What your AI will really cost

The three cost items of an enterprise AI

The first item is licences. This is the cost of SaaS assistants — ChatGPT Enterprise, Claude Enterprise — billed per user per month. The public price lists stop at the mid-tier plans, around 25 to 30 dollars per user per month; the Enterprise plans are quote-based on both sides, on the order of several tens of dollars per seat, decreasing with volume and length of commitment. It is a predictable cost, proportional to the number of equipped staff, and the easiest to overestimate: equipping everyone from the outset is the most common budget mistake.

The second item is the API. This is the cost of AI integrated into your products and processes: customer chatbot, document extraction, classification, agents. Billed per million tokens — on the order of 1 to 5 dollars for input and 10 to 25 for output for the frontier models according to the public price lists, five to ten times less for the lightweight variants. A variable cost, unpredictable if no one models it, highly optimisable if someone takes charge of it.

The third item is infrastructure. It only appears if you host your own models: GPU servers bought and amortised or rented, hosting, energy, redundancy. Order of magnitude: from a few thousand to a few tens of thousands of dollars per month depending on the size of the platform. This item never fully replaces the first two — the hybrid remains the rule.

Licences or API: which for which use?

The dividing rule is simple, and it shapes the budget. Licences equip humans: a colleague who writes, summarises, translates, prepares a meeting. The cost is per seat, independent of usage intensity — an intensive user costs the same as an occasional one, which argues for not equipping the latter.

The API equips processes: an automated task that runs without human intervention, a feature built into your product, an agent that processes cases. The cost follows the volume — it grows with success, which is healthy, but it must be modelled before it is discovered.

The classic mistake is to cover a process use case with licences — thirty colleagues manually copy-pasting into an assistant what an API task would do for a few dollars a month. The mirror-image mistake is to build in the API what a licence would solve straight away, and to spend six months of engineering rebuilding an assistant that already exists. Ask the question this way round: is there a human in the loop at every execution? If yes, licence. If no, API.

On-premise infrastructure: when it enters the calculation

The calculator's infrastructure option covers hosting open models on your own GPUs. It is only justified in two cases: a data constraint that rules out the cloud, or massive, repetitive volumes where the marginal cost per task becomes decisive.

The figure the calculator displays is deliberately incomplete, and we say so rather than hide it: it covers amortised hardware, hosting, energy and hardware maintenance. It does not cover the initial engineering — two to four months of a team for model serving, RAG, guardrails and evaluation — nor day-to-day operations, which mobilise two to four people at cruising speed or an equivalent managed-services contract.

Add these items by hand in your scenario: a private AI platform never costs the price of its hardware alone, and the gap with the real first-year total is measured in multiples, not in percentages. Our private AI versus cloud AI comparison details this full calculation.

The hidden costs no one budgets for at the start

Four items systematically escape projections, and they weigh heavily.

Adoption. An unused licence costs exactly the same as a used one. Count 10 to 20% of the licence budget in the first year for training, support and driving usage — it is the best investment of the whole programme, and the first that leadership teams cut.

Prompt and evaluation engineering. Once the assistant is deployed, someone has to build the prompt templates, measure quality, correct drift. A fraction of a role, never zero.

Governance. Usage policy, data classification, CNDP formalities under Law 09-08, review of connectors, logging. A modest compliance cost, but one that becomes very expensive if handled after the audit rather than before.

API volume drift. A feature that works sees its usage explode — that is good news, provided you have set caps, a cache and routing to lighter models for simple tasks. Without that, the bill triples without anyone having decided it.

How to read the calculator's results

Three words of caution. The first: these amounts are orders of magnitude, not quotes. Enterprise prices are not published; the ones we offer by default reflect what the market reports, and your negotiation will move them — volume and length of commitment are the two levers that matter.

The second: simulate at least three scenarios. A minimal scenario — the population whose daily usage is proven. A target scenario — the eighteen-month trajectory. A drift scenario — API volumes multiplied by three. The gap between the three will tell you whether your budget holds at scale, which a single figure never will.

The third: the dirham. The calculator converts at the rate you set, with 10 dirhams to 1 dollar by default. Your vendor invoices will be denominated in dollars or euros: exchange-rate risk is real on a three-year commitment, and it deserves a line in your budget. Once you have your quotes in hand, replace every default value — that is when the calculator becomes useful.