Somewhere in the last two years, a great many organisations signed off their first serious AI investment. Licences for the whole workforce, or a large slice of it. A number with several zeros in it, approved on the understanding that this was the year the organisation got AI.
Then, some months later, a team comes back with a proposal for something customer-facing — a self-service assistant on the website, a guide that helps people complete an application without phoning in. And there is an AI cost in the proposal. A recurring one.
The reaction is entirely fair, and I have heard versions of it in more than one room: isn't it all the same AI? Why are we paying twice? Does this ever stop costing money?
The word is doing too much work
The confusion is not a failure of intelligence. It is a failure of vocabulary.
"AI" arrives on the budget as a single line item, in the same way "cloud" did a decade ago and "digital" did before that. One word, one line, one signature. And the mind quite reasonably fills in the rest: we bought the thing, we have the thing, we should now be able to use the thing.
But underneath that one word sit at least two entirely different purchases, aimed at two different populations, sold on two different commercial models. Buying one of them does not give you the other, in the same way that buying every employee a car does not give you a delivery fleet. Same technology, different job, different procurement, different bill.
The first is AI for your people: a packaged assistant your staff use inside tools they already have open. Drafting, summarising, searching, catching up on a meeting they missed. It is sold as a licence, per person, per month.
The second is AI in your products: intelligence built into a service you own, aimed at someone who does not work for you. A customer, a citizen, an applicant. It is not a product you license. It is a capability you build, and you pay for it by how much work it does.
Why the first one cannot become the second
The instinct at this point is to ask why not. If we already pay for an assistant that can answer questions about our documents, why can it not answer a customer's question about the same documents?
Because a per-user licence is scoped to a user. That is not a technical limitation someone invented to sell you a second product — it is the entire commercial logic of the thing. The licence attaches to a named, signed-in member of your staff. It lives inside their applications. It works on content they already have permission to see.
A member of the public has none of that. No licence, no account, no identity in your tenant, no permission boundary already worked out. Serving them is a different activity: your own interface, your own rules about what may be answered, your own audit trail, your own decisions about what happens when the answer is wrong. None of that is inside a productivity licence, and no negotiation with the vendor will put it there.
What the second bill is actually for
Once the second purchase is understood as building a service rather than buying AI again, the recurring cost stops looking like a scam and starts looking like what it is.
You are not buying a licence to a finished product. You are commissioning a capability, on infrastructure you rent, that will do a variable amount of work every month for people you do not employ. The bill is recurring because the work is recurring. Every question a customer asks costs something to answer, in exactly the way every kilowatt-hour of light in a building costs something.
That is uncomfortable for a budgeting culture built on capital purchases — you buy the machine, you depreciate the machine, you are done. But it is not new or exotic. It is how every utility and every cloud service you already pay for behaves. AI simply arrived with enough hype attached that people expected it to behave like software from 2005.
Why the misunderstanding is expensive
If this were only a semantic problem it would not be worth eight articles. It is expensive in three specific ways, and I have seen all three.
Projects stall. A good idea for a customer service gets stopped at the budget gate, not because it is a bad investment but because the sponsor believes it is a duplicate of one they already approved. The organisation ends up with excellent internal drafting tools and nothing whatsoever facing the public.
Trust erodes. When the finance team believes IT is asking for the same thing twice, every subsequent technology request is read through that suspicion. That is a hard thing to repair and it usually gets repaired at the wrong time, in the middle of an unrelated negotiation.
The wrong tool gets forced into the job. Occasionally an organisation, determined not to pay twice, insists that the productivity licence be made to serve the public somehow. What follows is a workaround — staff manually relaying answers, a shared account nobody should be using, a process that quietly breaks the licensing terms and the audit trail at the same time. It costs more than the thing it avoided.
What this series does
The rest of this series takes the confusion apart, one piece at a time, in plain language and without asking you to become technical.
We start with the distinction that does most of the work — AI for your people versus AI in your products. Then the two pricing models underneath them, per-seat and per-use, and what a "token" actually is when someone puts one in a spreadsheet. Then reserved capacity, and why an organisation ends up paying to hold a lane it is not driving on. Then the shift from capital thinking to running-cost thinking, the genuine cost of keeping AI processing inside South Africa, the small set of controls that keep the bill governed, and finally a one-page way to decide which AI to pay for, for which need.
None of it requires you to understand how a model works. All of it is about what you are buying and what it will do to your budget.
The one thing to take from this article
Before you push back on the second AI bill, check whether it is actually the same purchase as the first. Almost always, it is not.
That does not make it automatically worth funding. It might be badly scoped, prematurely sized, or aimed at a problem nobody has. Those are all good reasons to say no. "We already pay for AI" is not one of them — and an organisation that says no for that reason will keep saying no, right up until a competitor or a neighbouring department says yes.
How CloudNala can help
We spend a lot of our time in the room where this argument happens, and the useful contribution is rarely more technology. It is separating the two purchases clearly enough that the finance conversation can proceed on its merits: what each one is for, what each one will cost at realistic volumes, and which of them the organisation actually needs first. Frequently the answer is that one of the two should wait — but for a real reason, not a misunderstanding.
Work with CloudNala
CloudNala helps organisations move from technology ambition to practical execution across cloud, AI, data, platform engineering and digital services.
Whether you are exploring AI, modernising your cloud environment, building a public-sector digital service, or turning an idea into a working MVP, we can help you shape the roadmap and deliver the next step.
Book an AI Readiness Workshop or write to us at consult@cloudnala.co.za