“How much will it cost?” is the first question on every call, and “it depends” is an annoying answer. So here is what it depends on, and how to reach a useful number.
The four cost drivers
Scope of autonomy. A tool that suggests is far cheaper than one that acts. Each step toward autonomy adds guardrails, approvals and testing.
Data condition. Clean, accessible data can be used right away. Scattered PDFs, undocumented databases and access restrictions add weeks of pipeline work.
Integrations. Every system the AI reads from or writes to adds effort. Modern APIs are quick. Legacy systems with no API can become the largest line in the budget.
Risk level. An internal drafting tool can tolerate mistakes. A customer-facing tool in a regulated industry needs far more evaluation, logging and review.
Typical project shapes
These bands are only a starting point, and real numbers depend on your scope:
| Project shape | Typical duration | What it includes |
|---|---|---|
| Strategy sprint and prototype | 2 weeks | Use-case ranking, working prototype |
| Focused internal assistant | 6 to 10 weeks | One team, one data source family |
| Workflow agent with integrations | 8 to 14 weeks | Several systems, approvals, monitoring |
| Customer-facing AI product | 3 to 6 months | Full product, safety, scale |
For a figure based on your own inputs, use the AI Project Cost Estimator.
The costs people forget
- Model usage. Tokens are cheap per call and surprising at scale. Model it early with the LLM Cost Calculator.
- Evaluation. Someone must write and maintain test cases.
- Subject-matter time. Your experts need to review outputs, especially early on.
- Change management. Training and process change decide whether people use it.
- Run costs. Monitoring, updates and model migrations go on for as long as the system lives.
How to avoid overpaying
Start with a fixed-scope sprint rather than a big commitment. Ask vendors to show their evaluation approach, not just a demo. Insist on a phased plan with an exit after each stage. A vendor who cannot name what could go wrong has not thought about your project.
How to get approval
Finance does not need certainty. They need honest assumptions and a conservative case that still works. Our AI Business Case Template gives you the one-page format.