FIELD NOTE
A strong proposal begins with movement, not models
The supplier should trace how work, information, and decisions move today. A proposal written before understanding users, exceptions, volumes, systems, and ownership is likely to optimise an imagined process.
Ask for a concise current-state map and a definition of the result that will be measured after deployment.
FIELD NOTE
Inspect the whole delivery system
A model is one component. The working product also needs interfaces, permissions, integration, data preparation, testing, monitoring, support, and change control.
- Named business and technical owners
- Baseline and target operating measures
- Data-flow and architecture diagram
- Security and privacy responsibilities
- Acceptance tests and failure scenarios
- Documentation, training, and support response
- Data export and supplier exit plan
FIELD NOTE
Separate evidence from theatre
Do not treat a smooth prototype as proof of production readiness. Request test cases using representative data, including missing fields, unusual orders, duplicate records, unavailable services, incorrect model output, and a staff member without permission.
Where approved client evidence is unavailable, the honest answer is a clearly labelled demonstration, not an invented case study.
FIELD NOTE
Price the life of the system
Compare discovery, implementation, usage charges, licences, hosting, monitoring, maintenance, provider changes, training, and internal operating time. The cheapest build can become expensive when no one owns exceptions or understands how to modify it.
FIELD NOTE
Choose for the second year
The right partner helps the organisation own a clearer process and better data after launch. Require a roadmap for measurement, iteration, provider changes, and retirement. An operating system is valuable because it can adapt without becoming opaque.
DIRECT ANSWERS
Questions operators ask
What should an AI automation proposal include?+
It should include the operating problem, users, current baseline, target measure, workflow boundary, data flow, architecture, controls, test plan, exception ownership, delivery stages, support, costs, and exit arrangements.
Should a business choose a product or custom AI system?+
Use a product when the process is standard and the product fits without harmful workarounds. Use custom integration or software when the differentiating workflow, data boundary, or system connections require it.
How can a buyer avoid AI vendor lock-in?+
Keep source records exportable, document integrations and prompts, separate business logic from one model where practical, define ownership contractually, and test the exit path before it is needed.
SOURCE LEDGER
Primary sources
Official material is linked directly. Claims are paraphrased and checked against the source status available on 2026-07-21.FRICTION EDITORIAL CONTROL
Original analysis. Visible limitations. No invented certainty.Prepared by Friction Research and reviewed against primary sources. This material is general information, not legal, tax, financial, or regulatory advice. Requirements can change; verify material decisions with the relevant authority or a qualified adviser.Read our editorial policyREADER EXCHANGE
Add to the conversation.
Name and email are required. Suspicious or abusive comments are held before publication.
Loading conversation.