- 01
Market
Saudi Arabia
- 02
Written for
Head of CX, CIO
- 03
Decision
Budget an AI voice agent across its build and its per-call operation
- 04
Languages
EN · AR
Anatomy of a call
The path a single call takes through an AI voice agent
Every stage below is a cost surface: some are built once, some are paid on every call. Seeing the path is what makes the two budget lines legible.
A call flows through four stages — arrival over telephony, understanding of speech and dialect, action in core systems, and closure — with a branch at every stage for handover to a person with context.
- 01
The call arrives
Telephony integration carries the caller from your numbers and phone platform into the agent — built once, metered per minute thereafter.
- 02
The caller is understood
Speech recognition and language handling interpret what was actually said — including the dialects your callers really speak, not the Arabic of the brochure.
- 03
The agent acts
It answers from approved content, or goes further and performs a transaction — a booking, a balance, an order status — inside your core systems, under authenticated, audited access.
- 04
The call closes
The outcome is recorded, the transcript kept, and the conversation feeds the tuning that keeps the agent improving against real usage.
Handover to a person
At any stage, the call can transfer to your team with its context attached — who is calling, what they asked, what the agent already did. Designing this branch well is a build cost; skipping it is a customer cost.
Two budget lines
What sets a voice agent's build cost and what sets its cost per call
Read each driver with one question in mind: is this paid once, or on every call? The distinction is the whole discipline of budgeting voice AI.
| Variable | Effect on price | Why |
|---|---|---|
| Call volume | The operating multiplier | Volume barely moves the build and entirely drives the running bill, because model usage and telephony are metered per call and per minute. Every design decision that shortens calls or resolves them earlier is therefore multiplied by your volume — which is why efficiency work that looks cosmetic in a demo is real money at scale. |
| Languages and dialect coverage | Arabic dialect handling is real work | An agent that understands the dialects your callers actually speak — not just modern standard Arabic — has to be tuned and tested against recordings of real calls, and every added language or dialect is scoped work with its own test set. A demo that performed beautifully in English or careful fusha has not yet met your callers. |
| Integration depth | Answering is cheap, acting is not | An agent that answers questions from approved content is the light end of the range. One that acts — books, cancels, checks a balance, changes an order — needs authenticated access to core systems, error handling for every way those systems fail, and an audit trail for every action. Each transactional capability is its own scoped piece of build. |
| Telephony integration | The line into your phone system | Carrying real calls means connecting to your numbers, your PBX or contact-centre platform, with transfer, recording and compliance handled properly. A voice agent demonstrated on a web widget has not priced any of that — and the telephony minutes themselves recur on the operating side for every call the agent takes. |
| Human-handover design | The escape hatch is a feature, not a failure | Deciding when the agent hands a call to a person, and delivering that person the context so the caller never repeats themselves, is designed and built work — and it is what separates an agent customers tolerate from one they hang up on. It also shapes staffing: the calls that reach your team after a good agent are the hard ones. |
| Ongoing model and telephony usage | An operating expense, not part of the build | Speech recognition, the language model and the telephony are metered services billed by use, so they belong in the operating budget, scale with volume, and continue for the life of the agent. A proposal that presents one merged figure has either padded the build or hidden the running bill — insist on seeing the two lines apart. |
Ask any supplier quoting this
- Which Arabic dialects has this agent been tested against, and with recordings from where?
- Is the agent answering from content only, or performing actions in our systems — and which actions?
- How does your figure split between building the agent and running it per call?
- What does the handover to a person look like, and what context travels with the call?
- How is telephony connected — to our numbers and platform, or to yours?
- How does the running cost move if call volume doubles during a seasonal peak?
The merged number
Why a voice agent needs two budget lines, not one
The build and the operation of a voice agent answer to different logic. The build is a project: it is scoped by dialects, integrations, telephony and handover design, it ends, and its cost is judged like any delivery. The operation is a rate: model and telephony usage metered on every call, continuing for the agent's life, scaling with your volume and moving with your call lengths. Merging them into one figure invites two mistakes at once — comparing suppliers on totals that assumed different volumes, and approving a project budget that quietly contains an open-ended subscription.
Kept apart, each line also becomes governable. The build line is controlled the way any scope is: fewer transactional actions at first, one language done properly before the next. The per-call line is controlled by design: resolving calls earlier, containing the ones the agent should never have taken, and reviewing the metered bills the way you would review any utility. The organisations that regret voice AI are rarely the ones that spent most — they are the ones that approved one number and only later learned which part of it repeats every month.
Our role
What we listen to before quoting an AI voice agent
- 01
Your real calls first
We scope from recordings of the calls the agent will actually take — their dialects, their intents, their lengths — because an agent designed from a call-centre report meets a politer caller than the one who phones.
- 02
An action list, agreed in writing
Which transactions the agent may perform in which systems is decided before the build is sized, because integration depth is the largest build variable and the easiest one to leave vague.
- 03
Two lines, both visible
Our proposal states the build figure and the per-call operating estimate separately, with the volume assumptions behind the second written down — so the recurring cost is a decision you took, not a bill you met.
Which calls to automate first, and how much of the phone experience to entrust to an agent at all, are service decisions that stay yours; we make sure both costs of the answer are visible before you make them.
Common questions
What buyers ask
What do you need from us before you can put a number on a voice agent?+
Three things, and they explain why no figure appears on this page. Recordings or samples of the calls the agent will take, because dialects, intents and call lengths set the language work. The list of actions you want it to perform in your systems, because integration depth is the largest build variable. And your call volumes, because they drive the per-call operating line. With those, a scoped build figure and an operating estimate follow quickly — without them, any number we gave you would be a guess about someone else's callers, and you would budget against it.
Why does Arabic dialect coverage change what a voice agent costs?+
Because the Arabic your callers speak is not the Arabic most systems were trained to expect. Callers greet, interrupt, code-switch into English mid-sentence and use regional vocabulary that varies within the Gulf itself — and an agent has to understand all of it reliably before it can do anything else. Getting there is concrete work: tuning against recordings of your real calls, building test sets per dialect, and measuring recognition honestly rather than assuming it. Each dialect done properly is scoped effort, which is why an agent for Gulf callers costs more to build than the same agent for one careful register — and why a demo in English or formal Arabic proves very little.
Is an AI voice agent a one-off project cost or an ongoing bill?+
Both, permanently, and treating it as either alone is the classic budgeting mistake. The build — language tuning, integrations, telephony, handover design — is a project that ends. The operation — model usage, speech processing and telephony minutes on every call, plus the tuning that keeps the agent accurate as your products and policies change — runs for as long as the agent answers the phone, and scales with volume. Budget it like a hire rather than a purchase: a cost to bring the agent in, and a running cost for the work it does every month thereafter.
In this cluster
A scoping session over recordings of your real calls that ends with the agent's action list, its handover design and both cost lines — build and per-call — stated separately.
Scope your agent with us
