Judge an agent on three things: where its data comes from, whether you can see why it decided something, and what it is allowed to do without asking. Features and model names tell you very little. The questions below are ordered so the ones that disqualify a vendor fastest come first.
Disclosure: Signl is an agent-native prospecting product, so we are in this category. This article names no winner and ranks nothing, deliberately. Where our own answers are relevant they are given, so you can hold us to the same checklist.
The three things that actually differ
Most comparison content in this category lists features. Features converge within a year; these three do not.
Where the data comes from. An agent is a reasoning layer over a data source. If the underlying data is a bought static database, no amount of agent quality fixes staleness. Ask what the source is and how old a given record can be.
Whether decisions are inspectable. When the agent says a person is a good fit, can you see why, in words? A confidence score is not an explanation. This is the difference between a tool you can correct and one you have to trust.
What it may do unsupervised. An agent that drafts is a different risk from one that sends. Both are legitimate; confusing them is how a team discovers that four hundred messages went out in their name.
Twelve questions
Ordered by how quickly a bad answer saves you time.
Data and sourcing
1. Where do the people come from, and how fresh is a record when I see it? A live query and a quarterly crawl are different products at the same price. Our own view of that split is in live query vs static database.
2. What happens when the agent cannot find something? You want "returns nothing with a reason". The wrong answer is a plausible guess, which in this category means an invented job title or an unverified email.
3. Is contact data charged per attempt or per result? Per attempt means you pay for the misses, and an agent is thorough, which makes it expensive in exactly the way a human is not. The bounce tax is the long version.
Qualification
4. Show me why this person was qualified, in a sentence. If the answer is a number, the agent cannot be debugged and neither can your ICP.
5. Can I see who it rejected and why? The rejected pile is where you find out the filter is wrong. A product that hides it is withholding the most useful output it has.
6. What does it do with missing data? Treating an empty field as a pass is the most common silent failure in qualification, and it produces confident nonsense.
Actions and boundaries
7. What can it do without asking me? Get the list. Sending, deleting, connecting an integration and anything irreversible should be confirmed. Signl confirms sending people, removing a source, deleting an autopilot rule and recording an objection.
8. What are the rate limits, and who enforces them? An agent with no cap is one bug away from an incident. Ours is 10 autopilot rules and 500 sends a day per account.
9. Does it send, or does it hand off to my tools? Handing off keeps deliverability and account risk in systems you already control. Sending itself is more convenient and concentrates the risk in the vendor.
Integration and control
10. Does it write back, and does anything come back? One-way integrations rot. Ask specifically which destinations return replies, meetings or deals, because that is what lets you tell which sources worked.
11. How do I turn it off at three in the morning? A real question. There should be an answer that does not involve emailing support.
12. What happens to my data if I leave? Export format, deletion policy, retention period. Signl deletes collected engagement 90 days after first collection by design.
Category boundaries worth checking
Vendors in adjacent categories all describe themselves as AI prospecting agents. The distinctions that matter commercially:
| Category | What it is for | Where it stops |
|---|---|---|
| Contact database with AI search | Finding people who match criteria | Says nothing about timing |
| Enrichment and orchestration | Filling in and routing data you bring | Does not discover people |
| AI SDR | Writing and sending at volume | Judgement, and anything off-script |
| Signal or listening tool | Finding people who did something | Volume. There is a ceiling |
| Agent-native data layer | Letting your own assistant do the work | Needs you to have an assistant and a workflow |
Work out which of those your bottleneck is before taking demos. A team whose problem is "we have lists and nobody works them" buys a sending tool. A team whose problem is "we do not know who to call" buys sourcing. These are frequently confused, and the published buyer's guides mostly assume the first.
A pilot you can run in a fortnight
Demos run on the vendor's data, where everything works. Design the pilot to break that.
Week one, one segment. Pick fifty real prospects from one tight segment of your actual ICP. Not your whole market.
Read every output before anything sends. The question is not whether the messages are grammatical, it is whether the reasoning is right about your market. Count how many you would be embarrassed by.
Deliberately include hard cases. Put in an agency that looks like a buyer, someone who just bought from a competitor, a job seeker with a senior title. A good agent rejects all three and tells you why.
Check the rejections, not just the approvals. Score the agent on both.
Week two, let it run on a small cap, then measure one thing: qualified conversations, not emails sent or contacts processed.
Two cheap tests worth adding. Ask it something it should not know and see whether it says so or invents an answer. And ask the same question twice an hour apart to see whether the answer is stable.
Pricing traps
- Per seat, with an agent in the team. An always-on agent either shares a human's seat or you pay for a phantom one. Ask directly.
- Credits that meter two things. Actions and data credits metered separately mean the sticker price is a floor, not an estimate.
- The platform fee is not the cost. Domains, inboxes, warming and data are routinely outside the headline figure.
- Annual contracts in a category with high churn. Published churn for AI SDR tools was put at 50 to 70 per cent a year by one vendor's own analysis; see AI SDR vs human SDR. Ask for monthly until it has worked for a quarter.
For reference: Signl is $99 a month for 25 sources with unlimited seats, no credits and nothing charged per person, with a 30-day money-back guarantee. The limits that make flat pricing possible are published on the fair use page.
If your team already uses Claude or ChatGPT
One question changes the shortlist: does it speak MCP?
An agent reachable over MCP is driven from the assistant your team already has open, rather than being another product to adopt. The practical difference is whether prospecting happens in the same conversation as everything else, or in a tab somebody has to remember to open.
What that looks like in practice is in Claude LinkedIn prospecting, the field is surveyed in the best MCP servers for sales, and the questions your security team will ask are in MCP security for revenue teams.
Limits
- This checklist ranks nothing. The right answer depends on your bottleneck, and a tool that fails nine of these questions may still be correct for a team whose constraint is the tenth.
- Pilots favour incumbents with good onboarding, which is not the same as good software. Weight the output, not the setup experience.
- Everything in this category is moving quickly. Re-ask question one and question seven at renewal, because both change quietly.
