By Shankar Ganapathy, Co-founder, Boomerang AI. Published November 2, 2026.
To evaluate an AI intro agent, test it on accounts where you already know the answer, check that every path it suggests shows evidence, confirm a person approves and sends every ask, and measure outcomes per intro rather than volume. An intro agent fails differently from an email agent. A weak cold email costs a reply. A bad intro ask costs a relationship, and it is your board member's or customer's reputation on the line, not just yours.
Below are 12 tests you can run in a trial with any tool, including ours.
Why do intro agents need different evals?
Most AI sales agent evals focus on message quality and volume. Adoption is still early: in SDR Leaders Global's State of Sales Development 2026, only 16% of organizations use AI SDR agents and 59% never have. And volume is the wrong yardstick here. For intros, the questions are different: did it find the right person, did the right person ask, and did the connector feel respected.
The 12 tests
Which metrics should a pilot track?
- Coverage: share of target accounts with at least one evidence-backed path.
- Approval rate: share of drafted asks the relationship owner actually sends.
- Edit rate: how much senders change drafts before sending.
- Accept rate: share of asks connectors accept, and decline reasons for the rest.
- Outcomes: meetings and opportunities per intro, and days from ask to meeting.
Decline reasons are the most useful eval data you will get. They tell you whether the agent picked the wrong connector, the wrong moment or the wrong sender.
How does Rudy handle these tests?
We built Rudy to pass them. Paths show their evidence, a shared employer alone is ignored, and scores weight persona, seniority and department. Rules let you leave out people, companies, domains or whole account categories. The relationship owner approves and sends every ask. Connectors can answer in Slack with Accept, Decline, Needs Clarification or Why This Intro, as shipped in the week of October 26. Every request and outcome is logged to the opportunity.
Run the tests on your own network: Starter comes with a 7-day free trial. For the approval model in more depth, read human in the loop for intro agents.
Frequently asked questions
How do you evaluate an AI intro agent?
Test it on accounts where you already know the best path, check that every path shows evidence, confirm a person approves and sends every ask, check that routing and leave-out rules hold, and measure outcomes per intro, such as accept rate and meetings, rather than message volume.
What metrics matter in an intro agent pilot?
Coverage of target accounts with an evidence-backed path, the share of drafts the relationship owner sends, how much they edit drafts, the connector accept rate with decline reasons, and meetings and opportunities per intro.
Why are intro agents riskier than email agents?
A weak cold email costs a reply. A bad intro ask costs a relationship, and it spends the reputation of your board member, customer or partner. That is why approval, routing rules and evidence matter more than speed for intro agents.
Should an intro agent send messages on its own?
No. The person who owns the relationship should approve and send the ask from their own account. The agent can find the path, draft the ask, follow up and log the outcome, but the decision to ask stays with a person.
.png)
