TL;DR: Don't set ICPs and forget them. After enough outcome data accumulates, ask Copilot to analyze what's actually converting and propose specific edits.
The 4-week rule
The first time you set up an ICP, you're guessing. After 2–4 weeks of outcomes flowing through, you have signal. The cadence:
- Week 1 — set up. Don't change anything.
- Weeks 2–3 — accumulate outcomes. Reply rate, meeting rate, signups by ICP × Persona × signal source flow into Insights.
- Week 4+ — run the refinement loop with Copilot.
The exact cadence depends on volume. High-volume motions can do this weekly; ABM motions may need quarterly.
The refinement prompts
Three prompts that surface different kinds of refinement:
"What's working"
> for the Mid-market SaaS ICP, what parameters correlate with the highest
reply rate over the last 30 days?
Copilot pulls outcome data, runs the correlation against each Score parameter, and tells you which ones predict replies. Often you'll see things like:
Records with uses Snowflake: yes reply at 14% vs 4% for the rest of the ICP. The Snowflake parameter currently has weight 5; recommend raising to 8.
Confirm or refine: > raise it to 8 and re-qualify the workspace.
"What's noise"
> for the Mid-market SaaS ICP, which Score parameters have NO correlation
with reply rate?
These are dead weight in your scoring. Copilot proposes either dropping them or moving them to Filter (if they're actually requirements you'd missed) or to Exclude (if they're inversely correlated).
"Where's the gap"
> what attributes are present in my closed-won customers but missing
from my Mid-market SaaS ICP definition?
This is the most powerful one. Copilot looks at your closed-won data and compares it to your ICP definition. If 80% of closed-won customers have a specific tech-stack tool you didn't include, Copilot suggests adding it as a Score parameter.
Applying the suggestions
Three modes:
- Accept directly —
> yes, make those edits and re-qualify. Copilot writes the changes to the ICP draft, prompts you to publish. - Stage as draft —
> stage those edits but don't go Live yet. Copilot updates the Draft version of the ICP; you review in the drawer before promoting. - Cherry-pick —
> apply only the Snowflake weight change, leave the rest. Copilot applies a subset.
Every change runs against the workspace's outcome history, so you can preview the impact: > what would the qualified count be if I made these changes?.
What NOT to change reflexively
Three traps:
- Don't add parameters for outliers. A single closed-won customer with a unique trait isn't a pattern. Wait for at least 3–5 matching outcomes.
- Don't tighten Filters early. The first instinct is "we got noise, let me add a Filter." But Filters are gates — they exclude records you might still want to score lower. Use Score with negative weight (Exclude mode for true never-touch, Score with weight 0 won't work — use Exclude).
- Don't refine more than once per 2 weeks. Each change re-qualifies the workspace and re-routes audiences. Constant tweaking thrashes destinations and pollutes the outcome data you'd use for the next refinement.
Persona refinement is parallel
Everything above works for Personas too. The prompt template:
> for the VP Sales Persona inside Mid-market SaaS, what title patterns
reply at the highest rate?
Often you'll find that "Chief Revenue Officer" replies at very different rates than "VP Sales", even though semantically they're close. Copilot can split the Persona, tighten the title patterns, or add a tenure parameter.
Tracking the refinement loop
Each ICP and Persona has a Modified by and Last modified field on the Qualification table. You can see at a glance what's been touched recently and what's stale. A stable refinement schedule is: monthly review of every Live ICP, quarterly review of every Live Persona. Copilot can prompt you: > which of my ICPs haven't been refined in the last 60 days?.