Deloitte’s Future of Service reports that 43% of organisations surveyed believe AI can cut contact-centre costs by at least 30% within three years. That is a finding about expectations, not about savings anyone has banked, and the difference matters most to the people who will be asked to deliver the number.
We read the publication as it stands, put it next to other research on realised returns, and explain how we think cost should be measured in regulated customer contact.
What Deloitte published#
“The Future of Service: The Age of Intelligent Experience” appeared on Deloitte’s US press pages on 16 February 2026. It is closer to a playbook and announcement than to a full research report, and the public summary we worked from does not set out the sample size or method in detail. Its headline figure, which Deloitte attributes to its then-forthcoming Global Contact Center Survey of service leaders, is clear, though: 43% of the organisations surveyed believe AI can reduce contact-centre costs by 30% or more within three years.
Two things follow from the wording. The figure measures belief about the future, so it cannot be cited as evidence that costs fall by 30%. And a majority of respondents did not report that expectation, at least not at that size or over that horizon. A fair summary is that a large minority is optimistic about a substantial cost reduction.
Why an expectation is not a saving#
Expectations drive budgets. When a board hears that four in ten peers expect a cut of a third, the number tends to migrate from a survey into a target. Somewhere in that journey the word “expect” is dropped, and a contact-centre director inherits a commitment that nobody measured.
The three-year horizon adds its own risk. A target set today will be judged against a contact centre that has changed in the meantime: different volumes, different customer mix, different regulation, including AI Act transparency duties that have applied since 2 August 2026. A saving that cannot be separated from those changes cannot really be claimed, however confident the original expectation was.
Other research suggests how wide the gap between expectation and outcome can be. Gartner reported on 4 November 2026 that just 4% of 275 customer-service leaders, surveyed in July and August 2026, had achieved the return they expected from their AI initiatives. Gartner has also forecast, in March 2025, that agentic AI could autonomously resolve 80% of common service issues by 2029, with a potential 30% cut in operating costs. That is a forecast about common issues, not all calls, and it is not a measurement either.
| Claim | Source | What kind of evidence |
|---|---|---|
| 43% expect ≥30% cost cut within three years | Deloitte, 16 Feb 2026 | Survey of expectations |
| 4% achieved the AI return they expected | Gartner, 4 Nov 2026; 275 leaders | Survey of reported outcomes |
| Agentic AI to resolve 80% of common issues by 2029; potential 30% cost cut | Gartner, 5 Mar 2025 | Forecast |
Each figure belongs to its own publication and population. They should not be combined into a single estimate.
Cost per what?#
Much of the confusion begins with the unit. A contact centre can cut its cost per minute and still spend more per result, if conversations get longer, reach falls or customers call back. It can also cut cost per call while the business outcome, a payment, a renewal, an attended visit, stays where it was.
We argued this publicly as early as 2022: a solution with a low unit price and a 2% success rate can end up costing more than twice as much per success as one with a higher unit price and a 30% rate. The arithmetic was illustrative, but the principle has held. The cost that matters is the cost of a single success, and that depends on effectiveness at least as much as on price.
If we are paid by the minute, we are not playing towards the same goal as our client. – Dawid Wójcicki, Founder & CEO of Vee
That is why new Vee agreements are no longer billed by the minute. Our model has a one-off setup, a base operating fee per case or account that does not depend on outcome, and an optional performance layer where the result can be measured in the client’s own systems. It is also why, in September 2026, Vee stopped regular reporting of minutes and case volumes in its public disclosures, in favour of reporting on economic value, operational effectiveness and process economics.
What realised savings look like in practice#
We do not publish cost-saving multiples, and we would treat any supplier who promises a fixed percentage with caution. What we can show are effects measured on specific processes, with their scope. In one lead pre-selection campaign, historically, interested leads rose by 26% and customer acquisition cost fell by 73%. In early-stage collections, historically, on-time payment frequency rose by 7–13% versus a human-agent comparator.
Neither figure is a forecast, and neither transfers automatically to another portfolio. They show the shape of a realised effect: one process, one comparator, one measurement window, one source of truth. Before any outcome-linked arrangement, what counts as success and how it is measured are set out in the agreement with the client.
There is a further reason to be careful with cost alone in regulated sectors. For a bank, the most painful part of early delinquency is not the cost of the call; it is the delay that pushes a customer into a later bucket and forces the bank to build reserves. A cheaper contact that arrives later can be the more expensive one.
Where cost really moves in an automated operation#
When savings do appear, they rarely come from the line item people expect. In our experience the larger effects come from three places. The first is capacity that follows infrastructure rather than hiring: in March 2020, when demand from large clients rose roughly fivefold, channel capacity at one mobile operator was raised from 250 to 900 without recruiting anyone. The second is the speed of change: one approved adjustment applies to every live conversation at once, with no retraining. The third is speed of contact, because a lead cools by the hour and a debtor who is not reached drifts into the next bucket.
None of these shows up neatly as a percentage cut in contact-centre cost. Some of them show up elsewhere in the business, in collections, in sales conversion or in a filled appointment slot. That is one reason why a single cost-reduction figure, whether expected or achieved, is a weak basis for a decision about customer contact.
Cost can also move the other way. Gartner has forecast that the cost per resolution of generative AI in customer service will exceed $3 by 2030, above the cost of many offshore human agents. It is a forecast and not a statement about API prices, but it is a useful corrective to the assumption that automation only ever gets cheaper. The unit that protects a buyer is the cost of a verified result.
How to use Deloitte’s figure well#
The 43% is a reasonable data point about sentiment. It can help explain why budgets for AI in service are rising and why expectations from boards are high. It should not appear in a business case as an estimate of what your contact centre will save.
- Quote it as an expectation. “43% of organisations surveyed by Deloitte believe…”, with the date.
- Separate it from outcomes. Put realised-return research next to it, with its own sample and date.
- Choose the unit first. Agree whether you are measuring cost per minute, per case, per contact or per success, and why.
- Insist on a comparator. Savings without a comparator on the same buckets or portfolios cannot be attributed.
- Count the outcome, not the activity. A promise to pay is not a payment; a booked visit is not an attended one.
What to ask a supplier#
If a supplier quotes an expected saving, ask where it comes from. If it is an analyst figure, it describes the market’s hopes, not the supplier’s results. If it is the supplier’s own result, ask for the process, the period, the comparator and the scope. And ask how the supplier is paid: a model tied to minutes rewards long conversations, while results often come from shorter ones.
Deloitte’s publication captures an optimism that is real and, for some processes, may prove justified. The way to find out is not to adopt the expectation but to test it: one process, on a sample, measured against a comparator, with the cost of a single success as the number that decides whether to scale. That is the order in which we work with clients.
Follow The First Second
Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.
Follow via RSS