Buyer’s checklist · 21 questions

Questions to Ask Before Buying an AI Sales Agent

Take these into every vendor call. Each question tests a claim rather than accepting it, says why it matters and what a good answer sounds like. They follow the five criteria behind the AI sales agents ranking.

Repeated-run results

A demo shows the agent's best run. You are buying its average one.

  1. Can you show results from the same task run many times, not one demo?Agents vary from run to run. A single good output says little about the next thousand.Good answer: A pass rate or a mean with a spread over many runs, and the task it was measured on.
  2. How often does a run stop early, skip a step or need restarting?Long-running work fails on stamina as much as on quality: the agent gives up or loses the thread.Good answer: A number, and what the platform does about it (resumable runs, schedules, alerts).
  3. Will you run it five times on our accounts during the trial?Your data and ICP are not the demo's. The only results that count are on your accounts.Good answer: Yes, with the outputs exported so you can compare runs side by side.
  4. When you change the model or the prompts, how do you know nothing got worse?Vendors update weekly. Without a regression suite, last month's good agent can quietly get worse.Good answer: A named benchmark or test suite every change runs through.

Data sources and freshness

Stale data does more damage than a weaker model.

  1. Where does each contact and fact come from?Your CRM, the vendor's database, third-party enrichment and web research have different failure modes.Good answer: A list of sources per field, and which wins when they disagree.
  2. When is a record re-checked, and how old can it be when the agent uses it?Roles change and people leave. A database refreshed quarterly is months out of date for some contacts.Good answer: A re-check on read or before send, with an age threshold you can see.
  3. Is every email verified before it is sent?Bounces hurt domain reputation for every rep who sends from it.Good answer: Verification at send time, not only at import.
  4. What does the agent do when two sources disagree?Agents tend to trust the database over the company's own website. The right answer is usually the fresher source.Good answer: A rule that prefers the more recent, first-party source, and flags the conflict.

Review steps

Decide where a person looks before anything reaches a buyer.

  1. Where does output wait for a person to approve it?Drafts, new contacts and CRM changes all deserve a check while you learn how the agent behaves.Good answer: A review queue for each, and per-task settings for what runs unattended.
  2. Can we see why the agent chose each account and each message?A reviewer needs the reasoning and the source to judge an output quickly.Good answer: The signal, the source and the fit reason shown next to every draft.
  3. What share of drafts do reviewers at your customers change or reject?It is the honest measure of how much work review saves and costs.Good answer: A number from live accounts, or a commitment to measure it in your trial.
  4. Does the agent learn from what we reject?Without feedback, reviewers fix the same mistakes every week.Good answer: Rejections and edits feed the agent's playbook, and you can see what changed.

Suppression and Do Not Contact

These must be rules the system enforces, not instructions the agent may follow.

  1. Where are suppression and Do Not Contact enforced: in the prompt or in the send layer?An instruction in a prompt can be ignored or forgotten. A check at send time cannot.Good answer: A check in the send layer that runs whatever the agent decided.
  2. Are current customers and contacts on open deals excluded automatically?Cold outreach to someone mid-negotiation or already paying is the costliest mistake an agent can make.Good answer: Automatic exclusion read from your CRM, kept in sync.
  3. How fast is an unsubscribe or opt-out honoured, across every channel?An opt-out on email should stop LinkedIn and calls too, and stop them at once.Good answer: Immediately, across channels, with a log you can audit.
  4. What sending limits, warm-up and domain protections are on by default?Agents can send far faster than reps. Your domains carry the risk.Good answer: Per-mailbox limits, warm-up and automatic pausing on deliverability problems.
  5. Can we test it? Put our own address on the list and try to reach it.The quickest proof that suppression works is a failed attempt to break it.Good answer: Yes, during the trial, with nothing reaching the suppressed address.

Pipeline outcomes and fit

Activity is not pipeline.

  1. How do you attribute meetings and opportunities to the agent's actions?Without attribution you cannot tell what the agent added from what would have happened anyway.Good answer: Replies, meetings and opportunities linked to the agent action that started them, in your CRM.
  2. Will you support a holdout test against our current process?Comparing to a group the agent did not touch is the only fair read on lift.Good answer: Yes, with the split agreed before the trial starts.
  3. Do we need to replace our CRM or sequencer to use it?Migrations add cost and risk. Some agents work on top of your stack, others replace parts of it.Good answer: A clear answer, and the integrations it reads from and writes to.
  4. What does it cost at our volume, and what happens when we go over?Credit, seat and per-contact pricing behave very differently as you scale.Good answer: A written price for your expected volume and the overage rate.

Answers are claims too. The way to settle them is a trial on your own accounts, run more than once and measured against pipeline: how to test an AI sales agent.