Key takeaways
- Pick a workflow that is frequent, repeatable and easy to check.
- Write the hypothesis and the success measure before anyone starts.
- Compare against a group doing the same work by hand.
- Keep actions that reach customers behind human approval during the test.
- End on the planned date and make a clear decision.
Which workflow should you pick first
Pick a workflow that happens many times a week, follows the same steps each time and produces output a person can judge quickly. Account research before a first call, finding contacts at target accounts and drafting first-touch emails for one segment all fit.
Avoid workflows that are rare, politically sensitive or hard to judge, such as pricing negotiations or executive outreach. A first experiment should teach the team how agents behave, not risk a key account.
How do you design the experiment
Design it like any sales test: one hypothesis, a test group, a comparison group, a fixed period and measures agreed in advance. Keep the segment and the sellers similar across both groups so the difference comes from the workflow.
Take an invented logistics software company with two SDR pods selling into the same region. Pod A uses an assistant to research each account and draft the first email; pod B works as before. For a month, both log time spent per account, how much each draft was edited and what happened after the first touch.
- Hypothesis: agent research and drafts cut preparation time without lowering reply quality.
- Test group: one pod using the workflow; comparison group: one pod working by hand.
- Measures: preparation time, editing needed, replies and meetings from the segment.
- Period: a fixed number of weeks, with the end date in the calendar.
Who should be involved
Involve a small group with clear jobs: one owner who answers for the experiment, a builder who sets up the workflow, the sellers who use it, and a sponsor who will make the final decision. The owner is often the sales manager of the test pod.
Tell security and legal early if the workflow touches customer data or sends anything externally. A short review at the start costs far less than stopping a live test.
What guardrails should you set
Set guardrails that keep people in charge of anything a customer sees. Run the agents in a mode where sends, record updates and lead searches wait for approval, review every draft at the start, and keep a simple log of anything the agent got wrong.
Also agree a stop rule. If drafts contain factual errors about a prospect or your own product more than a handful of times, pause, fix the context the agent works from and restart.
What do you do when the experiment ends
When the experiment ends, compare the groups on the measures you agreed, read the error log and make one of three decisions: scale the workflow to more pods, change it and run again, or stop. Write the decision and the reasons down.
If you scale, keep the same owner and add the next workflow only once the first runs without daily attention.
Where Pebbles Ai fits
- CapabilityResearch & analysisAn assistant reads a target company's website and public sources and returns a summary with inline numbered citations and a source list, so reviewers can check each fact.
- CapabilityProspecting & customer recordsFind fresh leads takes up to ten named companies, with the roles and locations you want, and lands the contacts in a CRM segment (in Beta), so the test group's accounts stay in one place.
- CapabilityAI workspaceIn Manual mode, starting a lead generation run, sends and CRM writes wait for your approval, which suits a first experiment.
Common questions
How big should the test group be in a mid-market team?
Big enough for the workflow to repeat many times in the period, and small enough that someone can review every draft. One pod of a few sellers, with a similar pod as the comparison group, usually fits.
Do we need a comparison group?
Yes, if you want a fair answer. Without one, a good or bad month in the market looks like the agent's doing.
What if sellers do not trust the agent's output?
Treat their edits as data. Ask them to note what they changed and why, then fix the context or instructions the agent works from.
About this page. Written by the Pebbles Ai team for people who run sales, marketing and growth. Last updated .
