AI implementation costs: budget for the whole workflow
Neaptide · September 7, 2026 · 6 min read
Data, integrations, evaluation and support belong in an AI budget. A fictional example compares the cost of an accepted answer, including human review.
On this page
To estimate AI implementation, choose a workflow with a clear input and result. “An employee assistant” is too broad. “Draft a response from current service terms, have an employee check it, then send it to the customer” is a process you can scope.
The budget covers process and data preparation, development, quality checks, launch and operation. Model usage is one line item. Human review, exceptions and knowledge updates also consume resources, even when they do not appear on the developer’s invoice.
What implementation should cover
- Workflow definition: inputs, expected outputs, exceptions and limits on autonomous actions.
- Data preparation: current documents, access, duplicate removal and update rules. Conflicting instructions need the process owner’s input.
- Integrations: receiving the request, retrieving information, delivering the draft, recording the outcome and handling failures.
- Evaluation: ordinary and difficult cases, expected answers, acceptance criteria and retests after changes.
- Handover: training, support owners, spending controls and a stop or rollback procedure.
A demonstration is not the whole workflow
A demo with a few straightforward questions can show that an idea is feasible. Day-to-day use also means handling outdated documents, ambiguous requests, restricted access, unavailable services and irrelevant search results. Ask the developer to include these situations in the estimate.
A narrow pilot tests assumptions before expansion. A valid outcome can be stopping: unreliable data or review time equal to manual work may undermine the case. Agree the stop conditions before testing.
A worked example: a cheap call can produce a costly answer
All figures are fictional, not model prices, Neaptide rates or customer measurements. Compare two AI-assisted drafting options. Assume both reach the same agreed quality after review; review time includes reading and corrections.
| Per accepted answer | Option A | Option B |
|---|---|---|
| Illustrative API cost for the entire workflow | €0.03 | €0.12 |
| Human review and corrections | 6 minutes | 2 minutes |
| Time valued at a hypothetical €24/hour | €2.40 | €0.80 |
| API plus review | €2.43 | €0.92 |
B has higher API cost but is €1.51 cheaper including review in this example. For 1,000 accepted answers, that is €920 versus €2,430. This depends on the assumed times and equal quality. Measure both in practice; an expensive model does not necessarily reduce review.
The comparison excludes development, hosting, storage, support and data preparation. Add those costs for the period you are assessing. Time saved does not automatically become cash saved: identify what the team could do with that time or which expense you could actually reduce.
Measure the pilot by accepted work
Count completed, accepted tasks rather than calls. Record system spend, review time, retries and rejection reasons. If 100 requests yield 70 useful results, do not spread the cost as though you delivered 100 finished answers.
Microsoft advises considering error impact and detectability when delegating to AI. Design human review into the workflow and measure its time.
Include operation after launch
Discuss model use, infrastructure, knowledge updates, failure analysis and regression checks. Changes to instructions, integrations or models may need retesting. Name the person who detects issues and the team with capacity to fix them.
Microsoft’s cost guidance covers both initial and recurring costs, and recommends comparing actual spending with your estimates. Use that approach to account for the work that continues after the demonstration.
Prepare a useful estimate request
Describe one workflow: how often the task occurs, who handles it, how long it takes and which errors would be unacceptable. Provide sample inputs with identifying details removed, examples of good outputs, the systems involved, relevant documents and a subject expert who can answer questions.
This supports a realistic Neaptide pilot discussion: what to test first, what implementation includes and when to expand. A broad price range without these details is less useful than a clearer scope.