Voice AI procurement: the CFO's 14 questions

By the time a voice AI deal reaches the CFO, it has usually already been declared a success. IT has confirmed the integrations. Legal has marked up the contract. Operations has watched a demo, run a pilot, and fallen a little in love with the call recordings. The deck says "ready to proceed." And then it lands on the finance director's desk, and stops.

This is not obstruction. It is the CFO doing exactly the job the rest of the business needs them to do. Everyone upstream has been asking does it work? The CFO is asking a different, harder set of questions: what does this actually cost us over three years, what is the real payback, what happens if it underperforms, and how do we get out? Most vendor materials — and most internal business cases — are built to answer the operational question. They are not built to survive financial diligence. That mismatch is why a programme that everyone wants can still die at the last gate.

This guide is shipped by the team behind Dilr Voice — enterprise voice AI live in 40+ countries — and the operators who sit on the other side of these finance reviews every week. For the consulting side, see DATS, our five-stage AI methodology.

This post is the checklist itself: the fourteen questions a finance team actually runs before signing a voice AI contract, grouped into the four areas a CFO cares about — unit economics, payback and return, risk and downside, and commercial structure and exit. For each one we give the answer a strong vendor and a well-prepared internal sponsor should be able to give, and the evasive answer that should make a CFO slow down. If you are the sponsor, build these answers before the review, not during it. If you are the CFO, this is the agenda.

Why the CFO's questions are different from everyone else's

It helps to understand why the finance lens looks so different from the operational one. The numbers explain it.

The cross-validated 2026 picture — from McKinsey's State of AI 2025 and BCG's The Widening AI Value Gap — is that almost everyone has adopted AI, only a third have it in production, and barely six per cent are seeing it move the profit line. A CFO has read those numbers, or lived them. They have signed off on AI spend before that did not change the EBIT, and they are not keen to do it again. Their scepticism is not anti-innovation; it is pattern recognition.

So while the COO is asking how the programme runs day to day — the operating cadence, the escalation rules, the weekly review — the CFO is asking whether the money is real. The two roles overlap but they are not the same agenda, and a sponsor who walks into the finance review with the operations deck will get sent away to do the work again. The questions below are the work. They sit downstream of the multi-stakeholder vendor evaluation — where IT, Legal, Finance and Operations each ask their own questions — and they are the deep dive on the finance column of that checklist. They are also the natural follow-on to the credit-attribution work covered in our piece on voice AI ROI attribution and the credit stack a CFO will sign: that post is about claiming the value after the fact; this one is about the diligence before the signature.

Group A — Unit economics: what does one interaction actually cost?

A CFO does not buy "a voice AI platform." They buy a cost per unit of work, repeated thousands of times a month. Before anything else, they want that number to be honest, complete, and stable as the programme grows.

Q1 — What is the fully-loaded cost per resolved interaction, not per minute? Vendors quote per-minute or per-call rates because they are small and they win demos. A CFO wants the cost per interaction that actually reached an outcome — a booking made, a query resolved, a payment taken — because an unresolved call that escalates to a human costs you twice. The good answer divides total programme cost by resolved interactions and shows both the AI-only number and the blended number including escalations. The evasive answer quotes a headline per-minute rate and changes the subject. Our breakdown of AI voice cost per call across human, hybrid and AI economics is the model to bring to this part of the review, because it counts the full loaded cost on both sides of the comparison.

Q2 — What is in the quoted price, and what is billed separately? This is where the headline number quietly grows. Telephony minutes, transcription, speech synthesis, large-language-model tokens, integration connectors, premium support, additional environments, and the analytics layer are all sometimes inside the price and sometimes line items on top of it. The good answer is a single page that shows what is included and what is metered. The evasive answer is "it depends on usage" with no worked example. The hidden lines are exactly what our piece on voice AI total cost of ownership and the costs vendors hide exists to surface — a CFO should assume the quote is the floor, not the ceiling.

Q3 — What is the run-rate at our steady-state volume, not the pilot volume? Pilots run at a few thousand calls a month and the economics look wonderful. Production runs at fifty or two hundred thousand, and the per-unit price often changes — sometimes down through volume tiers, sometimes up through overage. The CFO wants the recurring annual run-rate modelled at the volume the business will actually hit, with the tier breaks shown. Pricing a deal on pilot volume is one of the quiet reasons programmes stall on the way to scale, a failure mode we mapped in why most voice AI programmes get stuck in pilot purgatory and in our guide to designing the pilot-to-enterprise-scale transition.

Q4 — How does unit cost move as volume grows, and who absorbs overage? Related to Q3, but specifically about the shape of the curve and who carries the risk when forecasts are wrong. If you over-forecast, are you committed to minimums you will not use? If you under-forecast, what does overage cost per unit, and is it punitive? The good answer is a clear tiered table plus a stated overage rate. The evasive answer leaves overage "to be agreed" — which means it will be agreed when you have no leverage. That leverage question is the whole subject of our piece on buyer leverage in the 2026 voice AI funding cycle.

Group B — Payback and return: is the benefit real?

This is the heart of the review, and the part most business cases get wrong. Not because the maths is hard, but because the temptation to flatter the numbers is strong and a good CFO can smell it.

Q5 — What is the payback period, and what assumptions is it built on? Payback — the months until cumulative benefit covers cumulative cost — is the single number most CFOs anchor on for an operational technology purchase. They want it stated, and they want every assumption underneath it visible: volume, containment rate, loaded cost-to-serve, ramp time. A payback claim with no assumptions is not an answer; it is marketing. The discipline of building the benefit lines properly is the subject of our AI voice ROI framework covering total programme economics and the original enterprise business case for AI voice automation.

Q6 — What baseline are the savings measured against, and is that baseline real? Every saving is a delta from something. If you claim you are saving the cost of forty agents, the CFO will ask whether those forty agents existed, whether they will actually leave, and whether the work was going to grow anyway. Savings measured against an inflated or hypothetical baseline evaporate under scrutiny. The good answer ties the baseline to a number already in the management accounts — current cost-to-serve, current agency spend, current overflow rate. Defining what you measure, and measuring it before you start, is exactly the discipline in our guide to the KPIs an enterprise AI voice programme should track, and the containment number that drives most of the benefit is the one we benchmark in the 80% voice AI containment-rate procurement benchmark.

Q7 — How much of the benefit is hard cash versus reallocated capacity? This is the question that separates a credible business case from an optimistic one, and it is where you should volunteer honesty before the CFO extracts it. Hard cash savings are headcount or agency spend that genuinely comes out of the budget. Reallocated capacity — "the team gets twelve hours a week back" — is real value, but only if those hours are redeployed to something that generates or protects revenue. If they are not redeployed, they are not a saving; they are slack. A strong sponsor presents the hard-cash payback as the number to defend, and capacity as upside. This honest split is also the foundation of clean credit attribution later, which is why it connects directly to the ROI attribution credit stack, and why the capacity half so often depends on the change-management work that turns reclaimed hours into redeployed ones.

A worked payback model — the shape a CFO trusts

Here is an illustrative line-item model for a mid-market contact-centre deployment. The figures are representative, not a quote, and the point is the structure — the lines a finance team expects to see, and the deliberate separation of hard cash from capacity.

Illustrative model · internal · representative of engagements, not a guarantee of outcome.

The contrarian move here is to lead with the hard-cash payback of around fourteen months, not the nine-month figure that the capacity benefit produces. A CFO who is handed the optimistic number first becomes a sceptic for the rest of the meeting. A sponsor who hands over the conservative number first, and then shows the upside, earns the right to be believed.

Group C — Risk and downside: what happens if this goes wrong?

Operational sponsors model the base case. CFOs model the downside, because protecting the business from a bad outcome is half of what finance exists to do.

Q8 — What happens commercially if the system underperforms the SLA? A demo that works is not a guarantee that production will. The CFO wants to know what the contract does when containment falls below target, when latency spikes, or when accuracy drops — and whether there are service credits with teeth or just a polite apology. The good answer ties commercial remedies to the same metrics the programme is measured on. The evasive answer offers "best-efforts" support with no defined consequence. This is where the operational measurement framework and the commercial contract have to line up, and where an enterprise voice AI governance framework earns its place: the SLA is only enforceable if someone is watching the numbers that trigger it.

Q9 — What is the vendor's financial health and runway? The voice AI market in 2026 is well funded and consolidating fast, which cuts both ways. A CFO signing a multi-year commitment wants to know the vendor will still exist — and still own its own roadmap — at renewal. Is the company profitable or burning? When did it last raise, and how much runway does that imply? Is it an acquisition target, and what happens to your contract if it is acquired? We unpack how to read those signals in voice AI valuation signals for enterprise procurement and what consolidation does to the buyer in de-risking the buyer stack against vendor consolidation.

Q10 — How concentrated is our dependence, and what is the switching cost? The CFO is asking about lock-in. If everything — telephony, models, data, configuration, analytics — sits with one vendor, the switching cost at renewal is enormous and the negotiating position at year three is weak. The good answer is a clear-eyed account of what is portable and what is not, and a deliberate decision about how much concentration is acceptable for the simplicity it buys. This is the commercial face of the architecture decision we lay out in voice AI orchestration versus a managed platform and the broader build, orchestrate or buy decision for 2026 — and the same instinct that drives the wider enterprise voice AI platform selection criteria.

Group D — Commercial structure and exit: what are we actually signing?

The final group is where finance and legal sit side by side. It is worth being precise about who owns what. Legal owns the clauses — the words in the contract. The CFO owns the commercial terms those clauses express: the money, the duration, the escape routes, the treatment on the books. The clauses themselves are a subject in their own right, and we have written the full checklist separately in the 11 MSA clauses enterprise legal teams should require. What follows is the commercial lens on the same territory.

Q11 — What is the contract term, notice period, and exit cost? A three-year term with auto-renewal and ninety days' notice is a very different commitment from an annual rolling contract, and the CFO wants the real shape of it: minimum spend, early-termination cost, what happens to your data and configuration on exit, and how long transition assistance lasts. The good answer treats exit as a designed event, not an afterthought. The evasive answer makes leaving expensive and undocumented — which is itself information.

Q12 — How does price escalate at renewal, and is there a true-up clause? The first-year price is the marketing price. The CFO wants the renewal mechanics: is there a capped annual uplift, or is renewal "at then-current rates"? Is there a true-up that re-prices you upward if volume grew, with no symmetric reduction if it shrank? Unbounded escalation is how a sensible year-one deal becomes an expensive year-three one. The good answer is a stated cap; the evasive answer is silence.

Q13 — Who owns the data and the configuration, and what are the indemnities? Commercially, the CFO cares about three things here: that your call data is not used to train models you do not control, that you keep the configuration and prompts you have invested in building, and that the vendor indemnifies you against the liability its system could create — a non-compliant call, a data breach, a hallucinated commitment. The detail of how that data is held and for how long is covered in voice AI data retention under GDPR and enterprise voice AI data residency; the audit rights that make indemnities enforceable are the subject of voice AI auditability and the procurement gate most vendors fail.

Q14 — How is this treated on our books: opex versus capex, committed versus consumption? This is a genuinely CFO-specific question that operational sponsors almost never anticipate. A consumption-based contract is operating expenditure that flexes with usage; a large committed minimum or an on-premise build may be capital expenditure that has to be depreciated. The treatment affects the P&L shape, the cash profile, and sometimes which budget the spend even comes from — a distinction that matters enormously in, for example, public-sector procurement where capital and revenue budgets are separate. The good answer lets the CFO choose the structure that fits the books; the evasive answer assumes one model and will not flex.

How the 14 questions shift weight by sector

The questions do not change, but which one decides the deal does. A finance team in a regulated sector weights the contract differently from one buying outbound sales automation. Here is where the make-or-break question tends to sit.

Walk in with the answers already built — five steps

The single best predictor of whether a voice AI deal clears the CFO is whether the sponsor did this work before the meeting or is improvising inside it. Here is the preparation, in order.

Step 01 — Build the unit-economics model on cost per resolved interaction. Start from the number a CFO can defend: total programme cost divided by resolved outcomes, with the loaded human comparison alongside it. Use the cost-per-call model and pressure-test it against the hidden-cost checklist so nothing surfaces in the room that you have not already counted.

Step 02 — Agree the baseline before you claim a saving. Pin the baseline to a figure already in the management accounts and write down how you will measure the delta. If you cannot point at the baseline today, you cannot claim the saving tomorrow. The programme KPI set is where that measurement discipline lives.

Step 03 — Separate hard cash from reallocated capacity, and say which is which. Two lines, never one. Defend the payback on hard cash; present capacity as upside conditional on redeployment. This is the honesty that survives credit attribution six months later, when finance comes back to check whether the promised value actually landed.

Step 04 — Model the downside, not just the base case. Show the CFO what happens if containment lands ten points low, if volume undershoots, or if the vendor is acquired. Tie the contract's remedies to those scenarios. A downside you have already modelled is a downside the CFO does not have to fear.

Step 05 — Get the exit and escalation terms in writing before signature. Renewal caps, exit cost, data portability and indemnities are cheap to secure before you sign and expensive to negotiate after. Bring the clause checklist to legal early, and remember that your negotiating position is strongest precisely when the vendor still wants the logo — the dynamic we cover in buyer leverage in the funding cycle.

Where this sits in the wider decision

The CFO review is one gate among several, and it is most powerful when it connects to the others rather than standing alone. The finance questions sit downstream of the operational decision about how the programme will actually be run — the weekly cadence, the escalation rules, the cross-functional ownership that a COO holds, and that we treat in our AI operating model consulting. They sit alongside the legal review of the contract clauses. And they depend on the upstream architecture choices — build, orchestrate or buy; one vendor or several — because those choices set the cost base the CFO is being asked to fund. A finance review run in isolation from the operating model and the architecture will get the numbers right and the decision wrong.

Want to pressure-test this before your own finance review? Try Dilr Voice live (free, $20 credits), explore DATS, our five-stage AI methodology, or read about our approach to placing AI where the P&L actually moves.

Frequently asked questions

What is the single question voice AI deals most often die on at the CFO review?

Question 7 — the split between hard cash and reallocated capacity. Sponsors routinely present the full gross benefit as if it were all cash, and a CFO who spots one inflated line stops trusting every line. Lead with the hard-cash payback, present capacity as conditional upside, and the rest of the review goes far better.

Is voice AI usually treated as opex or capex?

Most consumption-based voice AI contracts are operating expenditure that flexes with usage, which many CFOs prefer because it carries no depreciation and scales with the benefit. A large committed minimum, a multi-year prepayment, or an on-premise build can shift the treatment toward capital expenditure. Ask the vendor to support whichever structure fits your books — Question 14 — rather than assuming theirs.

How short does the payback period need to be before a CFO signs?

There is no universal threshold, but for operational automation many CFOs look for a hard-cash payback inside twelve to eighteen months, with the full return visible across the contract term. The number matters less than the credibility of the assumptions underneath it — a defensible eighteen-month payback beats an unbelievable six-month one every time.

Should the CFO or Legal own the contract terms?

Both, on different layers. Legal owns the clauses — the wording, the liability, the data-protection terms, detailed in our MSA clause checklist. The CFO owns the commercial terms those clauses express: price escalation, exit cost, committed spend, and accounting treatment. They should review the contract together, not sequentially.

How do we assess a voice AI vendor's financial health?

Look at profitability or burn rate, time since last raise and implied runway, customer concentration, and whether the company is a plausible acquisition target. Filed accounts, funding announcements and the consolidation pattern in the market all help — we walk through reading those signals in voice AI valuation signals for procurement. Then make sure the contract protects you if ownership changes.

Written by the Dilr.ai engineering team — practitioners who ship enterprise AI in production and sit on the buyer's side of these finance reviews. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.