An AI automation budget needs two views: the cost to make one workflow usable and the cost to run it safely each month. Start with a specific unit of work, such as preparing a support reply for approval. Count successful, reviewed cases rather than model requests. One case may require retrieval, several model calls, a retry and a person who checks the result.
Separate setup from monthly operation
Ask each supplier to price the same boundary. Discovery covers the current process, data access and acceptance criteria. Implementation covers the integration, interface and approval path. Evaluation covers a representative test set and failure analysis. Launch covers monitoring, documentation and an agreed support arrangement. If one proposal excludes these items, its headline total is not directly comparable.
| Setup item | Hypothetical assumption | Illustrative cost |
|---|---|---|
| Discovery and workflow mapping | 20 hours × €80 | €1,600 |
| Integration and implementation | 60 hours × €80 | €4,800 |
| Evaluation and launch preparation | 20 hours × €80 | €1,600 |
| Setup subtotal | 100 hours × €80 | €8,000 |
| Planning reserve | 20% of €8,000 | €1,600 |
| Initial budget envelope | €8,000 + €1,600 | €9,600 |
Calculate the full cost per case
Assume 1,000 completed cases each month. Budget €0.12 per case for model and retrieval usage, including an assumed allowance for retries: 1,000 × €0.12 = €120. Add €180 for infrastructure and monitoring, plus four maintenance hours at €80 = €320. These technical operating costs total €620. Usage charges must ultimately come from the selected services and measured workload.
Now include approval. At two minutes per case and an assumed internal labor cost of €30 per hour, review costs 1,000 × 2 ÷ 60 × €30 = €1,000. Total monthly operation is €1,620, or €1.62 per completed case. Record exception handling separately if it exceeds the assumed review time. A low model bill can coexist with expensive manual checking.
Compare against a measured baseline
Suppose the same 1,000 cases currently take eight minutes each at €30 per hour. That baseline is €4,000 per month. The illustrative difference is €4,000 − €1,620 = €2,380 of monthly capacity value. Dividing the €9,600 budget envelope by €2,380 gives approximately 4.0 months of simple payback, assuming the volume, quality and time savings hold after launch.
Capacity value is not automatically cash saved. People may use the released time for other work, and a pilot may reveal cases that still need full manual handling. Include onboarding, parallel operation and ramp-up in a purchasing decision. Do not justify a contract with the simple payback calculation alone.
Stress-test the assumptions before signing
- If review rises to four minutes, monthly review becomes €2,000. Operation becomes €2,620, and the illustrative difference falls to €1,380.
- At 500 cases, assuming the same €500 fixed technical cost and unchanged variable costs, operation is €1,060 against a €2,000 baseline. The difference is €940.
- If source access requires extra work, revise setup before approval. A connector label does not prove that your records, permissions and history are accessible.
Ask for normal and adverse scenarios with a spending limit and a stop condition. Decide who receives cost alerts and who can pause the workflow. Record the cost of failed or abandoned cases too; excluding them makes a successful-case figure look artificially low.
Buy an evaluation before scaling
A useful pilot ends with a scored evaluation set, observed handling times, a cost report and an explicit decision. Agree acceptable errors and the review boundary before testing. Include missing documents, contradictory records and provider outages. Keep the manual route available when the system cannot support its proposal.
For a scoped starting point, see AI automation. If the main uncertainty is access to existing systems, start with AI integration and the integration planning guide. Bring your monthly volume, current handling time, example cases and intended approver; these inputs make the next estimate meaningfully narrower.