Where this checklist comes from
The best public evidence on whether AI agents can do real GTM work is PipelineBench, published by Frontrunner. It asks: given a company, a product, a CRM and a budget, how much qualified pipeline would an agent create as the first GTM hire for a year? Models run inside the same harness with the same tools, data and budget, 30 year-long simulated runs per model. Three of its findings shape the checklist below.
What PipelineBench found
One good run is not an agent. 1 in 10 model and task pairs passed one run and failed the next.
Fresh data beats a better model. With a year-old prospect database, models went with the stale number about half the time, even when the company's own site said otherwise.
Stamina separates agents. Below the top group, the gap was mostly stamina: one model stopped working before December in 8 of 30 runs, another in 13 of 30. No model came near the ceiling; the best reached 66% of what a run with perfect information gets.
Source: PipelineBench, usefr.com, published October 29, 2026. Results are simulated pipeline, not a customer's revenue.
1. Repeated-run reliability
- Run the same task on the same inputs at least five times. Count how many runs you would have accepted.
- Ask the vendor how they measure run-to-run variance, and for the number.
- Run a long task (a month of follow-ups, a full territory) and check whether the agent finishes it or stalls.
- Ask whether runs are resumable: if a run fails on day 12, does it pick up where it stopped?
- Ask what changes when the vendor swaps the underlying model, and how they regression-test it.
2. Data freshness
- Give it 20 contacts you know changed jobs recently. See whether it uses the old title or employer.
- Ask where in the flow records are re-checked: on import, on a schedule, or when they are read.
- Ask whether emails are verified before every send, or only when the record was created.
- Ask whether job changes arrive as signals, or the agent has to go and look.
3. Human review
- Confirm you can hold every draft, enrichment and CRM change for approval before it takes effect.
- Check the hold applies to automated and scheduled runs, not only to ones you start by hand.
- See what a reviewer sees: the output alone, or the output with the reason and the sources behind it.
4. Suppression and send-layer safety
- Add a test contact to your do-not-contact list, then ask the agent to email them. It should be blocked at send.
- Put a current customer and a contact on an open deal into a prospecting run. Both should be excluded.
- Ask whether these checks live in the send layer or in the agent's instructions. Only the first is a control.
5. Pipeline-outcome benchmarks
- Ask what outcome the vendor measures: replies, meetings, attendance, qualified pipeline or closed revenue.
- Ask how many runs, over what period, against what baseline. A single customer story is not a benchmark.
- Ask how they know the grading is right: a ceiling run, a do-nothing baseline, and shortcut strategies that should score low.
- Ask which model sits inside the product, and where that model ranks on a public benchmark such as PipelineBench.
How the agents compared here measure up
Frontrunner documents every item on this list: freshness checks when records are read, email verification before send, drafts held for review, suppression and customer and open-deal checks in the send layer, resumable runs on schedules and triggers, and a published benchmark that doubles as its regression suite. It is the agent to start with if you want work you can verify. It is not a CRM, a sequencer or a data provider: it works across the ones you have.
For Unify and Alta, most of these items are not published on their sites, so ask. See the comparison table, Frontrunner vs Unify and Frontrunner vs Alta.