Start with a unit you can count
Choose one process and describe a completed run: one enquiry classified, product record checked, or support reply drafted. Record volume, staff time, connected systems, and the extra work a mistake creates. A broad goal such as 'automate support' is too vague to estimate.
Separate setup from operation. Setup may include process mapping, data cleanup, integration, access controls, testing, training, and rollout. Ongoing work may include subscriptions, model usage, storage, monitoring, human review, exceptions, and integration changes. These are planning categories, not a quote; actual charges depend on your services, usage, and terms.
Providers publish pricing rules that can change. Use the current pricing page for the selected service and record the date and assumptions. Do not reuse an old proposal or multiply a headline rate without checking what the workflow sends and receives.
Count the calls and the human work
For each run, list model tasks, input and output size, and attempts. Include classification, extraction, drafting, summaries, tool calls, and retries. An agent may make several calls per task, so estimate the path you will actually test rather than one request in isolation.
Count non-model work too: reading and updating records, notifications, platform limits, subscriptions, implementation, and operations. Human review takes time; estimate it by category if only some cases need approval.
Create ordinary, busy, and exception-heavy scenarios without treating them as guarantees. Label what changes, such as message volume, document length, retries, or handoff share. Update the estimate when real usage arrives.
- Count completed tasks and estimate monthly volume from recent records.
- List model calls, connected tools, retries, and average input and output size per task.
- Add platform fees, implementation, support, staff review, and maintenance separately.
- Write down the source and date for every vendor price assumption.
Use a synthetic example to make the model concrete
A fictional shop wants to sort delivery questions. Today a teammate reads each message, looks up the order, checks status, and drafts a reply. A proposed workflow classifies the message, retrieves status, drafts approved wording, and waits for review. Count model requests, the lookup, platform fees, review time, and failed-lookup monitoring.
Also count manual cases: unclear order numbers, address changes, suspected fraud, disputes, cancellations, and refunds. Routing those to a person is a control with an operating cost. If drafts need frequent correction, include that time and investigate the source or prompt.
This is not measured savings or a quote. Replace assumptions with the shop's volume, provider terms, and observed review time. A pilot can show whether the workflow fits and where the estimate is wrong.
Compare the estimate with a baseline
Record today's cost in the same unit as the planned system. Include staff time, corrections, handoffs, and measurable delays. Compare with software, usage, review, support, and exceptions. Total cost may stay similar even if repetitive steps fall; consistency or staff capacity may still matter.
A faster draft is not a completed task. Define 'done' as a sent message, updated record, or decision made. Recheck the baseline after a pilot to compare actual volume, effort, and error patterns.
Ask a provider how usage is measured, how overages work, how data is handled, and what happens if the service changes. For a custom build, assign support, access, testing, and maintenance. UK procurement guidance also highlights requirements, support, and supplier dependency; the questions help private teams too.
Make the estimate a living worksheet
Set a review date before expanding a pilot. Track runs, usage, review time, handoffs, corrections, and failures, while following privacy and retention rules. Update assumptions when the process changes.
A clear estimate helps a team proceed, simplify, buy, build an integration, or keep the manual process. Publish assumptions so another person can reproduce and revise the calculation when usage or vendor terms change.
- Set a spending limit and an owner who reviews usage during the pilot.
- Compare actual monthly usage and staff review time with the estimate.
- Recalculate when volume, model, provider terms, or workflow changes.