The short answer
We do not publish prices for AI work, because "an AI assistant for our staff" can mean light configuration or a substantial build, depending on details nobody has looked at yet. What drives the cost is predictable, though:
- how clear the use case is and how many teams it touches,
- how ready your data is: current, organised and owned,
- how many systems the solution must read from or write to,
- how accurate it must be, and what a mistake would cost,
- how sensitive the information is, and
- how much it will be used, because model usage is metered.
Costs fall into two buckets: one-time work to assess, prove, build and launch, and ongoing costs to run, monitor and support. A proposal you can trust shows both, states its assumptions and lets you stop between phases.
Where the one-time cost goes
| Component | What it covers | Bigger when | Smaller when |
|---|---|---|---|
| Discovery and readiness | Use cases, data, systems, permissions, risks | Many teams, unclear goals, a regulated sector | One clear problem and a sponsor who decides |
| Proof of concept | Testing one use case on your data against agreed thresholds | Vague success measures, scattered data | Real examples ready, one data source |
| Build and integration | Connections to your ERP, Microsoft 365 or CRM, interface, permissions | Several systems, no usable APIs, writing back | Read-only access to one documented system |
| Data preparation | Removing outdated and duplicate content, structuring records | Conflicting versions, scanned paper, no content owners | Current sources with named owners |
| Evaluation and testing | Test sets, accuracy measurement, permission and security tests | High accuracy targets, outputs that affect people | A narrow task with easy-to-check results |
| Security and governance | Risk and data protection review (including a DPIA where required), logging, prompt injection testing, policy | Personal or sensitive data, data sent abroad, agents that take actions | Internal, non-sensitive content |
| Change management and training | Training, champions, updated procedures | Many users, a new way of working | A small team, a tool that fits current work |
The AI model barely appears. Most one-time effort sits in data, integration, testing and controls, which is where the cheapest quotes cut corners.
Ongoing costs, and how they scale with volume
Model usage. Most AI services meter usage by the token (a small chunk of text) or by the request, page or document processed. Microsoft notes that the billing unit and rate can vary by model, deployment type and meter (Microsoft Learn). Anthropic, for example, publishes separate per-million-token rates for input and output, with a discount for batch work that can wait (Anthropic). Rates change often, so check the vendor's current pricing.
Think of it as cost per transaction multiplied by transactions per month. Volume sets the second number. Design sets the first: how much document text is sent with each question, whether an agent makes several model requests per task, and whether a smaller, cheaper model handles simple steps such as classification. Two designs for the same use case can differ widely in cost per transaction, so measure it.
Uncontrolled usage is also a security issue: OWASP's Top 10 for LLM applications includes "unbounded consumption", where excessive or malicious use runs up the bill (OWASP). Spending limits and alerts belong in every production design.
Cloud hosting. Search indexes, databases, storage, application hosting and logs. Microsoft points out that AI service charges are only part of total solution costs, and that a deployed fine-tuned model incurs hosting charges even when unused.
Licences. Per-user licences scale with people, not volume. Microsoft 365 Copilot is generally sold as an add-on licence for users on an eligible Microsoft 365 plan, while web-grounded Copilot Chat is included with eligible subscriptions (Microsoft Learn). Who gets a licence first is a budget decision, which is why our Copilot service starts with a pilot.
Exchange rates. Most AI platform, model and cloud subscriptions are priced in US dollars, so exchange-rate movements change what they cost you to run in naira from month to month. Budget with that in mind and review running costs regularly.
Monitoring and support. Sampling accuracy, updating sources, re-testing when a vendor retires a model, handling incidents and helping users. Smaller than the build, never zero.
Self-hosted models. Running an open-weight model on your own GPU servers changes the shape of the bill: you pay for the servers whether they are busy or idle, plus the effort to patch, monitor and update them, instead of paying per token. That pays off at steady, high volumes or when data must not leave your environment. Our comparison of self-hosted models and AI APIs shows how to work it out.
Budgeting a pilot versus production
| Proof of concept | Production | |
|---|---|---|
| Scope | One use case, a small user group | All intended users and sources |
| Data | An approved sample, de-identified where possible | Live data with permissions enforced |
| Integration | Minimal or read-only | Supported connections, approvals for write-back |
| Usage cost | Small, but measured per transaction | Measured cost times realistic volume, with headroom for peaks |
| Result | Results against thresholds, go or no-go | A supported service with a runbook |
Budget the proof of concept as its own decision, and production as a second one informed by what it measured. Microsoft's cost guidance takes the same approach: run a small test workload with representative traffic and compare actual charges with your estimate before rollout. Our guide to running an AI proof of concept covers the thresholds.
Hidden costs to plan for
- Data clean-up. Three versions of the same procedure produce confident, wrong answers.
- Permissions. If files are overshared, AI makes them easy to find. Fixing access is often the first real cost of a Copilot rollout; see our Copilot readiness guide.
- Integration. API access can depend on your subscription tier, and older systems may need work first.
- People's time. Experts write test questions and judge answers, an owner decides, pilot users take part. It never appears on an invoice, so it is easy to forget.
- Re-testing and exit. New models and documents mean re-running tests. Your prompts, test sets and integrations should move with you if you change provider: put ownership in the contract.
Rather talk it through? If you need a realistic AI budget for leadership, we can scope your use case and give you phased estimates in naira (or US dollars where a licence or service is priced in dollars), with one-time and ongoing costs separated. Talk to a Promatics specialist
How Promatics structures an AI proposal
- A written scope. The use case, users, data sources, systems, success measures, assumptions and exclusions.
- Phases with decision points. An assessment, then a proof of concept with go or no-go thresholds agreed before it starts, then the build. Each phase is estimated separately and you can stop at any gate. The assessment is yours to use with any provider.
- One-time and ongoing costs separated. Build effort on one side; model usage, hosting, licences and support on the other, with usage estimated from measured results and your volumes.
- Estimates in naira (or US dollars where a licence or service is priced in dollars). Vendor subscriptions and usage charges are listed separately, because you pay them to the vendor, and dollar-priced items are marked as such.
- Support on clear terms. Monitoring, re-testing and fixes under a service agreement, with response targets set in writing. Managed-service clients get 24/7 monitoring and support.
When to bring in help
You do not need a consultant to budget a trial of Copilot Chat or an AI feature in software you already pay for: read the licence terms, check current pricing and set a policy on what data people may use.
Bring in help when AI will touch personal data, financial records or decisions about people; when data would be sent to an AI service abroad; when it must connect to your ERP, CRM or Microsoft 365 data; when a board needs a defensible budget; or when the quotes you have cannot be compared. Our AI strategy and readiness service gives you a scored shortlist of use cases with cost ranges that separate build from running costs. To compare providers, see choosing an AI partner.
Before you approve an AI budget
- The use case, its users, its success measure and its monthly and peak volumes are written down.
- Each data source has an owner, and its quality is known.
- Systems to connect are listed, read-only or write-back.
- The accuracy threshold, and who judges it, are agreed.
- Personal and sensitive data is identified, with a data protection review (and a DPIA where required) planned.
- One-time and ongoing costs are shown separately, with dollar-priced items marked.
- Cost per transaction comes from a measured test, with spending alerts planned.
- Internal time from experts, owners and pilot users is budgeted.
- Support, monitoring and re-testing after launch are priced.
Sources and further reading
Product capabilities and guidance change. These are the primary sources this article relies on, checked on the review date above.
- Plan and manage costs for Microsoft Foundry, Microsoft Learn
- License options for Microsoft Copilot, Microsoft Learn
- Pricing, Anthropic (Claude Platform Docs)
- LLM10:2025 Unbounded Consumption, OWASP Gen AI Security Project
This article is general information, not legal, accounting or security advice for your specific situation. Examples are hypothetical unless stated otherwise.