Most AI voice business cases die in finance review. Not because the technology fails — because the model the operator built does not survive the questions a CFO asks. Cost-per-call drops from £7 to £0.40 and the spreadsheet says ROI is infinite. Then the finance team asks where implementation sits, what the change-management line is, what the attrition risk is, what happens to expansion revenue, and the model collapses.
Forrester's 2025 Total Economic Impact study of one enterprise voice platform found a 391% three-year ROI and average savings of $10.3 million per deployment. Gartner expects conversational AI to remove $80 billion in contact-centre labour costs by 2026. These are real numbers — but they describe outcomes, not the build that gets a programme to them. The model in this post is the one our enterprise buyers actually defend in front of boards. It treats AI voice as a programme, not a tool.
This guide is shipped by the team behind Dilr Voice — enterprise voice AI live in 40+ countries. Or see DATS, our 5-stage AI consulting system used to place voice programmes inside regulated enterprises.
A finance-grade AI voice ROI model has five lines, not one. Containment savings are the headline. The four lines underneath — implementation, attrition, expansion, and risk — are what the CFO will challenge.
- Cost-per-call deltas alone overstate ROI by 30–50% versus a finance-grade build
- Implementation and change costs typically sit at 18–24% of Year-1 savings — not zero
- Expansion revenue (cross-sell on captured calls, recovered abandons) is the line most models miss entirely
- Attrition and rebuild risk is real but small — and worth modelling explicitly so the CFO does not pad it
What the cost-per-call view leaves on the table
The dominant ROI framework in vendor decks is a single line: human-handled calls cost £7–£12 fully loaded; AI handles them for £0.30–£0.50; multiply by call volume; declare victory. That maths is correct. It is also inadequate. The work our AI placement diagnostic does inside a buyer is to convert that single line into a five-line P&L the finance team can defend — and to surface where the savings actually arrive in the calendar.
There are five reasons cost-per-call alone fails as a board-ready model.
1. Containment is partial, not total
A live voice agent does not contain 100% of calls. UK enterprise programmes we see in production land between 48% and 72% containment in steady state, depending on intent complexity and the floor the buyer sets on escalation. The other 28–52% still hit a human, and the human-call cost line stays in the model — proportionally. A model that assumes full containment overstates Year-1 savings by roughly the inverse of the escalation rate. See our deeper breakdown of AI voice cost per call benchmarks for the per-channel split.
2. Implementation is not free, and most of it is people
The line vendors quietly omit. A typical UK upper-mid-market enterprise programme carries £180k–£320k of Year-1 implementation cost before the first call lands: integration build, conversation design, testing, GDPR DPIA (£1.5k–£3k alone), data-residency configuration, and the internal programme team's time. The DPIA and data-governance work alone is non-trivial — see our guide on EU data residency obligations for voice AI. Hiding implementation in a "Year 0" footnote is the move that gets a model rejected.
3. Change management is the line that decides whether savings stick
A voice agent that automates 60% of inbound traffic without a workforce plan creates a politically expensive surplus of contact-centre seats. Programmes that cost-out the change line — redeployment, severance, retraining, supervisor restructure — at 8–12% of Year-1 gross savings survive the board. Programmes that pretend the savings net to zero against headcount overnight do not.
4. Expansion revenue is missed entirely
The most under-modelled line. AI voice does not just remove cost on inbound — it captures revenue that human teams miss. Recovered abandons (calls a human queue would have dropped), structured cross-sell prompts on every call, and outbound coverage at zero marginal cost compound into a real expansion line. We see expansion revenue contribute 15–25% of Year-1 net programme value across our deployments. Modelling it as zero is the most common cause of a buyer underselling their own business case.
5. Risk-adjusting the model
A finance team will discount aggressive savings projections by 20–35% for execution risk. A model that does this explicitly — risk-adjusts containment, attrition, and expansion separately — is treated as serious. A model that does not, gets cut by the CFO arbitrarily and loses the conversation.
The five-line programme model
Below is the structure we run inside our AI placement diagnostic for upper-mid-market enterprises. Numbers shown are illustrative for a programme handling roughly 150,000 calls per month at £4.20 blended cost per call.
The same diagnostic logic underpins our AI placement diagnostic — a fixed-fee assessment used before any deployment commitment. Buyers we work with use it to convert a vendor cost-per-call pitch into a board-ready model in 4–6 weeks.
The decision tree above is what a finance team wants to see. Each gate is challengeable; each input is sourceable. A model that walks a CFO through this structure tends to clear approval in one cycle. A model that hides any of these gates tends to bounce.
How the finance team challenges every line
The second half of the model is anticipating where each line will be challenged. The table below is the actual question grid finance teams use on enterprise voice AI proposals — and the evidence we ship inside our enterprise voice AI evaluation framework to defend each line. McKinsey's State of AI 2025 found that while 88% of enterprises now use AI, only 6% are AI-mature and 14% report material EBIT impact — the gap is rarely model logic; it is model defensibility.
| Programme line | What finance challenges | Evidence required | Typical risk discount |
|---|---|---|---|
| Containment savings | "Why this rate, not lower?" | 30-day pilot data, intent split, escalation logic | 15–25% |
| Implementation cost | "What's missing from the line?" | DPIA, integration scope, internal team allocation | 10–15% |
| Change management | "Where's the workforce plan?" | Redeployment plan, severance modelling, supervisor cost | Hard cost — no discount |
| Expansion revenue | "Why isn't this zero?" | Recovered-abandon rate, cross-sell uplift, outbound coverage | 30–40% |
| Attrition / rebuild | "What if vendor fails?" | Multi-vendor architecture, data portability, exit clause | Hard cost — model explicitly |
The last line is the one most operators avoid. AI infrastructure is consolidating fast — see our analysis of the SoundHound omnichannel acquisition and what platform consolidation means for buyer durability. Modelling £60k–£180k of rebuild reserve (depending on integration depth) is the move that demonstrates the operator has thought through vendor failure. Finance respects that.
The contrarian point worth making: BCG's Widening AI Value Gap study (Sep 2025) classified only 5% of enterprises as "Future-Built" AI organisations and 60% as Laggards — and the difference between groups was rarely the technology selection. It was the operating model behind the deployment. A robust AI operating model — clear RACI, embedded governance, defined hand-offs — is the line item most cost-per-call models do not even include, and it is what separates the 5% from the 60%.
A practical note on payback: the public stats showing 30–90 day paybacks come from highly contained, narrow-intent deployments (often a single use case at high volume). Enterprise programmes that span multiple call types, integrate with regulated systems, and carry a real change-management line typically land at 9–14 months risk-adjusted payback. That is still excellent — but it is not 30 days, and a model that promises 30 days will be challenged correctly.
For programmes touching financial services workflows, layer in FCA AI governance obligations — the governance build adds cost but materially de-risks the savings line by making them defensible to the regulator. The same goes for building a voice AI business case, which sits upstream of this ROI model and defines the use-case boundary that determines containment in the first place. Talk to us via the DILR.AI contact page if you want to see the worked model on your call mix.
Some buyers ask whether to model AI voice as a tool or a programme. The honest answer: until you cross roughly 50,000 calls a month, treat it as a tool with a cost-per-call model. Above that, it becomes a programme — and the five-line model becomes mandatory. Our enterprise voice AI agents guide walks through the architectural threshold in detail. According to BCG's 2025 study, the operating-model gap (not the technology gap) is what separates leaders from laggards — which is why finance-grade modelling matters more than vendor selection at this scale.
Want to see this model run on your numbers? Try Dilr Voice live (free, $20 credits), book an AI placement diagnostic, see our DATS methodology, or read about our deployment approach for placing voice AI inside regulated enterprise systems.
Build the AI voice ROI model your CFO will sign.
30-min scoping call. No deck. Confidential. We'll tell you whether the five-line model fits your call volume — and where the EBIT actually moves.
Written by the Dilr.ai engineering team — practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.
Sources: