey of 275 customer-service leaders, conducted in July and August 2026 and published on 4 November 2026, found that just 4% had achieved the return they expected from their AI initiatives.
- BCG's “Agentic AI Is the New Frontier in Customer Service Transformation” (3 December 2025), based on 180 leaders, reports that only 28% see measurable business value from generative AI in customer service, and identifies the operating model as decisive.
- Deloitte's “2026 Global Contact Center Survey” (4 June 2026) reports that among less mature organisations, 72% cite integration problems. The study shows a correlation between maturity and outcomes, not a proof of cause.
None of these findings says the technology does not work. They say that buying and deploying it is not where value is created. Value is created in process design, integration, measurement and continuous improvement: the work that happens after the purchase order, and the work most likely to be missing from it.
When a platform is the right choice#
We should be fair to platforms, because they are the right answer for some organisations. If you have, or are prepared to build, a team that covers dialogue design, telephony, speech technology, integration, analytics, quality listening and compliance documentation, a platform gives that team flexibility and control. If your process is simple, inbound, low in regulatory sensitivity and modest in volume, the long road may be short enough to walk yourself.
The honest test is whether you can name the people. Not the project sponsor, but the person who will listen to unusual calls every morning, the person who will notice that reach fell in one region and know why, the person who will rewrite the opening sentence and get it approved, the person who will answer the auditor's question about which script version was live on a given date. If those names exist and they have the time, a platform may serve you well. If they do not, the software is already on its way to the drawer.
There is also a middle path that rarely works: buying a platform and paying an integrator to set it up. The integrator leaves when the project ends. The process, which needs most of its attention in the months after launch, is left with whoever remains.
Whichever route you take, count the cost of time honestly. A team learning to run outbound contact from first principles will learn, but it will learn on your customers and your portfolios. In processes where delay itself is costly, such as early collections, perishable leads and appointment slots that can still be refilled, the months spent on that learning curve belong in the comparison alongside licence fees and service fees.
What changes when someone runs the process#
A managed service is a different answer to the same need. Instead of buying the tool and assembling the team, you buy the outcome of a team that already exists, together with the technology it built for the job. In our case that means a few concrete differences.
- Accountability for the whole process. We design, launch, run and improve it, from telephony to reporting. If something on our side needs correcting, we correct it straight away; how changes are scoped and settled is set out in the agreement.
- People who listen. Every day, in every project, our quality team listens to selected recordings, especially unusual ones. Analysts look for correlations and propose tests, which run with the client's agreement.
- Changes that land at once. Vee employs no call agents; the conversations are executed by our AI. An approved change reaches every live conversation immediately, with versioning and approval in regulated set-ups.
- Guardrails built into the platform. In regulated processes, what customers hear comes from a closed, versioned library, identity is verified before any debt detail, and every conversation can be reconstructed.
- An incentive pointing the same way. New agreements are no longer billed by the minute. A base fee per account or case, with an optional performance layer, rewards results rather than talk time.
- A rhythm that does not end. Weekly meetings, a workshop for each new base and shared documentation, for as long as the process runs.
Scale is the other thing a platform cannot sell you. By 31 October 2026 we had handled 48,626,959 cases, cumulatively, according to our operating data. A case is a record assigned for handling, not a person or a call, and every one of them was executed by Vee's AI rather than by people on headsets. Many campaign types, very high volumes, mostly outbound, across regulated sectors: that is where the operating knowledge described in this article comes from.
Behind those differences is something harder to buy: a team that combines customer service with telecoms, speech technology, heavy back-end systems, data at the scale of hundreds of millions of log entries, compliance and customer experience. People with that mix cannot be hired off the shelf. We grew them, project by project, through every phase of contact automation since 2012, and they are the reason a process keeps improving after the launch rather than slowly falling out of use.
Questions to ask before you buy#
Whether you choose a platform, a call centre or a managed service, these questions will tell you who will really run the process.
- Who designs the contact strategy, meaning windows, attempts, retry spacing, numbers and SMS, and who changes it when reach falls?
- Who writes and approves what customers hear, and how is each version recorded?
- Who provides and operates telephony, and who answers when a carrier fails?
- Who listens to recordings, how often, and what happens with what they hear?
- How are results measured, against which comparator, and in whose systems?
- When the process is wrong because the supplier missed something, how fast is it corrected, and what does the agreement say?
- What does the supplier earn more for: longer calls or better results?
- Which named people will still be working on the process a year after launch?
If the answer to most of these is “your team”, you are buying software, and you should budget for that team before you sign. If you would rather the answer were someone else's accountable team, that is the service we built Vee to provide.
Follow The First Second
Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.
Follow via RSS