Most customer-service organisations that bought AI in the last two years are not getting the return they planned for, and the reason is operational rather than technological. The model usually works; what fails is everything that has to happen around it, every day, for the model to produce a business result.

Two surveys make the point with unusual clarity. Read together, they describe a market that has learned to deploy AI quickly and has not yet learned to run it. We have run telephone customer contact on our own AI since 2012, at production scale since 2017, so we have watched this gap open and close in many processes. Here is what the numbers say, and what we think they mean.

What the two surveys actually found#

In December 2025, BCG published Agentic AI Is the New Frontier in Customer Service Transformation, based on a survey of 180 customer-service leaders worldwide. Only 28% said their companies had achieved measurable business value from generative AI in customer service. BCG's own conclusion was that success depends on building the right operating model around the technology, and that the obstacles are organisational and strategic as much as technical.

In November 2026, Gartner published the results of a survey of 275 customer-service and support leaders, conducted in July and August 2026. Just 4% said they had achieved the expected return on investment from their AI initiatives. Gartner also offered a forecast: by 2028, organisations that build portfolios of linked use cases, rather than isolated ones, may get twice the value. That second figure is a prediction, not a finding.

Two surveys, one gap
BCGGartner
Published3 December 20254 November 2026
Sample180 customer-service leaders275 service and support leaders, July–August 2026
Finding28% report measurable business value from GenAI4% achieved the expected ROI from AI
TypeSelf-assessed surveySelf-reported survey; the 2028 figure is a forecast

Both are surveys of leaders reporting on their own organisations. They measure different things: value achieved versus return against expectation.

The two numbers should not be combined or compared directly. One asks whether any measurable value exists; the other asks whether the return met the plan. But, as we read them, they point the same way. Roughly one organisation in four can show value, and almost nobody got what they expected when they signed.

Deploying a model is not running a process#

A model is a component. A customer-contact process is a chain of decisions that starts long before anyone speaks and ends long after the call: which cases get contacted, from which number, at what hour, how many times, with what words, what happens when the person asks for a human, what gets written back to which system, and who checks that the result was real.

Most AI investments buy the conversational component and assume the rest will follow. It rarely does. In our experience the conversation is perhaps the most visible fifth of the work. The remaining four-fifths are operational, and they decide the result.

  • Reach. Telephony, number strategy, contact timing and retry rules, and SMS around the call, tuned for each market and portfolio. Without reaching the customer, nothing else matters.
  • Integration. Cases arrive from the client's systems and statuses go back, daily, in a form the client's analysts can act on: REST, SFTP flat files, VPN, SIP trunks. Integration adapts to the client, not the other way round.
  • Exception paths. A failed identity check, a request for a person, a withdrawn recording consent, a death or a legal dispute. Each needs a designed route: transfer, stop or referral.
  • Quality assurance. People listening to selected calls every day, especially the unusual ones, and turning what they hear into changes.
  • Measurement. Comparing results against a comparator on the same buckets and portfolios, in the client's systems, rather than counting what was said on the call.
  • Change control. Versioned content, approved by the client where regulation requires it, applied to every live conversation at once.

None of these is a model feature. All of them have to be owned by someone. When nobody owns them, the AI works in the demo and stalls in production.

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

Where the value leaks: a typical early-collections deployment#

Consider a bank that assigns accounts in early delinquency to an automated voice process. The aim is not the call; it is payment before the debtor slides into the next bucket and forces the bank to build reserves. Now follow the value through the chain.

If the dialling strategy gives up too early, a large part of the assignment is never reached, and the model never gets a chance. If the opening sentence is badly worded, people hang up after hearing the company name. If identity verification is not enforced by the platform before any debt detail is disclosed, compliance will stop the project. If the conversation records a promise to pay but nobody checks whether the payment arrived, the reported success rate is a declaration, not a result. And if statuses do not reach the bank's systems the same day, its own team cannot act on them.

Each of these leaks is small. Together they explain why an organisation can deploy a capable model, see impressive conversations in testing, and still fail to find the return in its books. This is why we measure every process on a four-rung ladder: reach, continuity, key information, business outcome. A gain at the top rung is only possible if the rungs below it hold.

Why the gap is wider in regulated sectors#

Banks, insurers, healthcare providers, telecoms and energy companies face a further layer. Every automated conversation has to be explainable after the fact: the recording, the logs, the statuses and the script version that was live at that second. Article 50 of the AI Act has required transparency about AI interaction since 2 August 2026. GDPR and, for financial clients, DORA shape how data and suppliers are handled.

These requirements do not block value, but they change where it comes from. In regulated processes we let models recognise what the customer means and select from a closed, versioned library of approved replies; they do not invent answers. In our experience that design has not cost results, and it means the process can survive an audit. Organisations that deploy a generative model first and ask the governance questions later often discover that the answer is a redesign, and the redesign is where the expected return goes.

We have seen this cycle before#

The current gap is not new. In a white paper on voicebots published by a Polish speech-technology company in December 2020, our CEO argued that market growth would depend on the quality of the technology and the experience of the teams deploying it, not on the availability of tools. A year later, in a December 2021 year-in-review in the marketing title BRIEF, he described companies that had rushed into cheap, ineffective bots and been disappointed.

Large language models made conversational software far easier to build in 2026. That is good for the market. It also means more organisations can reach the deployment stage without having solved the operating one, which, in our view, is what the 4% and the 28% describe. The barrier to building has fallen; the barrier to running has not moved.

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

What closing the gap looks like in practice#

In our experience, the organisations that find value tend to do a few unglamorous things consistently. They start on a controlled sample of the database, never the whole of it, and agree in advance what success means and where it will be measured. They compare against a comparator on the same buckets rather than celebrating an absolute rate. They keep allocation strategy with their own analysts and let the supplier optimise reach and conversation. They meet weekly, listen to calls, and change one thing at a time.

This is also why we run the service ourselves rather than sell a platform. When the same team answers for telephony, dialogue, integration, quality and reporting, there is no gap between components for value to fall into. Where we have published results, they carry their scope: for example, +7–13% in on-time payment frequency in early-stage collections, historical, against a human-agent comparator. That is an increment on a defined base, not a headline rate, and it is the kind of number that survives a finance review.

Questions to ask before the next AI investment#

  1. Who owns reach: numbers, call windows, retry logic and SMS? If the answer is “the client”, budget for it.
  2. What is the definition of success, when is it measured, and in which system? A recorded promise is not a payment.
  3. What is the comparator, and is it on the same buckets and portfolios?
  4. How are exception paths designed, and who approves changes to them?
  5. Can every conversation be reconstructed step by step for an auditor?
  6. How often does someone listen to real calls, and what happens with what they hear?
  7. Is the supplier paid for capacity, such as minutes, or for progress on the KPIs you care about?

If these questions have clear owners and answers, the model choice becomes a secondary decision. If they do not, a better model will not close the gap. That is the lesson the surveys are teaching the market in 2026, and it is the one we have been working on since long before the word “agentic” arrived.

Follow The First Second

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

Follow via RSS