Short answer: Budget the complete experiment: discovery, data preparation, integration, model usage, evaluation, human review, security, internal time, and a controlled shutdown.
Budget the experiment, not only the model
An AI API can be inexpensive for a small test, but the API does not define the workflow, prepare data, build an integration, evaluate quality, investigate failures, or train the responsible team. A useful budget separates one-time learning work from recurring operating cost.
One-time pilot cost drivers
- Discovery and scope. Define the task, baseline, evaluation set, risks, acceptance thresholds, and decision process.
- Data preparation. Collect representative examples, remove prohibited information, correct obvious quality problems, and document permitted use.
- Prototype and integration. Build the smallest controlled route from input to reviewed output without granting unnecessary production access.
- Evaluation. Run normal, difficult, incomplete, multilingual, and adversarial cases; record results and human corrections.
- Security and privacy work. Review access, retention, provider terms, logs, secrets, incident handling, and any required assessment or agreement.
- Training and handover. Teach reviewers how to identify errors, escalate problems, and stop the pilot.
Recurring operating costs
- model, API, search, database, hosting, monitoring, and storage usage;
- human review and exception handling;
- prompt, evaluation, integration, and documentation maintenance;
- security review and access administration;
- model or provider changes;
- support, incident response, and periodic re-testing.
Create three scenario budgets
Use the same worksheet for a low, expected, and high case. Vary transaction volume, input/output size, retry rate, retrieval calls, human-review time, and exception rate. Keep provider prices as dated inputs linked to the provider’s official pricing page; do not turn them into permanent claims in a proposal.
| Budget line | Low case | Expected case | High case |
|---|---|---|---|
| Discovery and evaluation preparation | Defined hours | Defined hours | Defined hours |
| Build and integration | Smallest path | Expected path | Additional exception work |
| Provider usage | Dated calculation | Dated calculation | Dated calculation plus retries |
| Human review | Best observed rate | Median rate | Difficult-case rate |
| Contingency | Explicit amount | Explicit amount | Explicit amount |
The table should contain the client’s actual assumptions. Generic global “AI pilot prices” are not a substitute for scope.
Compare quotes fairly
Ask whether the quote includes the evaluation set, source and integration code, provider configuration, security work, usage charges, project management, documentation, training, warranty, and production transition. Confirm what happens if the pilot stops and whether the client can export its data and work products.
Set a learning limit
Approve a maximum amount the business is willing to spend to answer the pilot question. Release money in stages: discovery, controlled build, evaluation, and decision. Stop when evidence shows the expected value cannot justify the risk or ongoing operating cost.
A good pilot budget buys a reliable decision. It does not guarantee that the answer will be to proceed.