Start with the cost of the workflow, not the model
The model API is often one of the smaller costs in a serious automation project. The expensive parts are discovering the edge cases, integrating source systems, handling failure, building review paths and operating the workflow after launch.
A useful budget discussion starts with how many hours, delays, errors or missed opportunities the current process creates. If the workflow is not economically meaningful, it is usually not a good automation target.
Four cost buckets
Most projects can be understood through four buckets.
- Process discovery and solution design
- Integration and application engineering
- Model usage, evaluation and guardrails
- Production monitoring, maintenance and support
Why a fixed-scope sprint can be better first
For a first engagement, a narrow production workflow is usually more useful than a giant AI transformation roadmap. Pick one measurable process, integrate it end to end, observe it in production and use the evidence to decide what deserves automation next.
AM12 typically starts with a workflow teardown and a bounded implementation rather than asking a business to commit to a broad transformation programme.
The question to ask a vendor
Ask what happens when the model is wrong, an API is down, the input document is weird, or the human disagrees. If the answer is only about prompts or model choice, you are still looking at a demo rather than an operational system.