The cost of adding AI to a business is three bills, not one: a build you pay once, a run you pay every month for as long as the system lives, and a hidden bill that decides whether the first two were worth paying. This post opens up each bill, prices one realistic first project end to end, and shows where owners get quoted the wrong number.
Most vendors answer “how much does AI cost?” with it depends. Here is exactly what it depends on.
The three bills, and why you usually only see one
A quote covers the build. The invoice that arrives every month afterwards is the run. The hidden bill never appears on paper; it shows up as a stalled project, a tool nobody adopts, or a second build to fix the first.
flowchart TD A(["What adding AI really costs"]) --> B(["Build: paid once"]) A --> C(["Run: paid every month, for as long as it lives"]) A --> D(["Hidden: paid when nobody budgeted for it"]) B --> E(["Integration, data preparation, testing, rollout"]) C --> F(["Usage, hosting, upkeep, oversight"]) D --> G(["Bad data, change, compliance retrofit, doing it twice"])
If a proposal does not cover all three, the number on it is not the price.
Bill one: the build, paid once
In our experience the model itself is rarely the expensive part of the build; the work around it is.
- Integration. Connecting the AI to the systems you already run: the ERP (SAP, Oracle, Odoo, whichever it is), the CRM, WhatsApp Business. This is engineering, and usually the largest line.
- Data preparation. Getting your catalog, contracts, FAQs, or transaction history into a shape the model can use, and deciding what it must never see.
- Access and permissions. Who can ask it what, which records it can touch, how requests are logged. Cheap up front, expensive to bolt on later.
- Testing on your real cases. Running a few hundred of last month’s actual enquiries through it before a customer does.
- Rollout. Training the people who will use it, and naming an owner for after go-live.
Ask for the build quote broken into those lines, not one number for “implementation”.
Bill two: the run, paid every month for as long as it lives
AI is metered: you pay for how much it is used, so the cost grows with your volume, and it never stops.
- Usage. Every question answered and every document read is a call to a model, charged per call or per unit of text. More customers, more cost.
- Hosting. The servers and storage it runs on. If customer data has to stay in the Kingdom, that means an in-Kingdom cloud region, at that region’s pricing.
- Upkeep. Models get retired, integrations break when the system on the other side updates, prompts drift. Someone has to own it.
- Oversight. The human who reviews the exceptions the AI hands back. A real salary line, and the pitches that skip it are the ones whose savings never materialize.
The rule that matters most: a cheap build with an expensive run costs you more over a year than the other way around. Price both, and price the run at the volume you expect in month twelve, not month one. Our post on saving costs with AI works through the gross-versus-net subtraction; the run cost is why those two numbers differ.
Bill three: the hidden costs nobody quotes
- Bad data. If your information is scattered across spreadsheets, an old system, and people’s inboxes, cleaning it is part of the cost, and in our experience it is often the largest hidden line of all. Gartner names poor data quality among the reasons it predicts at least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025, alongside inadequate risk controls, escalating costs, and unclear business value (Gartner, 2024). Every one of those four is a cost that was not on the quote.
- Change. Software your team does not adopt costs the same as software they do, and returns nothing. Training and a manager who insists on the new way are real money, and usually left out.
- Compliance retrofit. Saudi Arabia’s Personal Data Protection Law (PDPL) took effect on 14 September 2023, full compliance was required by 14 September 2024, and SDAIA is the regulator (Morgan Lewis, 2024). If your assistant handles customer data and nobody checked where that data goes, you pay to re-architect after launch, which costs far more than asking before the build. We covered the mechanics in using your company’s own data with AI, safely.
- Doing it twice. The most expensive AI project is the one built wrong the first time, then paid for again to fix.
A worked example: pricing one first project end to end
Say you run a retail business and get 6,000 customer enquiries a month across WhatsApp and email: where is my order, can I return this, is it in stock. The numbers below are made up to show the shape, not to quote you.
- The build. Integrating an assistant with your order system and WhatsApp Business, preparing the FAQ and returns policy, a PDPL review of which customer data it may see, and two weeks of testing on real enquiries. Call it SAR 90,000, once.
- The run. Model usage at that volume, about SAR 2,500 a month. Hosting in an in-Kingdom region, SAR 1,500. An upkeep retainer, SAR 2,000. Total run: SAR 6,000 a month. Plus oversight: one agent spends around 30 hours a month on the enquiries the assistant hands back, call that SAR 1,500 of salary.
- The payoff. If the assistant handles 65% of enquiries end to end, that is 3,900 conversations a month that no longer take an agent five minutes each: roughly 325 hours, about two agents, say SAR 16,000 a month.
- The real number. SAR 16,000 saved, minus SAR 6,000 run, minus SAR 1,500 oversight, leaves SAR 8,500 a month. Against a SAR 90,000 build, the project pays back in about eleven months.
Now change one line. Suppose that instead of a build, you bought a platform that charges SAR 2 per conversation. The build drops to almost nothing, which looks wonderful. But the run is now SAR 12,000 a month, the net falls to SAR 2,500, and the payback stretches to years while the fee climbs with every customer you win. That is “cheap build, expensive run” in a real spreadsheet.
Where owners get quoted the wrong number
- The demo price. A quote for a proof of concept is not a quote for a production system; the pilot skips the integration, the permissions, and the testing. Our guide to running an AI proof of concept that leads somewhere covers how to keep the pilot’s number honest.
- Volume priced at month one. A run cost estimated on the pilot’s traffic understates the real bill. Price it at the volume you expect once every customer is on it.
- Paying for AI when a rule would do. If the task is “send every invoice over SAR 50,000 to the finance director,” that is a rule, not a model, and a rule costs nothing to run.
- The platform before the use case. Buying an enterprise AI suite and then looking for something to do with it puts the whole bill in front of any proof.
- No exit. If your data, prompts, and history live inside one vendor’s product, leaving means a rebuild. Ask up front what you can export.
How to start without overpaying
The boring pattern works. Pick one process that is high volume and repetitive, where a wrong answer gets caught by a person rather than a customer. Measure what it costs you today in hours, honestly. Get a quote split into build and run, with the run priced at real volume. Pilot that one thing, then measure the net saving, not the gross. If it pays back, you have your model, and the second project costs less because the integration and data work are already done. If not, you spent a little to learn something cheap.
Before you sign anything, put three questions to whoever is quoting you.
flowchart LR
Q1{"All-in cost to build it?"} -->|straight answer| Q2{"Monthly run cost at our real volume?"}
Q2 -->|straight answer| Q3{"What is it worth to us if it works?"}
Q3 -->|straight answer| Y(["Fund a small pilot"])
Q1 -->|vague| N(["Walk away from the vendor, not from AI"])
Q2 -->|vague| N
Q3 -->|vague| N
If you cannot get a straight answer to all three, that is a red flag about whoever you are talking to, not about AI. The right answer to “how much does AI cost” is a number for your business, with all three bills visible.
Want a real number for your situation, not “it depends”? SDCG’s AI consulting gives owners honest, independent estimates: build cost, run cost, and the payoff, for AI projects scoped to your business. We don’t resell software, so the numbers aren’t bent toward a product. Book a free 30-minute review.