Accenture describes an intelligent service centre that, for US federal agencies, could deliver a potential 40% reduction in operating expenses, and in a separate report argues that as AI gains autonomy, control and trust must grow with it. Both points are sound. Neither is a result, and the distance between a projection and a running service is where most of the work lies.
We run an AI call centre for regulated enterprises: conversations executed by Vee’s AI, with the people at Vee designing, listening, analysing and improving the process. Below we read Accenture’s two publications as operators and set out what they get right, what they cannot show, and what we would add.
The two publications#
| “The Intelligent Service Center” | “Technology Vision 2025: AI – A Declaration of Autonomy” | |
|---|---|---|
| Date | 22 July 2025 | 7 January 2025 |
| Audience | US federal government | Business and technology leaders broadly |
| Main point | Potential 40% cut in operating expenses through a phased approach; pressure to modernise infrastructure | Growing AI autonomy requires control, and the trust of employees and consumers |
| Evidence type | Projected potential; derivation not published | Broad technology-trends report |
| Main caveat | The 40% is a potential for US federal agencies, not an observed result | Not specific to telephone contact or customer service |
Sources: Accenture, 22 July 2025 and 7 January 2025. Summarised; read the originals for the full argument.
Accenture is a consultancy and systems integrator that sells transformation services, which is worth bearing in mind when it describes the potential of transformation. That does not make the analysis wrong; it makes it an argument for a course of action rather than a neutral measurement.
“A potential 40%”: how to read a projection#
The service-centre piece says that a phased approach to an intelligent service centre can lead to a potential 40% reduction in operational expenses for federal agencies, and elsewhere calls that cut a goal. Three words carry the weight: “potential”, “phased” and “federal”. It is a projection whose derivation the page does not set out, it is a target rather than a measured result, and it is set in a public sector with its own cost structures, legacy systems and procurement rules.
We would not carry that figure into a European bank’s or insurer’s business case, and we do not borrow analyst savings figures for our own claims. A cost reduction is only real when it has been measured on a specific process against a comparator, and even then it belongs to that process. What the projection does usefully is show the direction of pressure: service organisations are being asked to do more with ageing infrastructure, and that pressure is not unique to government.
What an intelligent service centre needs in practice#
The phrase “intelligent service centre” suggests a technology stack. From the inside, it looks more like a set of responsibilities that someone has to carry every day. In our managed service those responsibilities cover the whole chain, because a customer contact fails at its weakest link, not its strongest.
- Telephony. Numbers, SIP trunks, call-attempts-per-second limits, routing, transfers and secure connections. We run our own.
- Dialogue and process. The conversation design, its exception paths and the agreed transfer rules.
- Contact strategy. Dialler settings, schedules, retries, number strategy and SMS before or after a call.
- Data exchange. REST API, SFTP flat files or HTTPS, so that statuses reach the client’s systems.
- Reporting and analysis. Vee Analytics, daily case-level statuses and correlation work by analysts.
- Support. An incident team working to an agreed SLA, and a QA team listening to calls every day.
None of these is new, and none is solved by buying software. An intelligent service centre is intelligent only if someone owns the whole chain and answers for the result. Where responsibility is split between a platform vendor, an integrator, a telecoms provider and an internal team, the gaps between them are where customers fall.
Ownership also has a time dimension. Our delivery path runs from a performance diagnostic, which includes listening to the client’s own call recordings and saying plainly when a process will not pay back, through contract and data-protection work, a three-part specification, testing and a controlled pilot on a sample of the base, into managed production. Production starts with an intensive rhythm and settles into weekly and then monthly reviews, and it continues for years as new buckets, portfolios, processes and countries are added. A service centre that is designed once and then left alone does not stay intelligent for long.
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
Autonomy needs control: Technology Vision 2025#
Accenture’s Technology Vision 2025, subtitled “AI – A Declaration of Autonomy”, makes a broad argument: as AI systems act with more independence, organisations need stronger control, and they need the trust of both employees and consumers. The report is not about contact centres, and we draw nothing from it beyond that central point.
The point is one we agree with, and in regulated customer contact it has a concrete shape. We do not describe our service as autonomous. Processes include agreed transfers to the client’s own staff, and the boundaries of what the AI may do are set in advance and approved by the client.
What control looks like on a live call#
- Bounded replies. In regulated processes replies come from a closed, versioned library. Language models may classify intent; they do not compose what the bank or insurer says.
- No learning in production. The system’s behaviour changes only through approved, versioned releases.
- Verification first. No debt detail is disclosed before identity is verified; this is a platform rule, not a script line.
- Full reconstruction. Every call can be rebuilt from recording, logs, statuses and the script version in force.
- Approved change. If something on our side needs correcting, we correct it straight away; client-side changes are approved by the client; every content change in a regulated set-up is versioned and approved.
Our own technology reflects the same principle at architectural level: deterministic where it matters, generative where it helps. Variants range from fully deterministic dialogue with no language model, through small models used only to classify intent, to generative variants for low-sensitivity processes, which are still being built. The choice is made per process, not per fashion.
Trust of employees, trust of customers#
Accenture names two audiences whose trust matters. On the customer side, the starting point is honesty: Vee always introduces itself as AI, and since 2 August 2026 Article 50 of the AI Act has required transparency for AI interactions in the EU. On the employee side, the relevant people in our model are the client’s own staff who receive transfers, and the client’s analysts who keep control of allocation strategy, meaning which accounts are contacted and when.
That division is deliberate. We run the contact; we do not take over the client’s credit policy, clinical judgement or sales targets. Trust is easier to build when each side knows exactly what it owns.
What we see that a projection cannot#
Projections describe a system at rest. Running one shows how it behaves under load and over years. Our cumulative operating data to 31 October 2026 records 48,626,959 cases (a case is a record assigned for handling, not a person or a call) and 298,684,764 single dial attempts. Numbers on that scale teach things a model does not: that reach is the first constraint on everything else, that time windows and number strategy differ by market, and that small changes to an opening sentence move hang-up rates measurably.
They also teach patience about savings. The effect we can show is incremental and scoped: in early-stage collections, a 7–13% increase in on-time payment frequency against a human-agent comparator (historical). That is a different kind of claim from “a potential 40%”, and deliberately so.
What to ask a supplier of an intelligent service centre#
- Who owns the whole chain, from telephony to the status in our system, and answers for the result?
- Which savings figures are measured on a comparable process, and which are borrowed from projections?
- What may the AI decide on its own, what is fixed in advance, and who approves changes?
- Can any single call be reconstructed for an auditor?
- How is the client’s own staff involved in transfers and exceptions?
Accenture is right that service operations are under pressure and that autonomy without control is a liability. The practical answer is a service run by people who own it end to end, with boundaries set before the first call. That is the kind of service centre we run.
Follow The First Second
Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.
Follow via RSS