Connect the stack to the work.
Start with the outcome and the systems you already use. This guide helps product, IT, and engineering teams identify the decisions behind an AI deployment before assembling a list of tools.
Six areas to decide together
These are responsibilities to cover, not six products to purchase. An existing application may cover several of them.
Where will people use the result?
Name the task, its current baseline, the user, and the person accountable for mistakes. Start with one workflow.
Which information is reliable enough to use?
List source systems, content owners, update frequency, and deletion rules. Resolve contradictory or obsolete material before expanding retrieval.
Whose permissions apply?
Define user identity, service permissions, and tenant boundaries. Test denied access and revoked access, including cached answers.
What needs to be generated or found?
Compare existing capabilities before adding infrastructure. Measure retrieval quality separately from answer quality when the workflow depends on company documents.
What can the system change?
Separate reading, drafting, and writing. Define approval, duplicate-action handling, and recovery for every external action.
How will the team know it is working?
Assign test ownership, acceptance criteria, monitoring, spending limits, incident response, and a manual fallback. Include human review effort in the cost.
Choose the smallest workable approach
| Approach | Worth investigating when | Resolve before committing |
|---|---|---|
| Improve the current process | The main problem is missing ownership, inconsistent content, or unnecessary handoffs. | Whether cleaner inputs or simpler rules solve the task without AI. |
| Use an existing platform | The workflow and data already live in a system with relevant capabilities. | Actual license entitlement, access behavior, quality, export, and operating cost in your environment. |
| Buy a specialist product | The workflow has needs your current tools do not meet. | Integration effort, proof against your test cases, data handling, and exit options. |
| Build a custom workflow | Requirements need control that available products cannot provide. | Who will own maintenance, evaluation, security review, and incidents after the prototype. |
Two examples from platform documentation
Microsoft documents that Copilot respects existing permissions and recommends addressing oversharing in SharePoint and OneDrive. Access cleanup belongs in the deployment plan. Microsoft: prepare SharePoint for Copilot.
AWS distinguishes evaluating retrieved information from evaluating generated responses. A useful answer test should not hide a weak retrieval step. AWS: evaluate RAG sources.
Apply the guide to one workflow
Prepare a pilot review in the Workbench →
The Workbench organizes your decisions and open questions. It does not inspect your systems or certify readiness.