AI automation procurement
How do you brief an AI automation project and compare proposals?
Compare proposals for one complete internal workflow, with the same inputs, permissions, acceptance cases and operating responsibilities. Use this brief to distinguish a useful pilot from a convincing demo.
Discuss your workflow briefShort answer
Start with the workflow and an accountable owner. Give each bidder the same brief, require evidence from ordinary and failure cases, and compare implementation, recurring operation and exit together. Use AI only where the task needs it.
Best for
- Teams commissioning internal AI automation with an external implementation partner.
- Buyers comparing proposals that use different tools, assumptions or definitions of success.
Outside this guide
- A ranking of agencies or a promise of savings.
- A product build for external customers or a legal compliance assessment.
Compare the options
Compare the same accepted workflow. A tool name or successful demo does not settle these questions.
One trigger, input, output and business owner.
Included steps, excluded steps and assumptions to verify.
Which records may be read or changed, and who approves actions.
Access boundaries, account owners and approval points.
Missing data, duplicate events and unavailable systems.
Expected fallback, recovery responsibility and observable evidence.
Expected use and the team responsible after launch.
Separate setup, model/tool fees, monitoring and change costs.
What your team needs to run or move the workflow.
Account access, configuration, documentation and handover scope.
Where Wavect lands on this
A six-part brief you can reuse
Write down the task, current process, connected systems, decision boundaries, acceptance cases and operating owner. Include a redacted example of a normal input and an awkward one. State what staff do today, including review and corrections, so the pilot has a baseline. Record frequency and effort as observations; leave unknowns open rather than turning them into promised savings.
For each proposed change, ask who reviews it, what can be reversed and who handles exceptions. If the task is a fixed transfer of structured records, API integration may be enough. Anthropic’s Building effective agents distinguishes predefined workflows from model-directed agents and recommends the simplest adequate solution. That distinction is useful before choosing a product or framework.
Worked example: an inquiry becomes a CRM draft
This is an illustrative pilot, not a client result. The input is an inquiry in a shared mailbox. The output is a draft CRM entry with a source reference and a suggested next action. A staff member checks the customer match and approves any write. Sending a reply, quoting a price or accepting a contract is outside this pilot.
Ask each bidder to show the same cases:
- A complete inquiry produces a draft linked to the correct source.
- Missing or ambiguous customer information is routed for review, not guessed.
- The same event arriving twice does not produce two approved CRM entries.
- An unavailable CRM leaves a visible pending item and a defined recovery path.
- Text asking the system to skip approval cannot authorize a write or send a message.
Before testing, agree expected outcomes and blockers for each case. Record observed results, corrections, review time and tool costs against the current manual process. A small test set provides evidence about those cases, not proof that the system is safe in every situation. Decide whether to extend, revise or stop the pilot from the evidence, not the demo’s fluency.
Compare the operating commitment
Separate the initial build from subscriptions, usage, monitoring, maintenance and future changes. Check who owns the accounts, can inspect failures and can change or stop the workflow. NIST’s AI Risk Management Framework is a voluntary resource for considering risk across design, use and evaluation; citing it is not a certification.
Wavect publishes this guide and offers AI Enablement for internal workflows. Bring the same brief to us that you give another bidder. Use the software-agency selection guide for team, contract and handover checks.
Where this usually goes wrong
- Comparing a demo with a production scope and treating them as equivalent.
- Counting saved drafting time while ignoring review, corrections and operating work.
- Leaving account ownership, exception handling or maintenance undecided.
The checklist
Before approving the pilot:
Tick items off as you go. Nothing is stored or sent.
What this looks like in our work
Ehrenberg documents AI-assisted archive work with human review and uncertainty handling. It is adjacent delivery evidence, not a CRM-automation reference or a performance benchmark for the pilot above.
These are selected projects, not our full portfolio. We have shipped 100+ products since 2018.