July 14, 2026

Catch up on Part 1 and Part 2 if you're just joining — we left off with a list of ~20 tasks worth investigating for optimization.

The Final Filter, and Measuring What Matters

Step 7: [Filter 1] Does it move a feeling?

Before AI enters the conversation at all, run your ~20 tasks back through the four outcome factors from Part 1:

• Cross-functional audience — how information reaches non-R&D teams
• Stakeholders feel confident — predictable timing, clear scope, tested use cases
• Don't burden PMs — PMs aren't the bottleneck or the translator
• Minimize hand-holding — the process is repeatable and self-serve

The hypothesis is simple: if changing a task impacts one of these statements, it's worth pursuing — regardless of whether AI is involved. This step alone will tell you which tasks need to be addressed. AI appropriateness is a separate, second question.

Step 8: [Filter 2] Should AI touch this?

For the tasks that survive the first filter, run a second pass — this time asking whether AI is actually the right tool:

• Blast radius — if this goes wrong, how big a deal is it? (Michael Domanic, Head of AI at Section has a good article about these risks)
• Reversibility — is this a two-way door I can back out of?
• Access to tooling — is there a real interface for an agent to act through?
• Access to data – do I have access to the right connectors, cleanly?
• Usefulness of data — is the contextual data in a state an agent can actually use?
• Data privacy compliance — am I allowed to use this data?
• Generative response risk — can this output tolerate some creativity, or does it need to be deterministic?

Worked example