Voice AI run cost optimisation: the post-launch playbook

Most enterprises optimise the cost of a voice agent exactly once, before it launches. The business case gets modelled, a platform gets chosen on price, and the programme goes live. Then the meter starts running and nobody looks again. In 2026 that is an expensive habit: McKinsey's State of AI found 88% of organisations now use AI, but only 33% have it in production and just 6% are AI-mature, so the enterprises that already run voice agents at scale are the ones with the most unmanaged spend sitting in plain sight.

Run cost is the money a live agent burns on every call: the inference tokens the language model reads and writes, the telephony minutes that carry the audio, and the orchestration around both. It drifts upward quietly. Prompts grow as teams patch edge cases, context windows fill with retrieved knowledge, defaults route every turn through the largest model, and dead air on the line is billed the same as speech. None of that shows up in the original business case, which is why the per-call number a programme launches with is almost never the number it settles at.

This is the post-production playbook for pulling that cost back down. It is not the human-versus-AI comparison that decides whether to deploy at all; our guide to AI voice cost per call owns that three-way business case. It is not the cost-structure breakdown in our voice AI total cost of ownership guide, nor the price-volatility hedging in our inference cost volatility playbook. This post owns one thing: the levers that reduce what a voice agent you already run costs to run.

This guide is shipped by the team behind Dilr Voice, enterprise voice AI built for regulated deployments. Or see DATS, our five-stage AI consulting system.

What is voice AI run cost optimisation?

Voice AI run cost optimisation is the post-launch discipline of reducing what a live voice agent costs to handle each call, without lowering answer quality. For Dilr Voice, it means right-sizing the model to each intent, trimming prompt and context tokens, caching stable responses, routing simple turns to cheaper paths, and cutting silent airtime, with every change regression-tested before it reaches production traffic. It is continuous FinOps applied to a conversation, not a one-off procurement exercise.

The discipline has a formal home. The FinOps Foundation, the industry body for cloud and AI financial management, added a dedicated scope for it. As their framework puts it:

"FinOps for AI focuses on addressing the cost complexity, faster development cycle, spend unpredictability, and the need for a greater degree of policy and governance to support innovation through allocation, forecasting, and optimization decisions that align consumption, investment, and business value." FinOps Foundation, FinOps for AI

The five levers below are how that intent becomes cost per call. They rank roughly by leverage, and they compound, but the order matters less than the guardrail: no cut ships until it clears the same quality gate.

Why does per-call run cost keep climbing after a voice agent goes live?

Per-call run cost climbs because a live voice agent is a moving target, not a fixed asset. Adoption grows, so volume rises; teams patch prompts to handle new edge cases, so every turn carries more tokens; knowledge bases expand, so more context is retrieved per call. Voice AI cost scales with usage and behaviour rather than with a provisioning decision, which is precisely why the FinOps Foundation treats spend unpredictability as the defining problem it exists to manage.

There is a second reason: nobody owns the number after launch. In most programmes the business case is written by finance, the platform is chosen by procurement, and the agent is then handed to an operations team measured on containment and customer satisfaction, not on cost per call. The lever sits idle because it is nobody's job. Standing up that ownership is part of a working AI operating model, and it is the difference between a cost that drifts and one that is actively managed. Internal recharge, covered in our voice AI chargeback guide, is one way to make a department feel the number it is generating.

Which lever cuts voice AI run cost the most?

Right-sizing the model to the intent is usually the single largest lever, because published token prices differ several-fold across a model family and most voice traffic is simple. Confirming a booking or answering an opening-hours question does not need a frontier reasoning model; a smaller, faster tier handles it at a fraction of the output-token price. Right-sizing sends each turn to the smallest model that answers it correctly, holding the large model for genuine reasoning.

The price gap is public and easy to verify. Across Anthropic's Claude family, published output-token list prices on 30 August 2026 run from $25 per million tokens for the largest tier down to $5 for the fastest, a five-fold difference that a right-sizing pass converts directly into cost per call. OpenAI and Google publish the same large-to-small tier structure across their model families, so the lever is not vendor-specific; it is a property of how frontier models are priced.

A note on where this lever stops. Right-sizing is about matching model to intent, not about chasing the lowest sticker price regardless of quality; a model that mishandles one turn in twenty and forces a callback is not cheap. That is why the gains here only count once they survive the regression gate described below, and why our proof of concept to production guide treats quality parity as a release condition, not an afterthought.

How do you cut token cost per call without hurting quality?

You cut token cost by attacking two multipliers on every turn: how many tokens you send, and how often you send the same ones. Trimming means shortening bloated system prompts, pruning few-shot examples the model no longer needs, and retrieving only the context a turn actually uses rather than the whole knowledge base. Caching means never paying full price twice for stable content. Together they shrink the token bill without changing what the caller hears.

Caching is the quieter win of the two. Model providers price a cache read at a steep discount to a fresh read: Anthropic, for example, charges prompt cache reads at 10% of the base input price, so a stable system prompt or a fixed knowledge block that is read on every call is charged at roughly a tenth once it is cached. For work that is not real-time, such as nightly knowledge refreshes, transcript summarisation or offline evaluation, batch processing runs at 50% of standard price. The rule of thumb: any content that repeats across calls and rarely changes is a caching candidate, and every token you can stop sending is a token you never pay to process.

When should a voice agent route to a cheaper model or a deterministic flow?

A turn should leave the expensive path the moment it does not need reasoning. Route to a cheaper model when the intent is simple and well-defined, and route to a deterministic flow, a scripted branch with no model call, when the turn is pure transaction: reading a balance, playing opening hours, capturing a reference number. The classifier that decides scope is cheap; the reasoning model it protects is not.

The discipline is to make routing an explicit design decision, not an accident of defaults. Map the intents your agent actually handles, mark which need reasoning and which do not, and wire the cheap paths first. This is the same routing logic that AI operating model consulting makes explicit, and it is where a lot of run cost hides, because an agent that routes everything through one large model is paying reasoning prices for clerical work.

How much of voice AI run cost is telephony and silent airtime?

Telephony is a real slice of voice AI run cost, billed by the minute whether the caller is speaking or not. Two things drive it: the carriage route the audio travels, and how many billed seconds each call actually consumes. Both are tunable after launch, and both are usually left at whatever the platform defaulted to on day one.

Carriage prices vary sharply by route. Twilio's published UK Programmable Voice rates on 30 August 2026 charge $0.0040 per minute over a SIP or BYOC trunk, against $0.0158 to a UK landline and $0.0305 to a UK mobile. The trunking route is several times cheaper than mobile termination, so how audio reaches your agent is itself a run-cost decision.

Silent airtime is the waste hiding underneath. Every second the line is open while the agent processes, waits on a slow tool call, or sits through a caller's pause is a billed second. Tightening endpointing so the agent responds faster, shortening synthesised responses, and resolving turns in fewer round trips all cut billed minutes as well as tokens. The lever here is latency: a faster agent is a cheaper agent, and a voice AI agent tuned for responsiveness bills less airtime per resolved call than a sluggish one, even at the identical per-minute rate.

How do you cut run cost without breaking the agent?

You cut safely by putting every proposed change through the same gate: a golden-set regression test that replays a fixed suite of real calls against the cheaper configuration and compares both quality and cost before anything ships. A right-sizing swap, a trimmed prompt, a new cache boundary or a routing rule is a hypothesis, not an improvement, until the golden set confirms the answer held.

Our golden-set regression testing guide covers how to build that suite; here it is simply the non-negotiable guardrail on every lever.

Changing one lever at a time is what makes the gate meaningful. Bundle four cuts into one release and a quality regression tells you nothing about which cut caused it. Ship them singly, measure each against the golden set, and you build an audit trail of what each lever actually bought, which is also what a mature AI execution office keeps on file when finance asks where the savings came from.

What is the best way to optimise voice AI run cost in 2026?

The best approach in 2026 is to treat run cost as a governed, continuous practice rather than a one-off tuning sprint, and to match the tooling to who owns the levers. An in-house team with FinOps discipline can pull most levers on a self-serve platform; a regulated enterprise that would rather buy the outcome than staff the practice is better served by a managed, pre-tuned one.

Self-serve platforms such as Vapi, Retell AI, Synthflow and Bland AI expose model choice, prompt configuration and caching, so the levers are yours to pull. Managed and pre-tuned platforms such as PolyAI and Dilr Voice own the optimisation on the customer's behalf, which suits regulated deployments that value the outcome over the operational overhead.

The honest concession is that a low-volume agent handling a single simple flow may not repay a formal optimisation programme; the effort can cost more than it saves, and the right move is to right-size the model once and move on. The threshold is call volume and intent variety. Above it, where an agent handles tens of thousands of calls across many intents, the levers compound and the practice pays for itself quickly. Finding that break-even is exactly the kind of question a scoping call answers, and it is a better first question than which platform to buy. To compare that build-versus-buy decision itself, our operating model guide weighs in-house against managed delivery.

The run cost optimisation operating cadence

A run-cost practice that survives contact with a busy operations team is light and repeating, not heroic and occasional. Set a monthly right-sizing pass that reviews which intents route to which model and whether the split still holds as traffic shifts. Run a weekly ten-minute review of cost per call alongside containment and customer satisfaction, so the number has an owner and a trend. Keep a standing backlog of candidate cuts, each waiting its turn through the golden-set gate. The cadence is deliberately dull, because dull is what compounds; the enterprises that win on run cost are the ones that pull one lever, prove it, and move to the next, month after month. That rhythm is the same operating discipline our approach brings to every deployment, and it is worth reading about Dilr.ai and the wider voice AI strategy library to see how it connects to the rest of a programme.

Does model right-sizing reduce answer quality?

Not when it is done properly. Right-sizing sends each turn to the smallest model that answers it correctly, which for most voice traffic is a smaller tier, and it holds the large model in reserve for genuine reasoning. Quality only drops if you right-size blind, without the golden-set regression test that confirms the cheaper model handled the same real calls to the same standard before the change reached any live caller.

How often should you re-run voice AI cost optimisation?

Treat it as continuous, with a monthly checkpoint. Traffic mix, prompt content and model pricing all move, so a split that was optimal at launch drifts within a quarter. A monthly right-sizing pass plus a weekly cost-per-call review keeps the practice current without turning it into a project. Dilr Voice runs this cadence for managed deployments so the enterprise sees the savings, not the workload behind them.

Want to see this in production? Try Dilr Voice live, book an AI placement diagnostic, see our DATS methodology, or read our voice AI cost of ownership guide.

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.