What an AI call centre actually costs in Saudi Arabia
A per-minute price is not a business case. How to model the real cost of an AI call centre in the Kingdom: the four line items vendors leave out, why containment is the denominator, and how to build a number your CFO will accept.
The first number every vendor gives you is a price per minute, and it is close to useless on its own. A minute is not the unit you care about. The unit you care about is a resolved call, and the gap between those two is where every AI call centre business case is won or lost.
This is how to build a cost model that survives a finance review, using numbers you can get from your own contact centre in an afternoon.
Two cost shapes, and why they behave differently
A human contact centre is priced per seat: a salary, a supervisor ratio, a workstation, licences, training, recruitment, and the cost of attrition. It steps up in discrete chunks. Going from 40 to 41 concurrent calls at peak may mean hiring five people, because you cannot hire a fifth of a person or roster them at three in the morning.
An AI call centre is priced per minute of conversation. It is close to linear and it does not care what time it is. That difference is the actual argument, and it is strongest exactly where seat-based cost is worst: overnight, weekends, Ramadan hours, and the volume spikes that currently produce a queue.
| Cost driver | Seat model | Minute model |
|---|---|---|
| Scaling to peak | Steps up in whole people; peak capacity is idle off-peak. | Linear; you pay for the peak minutes and nothing between them. |
| Out-of-hours cover | Shift premiums, or the calls go unanswered. | Identical unit cost at 03:00 and at 11:00. |
| Attrition and training | A standing cost that rises with headcount and churn. | Zero, replaced by a much smaller standing cost of call review. |
| Adding a new intent | Retrain everyone; the cost scales with headcount. | Change once; the cost is engineering hours, not headcount. |
| Quality drift | Managed by supervision you are already paying for. | A real ongoing cost that vendors routinely leave out. |
The four line items that are missing from the quote
A per-minute rate covers the conversation. It does not cover the four things that consume the budget in the first year, and a business case built without them will be wrong by a wide margin in the direction that gets people fired.
1. Integration into your systems
A voice agent that cannot write to your CRM, ticketing and scheduling systems is a very expensive answering machine. The integration work is the difference between a call being answered and a call being resolved, and it is where most of the engineering time goes. Budget for it explicitly, per system, and make sure someone has actually looked at the API surface of your ticketing tool before a number is quoted.
2. Dialect and vocabulary tuning against your own calls
Every contact centre has vocabulary the general model has never seen: product names, internal system names, site and building names, the particular way your customers describe a fault. Tuning against real recordings from your own lines is not optional, and the cost is proportional to how many distinct intents you are covering, not to call volume.
3. Escalation design
Deciding what the agent must never attempt, when it hands over, and what the human receives on transfer is a policy exercise involving your operations team, not a configuration screen. It is usually a week of somebody senior's time and it is the single highest-leverage week in the project.
4. Weekly call review, forever
Voice agents drift. Prices change, a product is renamed, an upstream model updates and a prompt that worked stops working. Somebody has to listen to real calls every week and act on what they hear. Whether that is your cost or the vendor's is the most consequential commercial question in the contract, and it is rarely on the quote at all.
Containment is the denominator
Here is the arithmetic that matters. Take your fully loaded cost per human-handled call. Take the vendor's cost per AI-handled minute and multiply by your average handle time. Then divide the AI figure by your containment rate, because every call the agent fails to contain costs you the AI minutes and the human call.
That division is brutal and it is why containment, not price per minute, is the number to negotiate on. An agent at half the per-minute price with 45% containment is more expensive than one at full price with 80% containment, and the cheaper quote is the one that will be in front of your CFO.
Which is also why containment has to be measured per intent rather than blended. Route the intents where containment is high and keep humans on the rest; the blended figure will look after itself. A vendor quoting one overall containment number across every call type is either not measuring it properly or is hoping you will not ask.
Building the number, in an afternoon
- Pull last month's call volume broken down by intent, and by hour of day. The hour-of-day split is where the out-of-hours argument lives.
- Compute fully loaded cost per human-handled call: total contact centre cost, including supervision, recruitment, attrition and premises, divided by calls handled. Not just salary.
- Identify the three highest-volume intents that are fully answerable from systems you already have. These, and only these, are your phase one.
- Estimate containment per intent conservatively, then halve it for the first month. You will be closer than the vendor's estimate.
- Add the four missing line items above as a first-year figure, then a smaller recurring figure for review.
- Compare against the counterfactual that is actually on the table: usually not “fire people”, but “stop hiring”, “stop abandoning calls at peak”, or “answer the phone overnight at all”.
That last point is worth dwelling on. The strongest business cases in the Kingdom are rarely headcount reduction. They are abandoned calls that currently become lost revenue, out-of-hours demand that currently goes unanswered, and peak-hour queues that a seat model cannot economically staff for.
Where Voho fits
We price per minute like everyone else, and we would rather you negotiated us on containment. Our published rate is on the pricing page and an account starts with credit on it, so you can measure your own numbers before a commercial conversation rather than after.
On the four missing line items, we take on the first three directly. Voho sends forward deployed engineers into the customer's contact centre to do the integration, the tuning against your real call recordings, and the escalation design, sitting with your operations team rather than working from a specification. That is a cost we carry into the deployment rather than an invoice that arrives in month three, and it is the main reason our first-year numbers look different from a self-serve platform's.
The fourth, weekly call review, is a shared cost and we would be suspicious of any vendor who claims otherwise, ourselves included. Somebody has to listen.
The rest of the position is straightforward: AI-native rather than a contact centre suite with a model bolted on, six Arabic dialects with what we believe is the best Saudi Arabic available, multilingual within a single sentence, and all data hosted in the Kingdom, or inside your own network where it cannot leave at all.
Frequently asked
- How much does an AI call centre cost in Saudi Arabia?
- Conversation is priced per minute and rates vary by vendor, but the per-minute rate is the smaller half of the real cost. Budget separately for integration into your CRM and ticketing systems, tuning against your own call recordings, escalation design with your operations team, and ongoing weekly call review. Then divide the running cost by your containment rate, because uncontained calls cost you both the AI minutes and the human call.
- Is an AI call centre cheaper than hiring agents?
- It depends almost entirely on containment and on where your volume sits. The economics are strongest out of hours, at peak, and on high-volume repetitive intents, because a seat model scales in whole people and cannot be rostered efficiently against those patterns. On complex low-volume intents, humans are usually still cheaper as well as better.
- What is a realistic containment rate?
- It varies so widely by intent that a single figure is not meaningful. Appointment booking, balance and status enquiries and simple service requests contain far better than billing disputes or anything requiring judgement. Insist on containment reported per intent, define contained as the outcome existing in your system of record, and count abandoned calls and 24-hour callbacks as failures.
- What is the real business case if we are not reducing headcount?
- For most Saudi enterprises it is not headcount at all. It is calls currently abandoned at peak that become lost revenue, demand outside working hours that currently goes unanswered, and the ability to stop hiring against volume growth rather than to cut existing staff. Those are easier numbers to defend and they do not require an argument about redundancies.
Keep reading