Clay + Flint / November 17, 2026 / 4 minute read
Start with the work, then choose the technology.
The most useful AI question is often ordinary: where does the team spend time turning known information into a repeatable output? A strong first candidate has a clear input, an observable result and a person who can judge whether it is correct.
Examples include preparing an interview outline from approved research, turning a verified expert transcript into a draft article, or assembling an internal lead brief from permitted sources. Begin with a task you can inspect before giving a system more freedom.
Know which kind of workflow you need.
| Approach | Useful when | Example |
|---|---|---|
| Standard automation | The rules and route are predictable. | Create a follow-up task when a qualified form arrives. |
| AI assistance | Interpretation or drafting is useful, with a person reviewing the output. | Summarize an approved interview into a source-linked article outline. |
| Agentic workflow | The system must choose among permitted steps or tools to reach a bounded goal. | Gather approved account context, check required fields and prepare an internal handoff for review. |
An agentic workflow is not automatically a better solution. More possible actions create more behavior to evaluate. Use the simplest approach that reliably completes the job.
Draw the workflow before building it.
Record the trigger, allowed sources, permitted tools, expected output, reviewer and completion condition. Include the exception path: what happens when data is missing, two sources conflict or a connected service is unavailable?
Define quality in examples.
Create a small set of representative cases before optimizing the prompt. Include ordinary inputs and difficult ones: a blank record, ambiguous company names, stale information, contradictory instructions and content trying to redirect the workflow.
Write the expected behavior for each case. A reliable workflow should flag missing evidence, respect its permission boundary and escalate when it cannot complete the task. “Sounded good” is not a sufficient acceptance test. Repeat the evaluation when prompts, tools, permissions or models change.
Treat sources as evidence, not instructions.
A webpage, email or uploaded document may contain text that asks a system to reveal information or take an unrelated action. The workflow should treat that material as content to analyze. Operating permissions and approval rules should come from the trusted workflow configuration.
Restrict access to the records and actions needed for the task. A summarization pilot rarely needs deletion rights, bulk exports or permission to send messages. Verify provider and data-handling arrangements before introducing confidential or personal information.
Pilot with a real reviewer.
Run the workflow alongside the current process before relying on it. Log what it attempted, what it produced, where it stopped and what the reviewer changed. Assign someone to inspect exceptions, maintain credentials and pause the workflow if quality degrades.
Keep consequential actions—such as publishing, customer messages, contract changes and spending—behind agreed approvals. The appropriate level of autonomy follows evidence from the pilot and a separately approved operating scope.
Measure the entire cost of completion.
Measure total time through accepted completion, including preparation, review, corrections and exception handling. Record accepted outputs, unsupported claims, tool costs and failed runs. Compare that with the same task completed through the existing process.
An early pilot can establish feasibility without proving ongoing savings. If review effort consumes the time saved, narrow the task, improve the sources or keep the current process. A useful strategy includes reasons to stop.
What should remain after the pilot?
- A workflow map with the owner, sources and permissions.
- A reusable evaluation set and agreed pass conditions.
- A log of reviewed outcomes, costs and unresolved exceptions.
- A runbook explaining approvals, monitoring, recovery and shutdown.
- A decision to expand, revise or stop, with the evidence behind it.
Our Agentic AI strategy approach connects opportunity selection, a bounded pilot and a practical adoption plan. Implementation is scoped once the workflow and success criteria are clear.
Put the thinking to work.
Use the guided workbook to turn your business constraint into a measurable 90-day plan.
Get the free workbook More field notes