How to test an AI sales agent before you give it your accounts
Three tests to run on any AI SDR before it touches your LinkedIn or inbox: your voice, a hard conversation, and the full path across channels.
Every AI sales agent says it writes in your voice, holds real conversations and runs your whole outreach across channels. Ask any vendor "can it do this?" and the answer will be yes. The answer tells you nothing. What tells you something is a test you set up yourself, on your own messages, before the agent gets access to your LinkedIn, your mailbox or your WhatsApp.
This is the test we would run on any vendor, including us. It takes about half an hour on a call and a week in a pilot, and it checks the three things that decide whether an agent helps or quietly costs you deals: whether it sounds like you, whether it can stay in a hard conversation, and whether the full path from a list to a meeting actually works across channels.
Why a demo proves almost nothing
A demo is a conversation the vendor has rehearsed. The prospect in it asks the questions the agent handles well, the replies are short, and nobody says "send me the price and I will think about it". Real conversations are not like that. Real prospects go quiet for two weeks, answer on a different channel, ask for a discount in the second message, or say "we already have a vendor".
So the rule for the call is simple: do not ask "can you", ask "show me", and bring the material yourself. Twenty of your own sent messages, three real conversations that went badly, one list of companies you actually want to reach. If the vendor cannot work with your material on the call, you have learned the most important thing already.
Test 1: does it actually sound like you
An agent that writes from your account in someone else's voice is worse than no agent. Your clients notice when your two-line LinkedIn replies turn into polite paragraphs.
Ask these questions, and listen for specifics:
- What does it learn from? The answer you want: your real replies, paired with the messages you were answering. Not a "tone" dropdown with options like friendly or formal.
- Is the voice per channel? People write differently in email and in a chat. One voice for everything is a sign the vendor has not looked closely.
- How do they measure the match? "Our customers love it" is not a measure. A good answer is a score against your own replies that the agent did not learn from, so it cannot just copy.
- Can you test it blind? Ask to see your real reply next to the agent's reply with the same meaning, and try to tell them apart.
- What happens before there is enough history? A new account has nothing to learn from. The agent should say so, not pretend.
Then do the live part: give the agent five messages your clients really sent you, the uncomfortable ones, and read what it would answer as you. Nothing has to be sent to anyone for this.
Test 2: can it hold a long, difficult conversation
First replies are easy. The hard part is message eight, when the person asks something the agent was never told, or pushes on price, or has gone silent for a week.
Ask for a real conversation, anonymized, of ten to fifteen messages, where the person doubts, asks unexpected questions and is not ready to book a call. Then ask:
- When does it hand the conversation to me? A good agent has an explicit list: money and terms beyond what you allowed, an unhappy person, a question outside what it knows, a person who suspects a bot. "It figures it out" is not a list.
- What happens when I write in the thread myself? The agent should go quiet there, not answer over you.
- Can I set how much it sends on its own? You want a mode where you approve the first letters, and a way to loosen it later.
- What does it read before replying? The whole conversation, the contact's card, your CRM, your notes, or just the last message.
Answers worth hearing
- A named list of reasons it calls you
- It goes quiet when you write
- First letters wait for your approval
- It reads the whole thread and your CRM
Red flags
- It handles everything itself
- No way to stop it in one conversation
- It negotiates price on its own
- Only the last message counts
Test 3: does the whole path work across channels
The most expensive failures happen between steps: the list is wrong, a person answers by email to a LinkedIn message and the agent starts from zero, or an account gets restricted because the tool sent too much.
Ask to see the full path on your own example: from a description of who you sell to, to twenty relevant people, to letters written for each of them, to a reply, to the next message in a different channel. Then ask about the joins:
- How does it know two profiles are one person? Exact matches of an email or profile are safe. Guessing by name is how you end up writing to the wrong Alex.
- Does the agent see what was said in the other channel? If not, it will repeat itself.
- What limits does it hold per account? A good answer ties volume to the health of each account, not to the plan you pay for.
- Which channels have restrictions? For example, writing first on WhatsApp to strangers gets numbers banned. A vendor who says "no restrictions" is not being straight with you.
A one-week pilot that tells you the truth
Pilot before you trust it
- Day 1: connect one account per channelLet the agent read your history. Check that it says when the voice is not ready yet.
- Day 2: test the voice on hard messagesFive real client messages, read every answer. Try to tell yours from the agent's.
- Days 3–6: one narrow segmentA few hundred people, first ten letters under your review. Read every letter.
- Day 7: judge by meetingsCount meetings that took place, not letters sent or replies.
Keep the pilot small on purpose. A few hundred people is enough to see the quality of letters and conversations, and too few to damage a domain or a LinkedIn account if something is wrong. We wrote more about where agents earn their money and where they fail in do AI SDRs actually work.
How Nitrino answers these questions
We wrote down our own answers in public, so you can check them before a call rather than take them on trust:
- How Nitrino learns your voice: a voice per person and channel, a score out of 100 against your replies, a check where nothing is sent, and a blind "spot your reply" test.
- Long conversations and handoff: the reasons the agent calls you, what happens when you write yourself, and the four modes of how much it sends alone.
- One person in every channel: when two profiles are joined, and what the agent sees after a person moves channels.
- From one sentence to a meeting: the full path, the limits we hold to, and the same pilot as above.
Two weeks free with a code word
You can run the pilot above with us for free. Leave an application on the site and write the code word AGENTTEST in the Comment field: we will open two weeks of work at no charge, on your own accounts and your own segment.
Whoever you end up choosing, bring your own messages to the call. Half an hour with your material will tell you more than an hour of slides.
