Is your AI sales agent writing to the right person?
For Heads of sales, CTOs and product owners who are letting an AI agent pick and contact real people.
Our sales agent, AI SDR, finds the right person at a target company. On a run of 47 small software firms it picked someone at a different company 11 times, often with a verified email. It had been given the company's website and never used it.
- AI sales agents
- Our own AI sales agent
- People
- Names
- Records
- Tests
What was at risk
A stranger gets the email
The agent unlocked someone at a different company, mostly with a verified email that a caller would have mailed.
A look-alike firm gets matched
A firm was matched with an accounting firm of a similar name. Another was matched with a firm one letter away.
Old wrong answers served again
Caches hold earlier picks. They now serve a saved pick only if it passes the same rule, and old wrong answers are searched again.
Fixtures that hide the bug
A step that never worked passed its tests because the tests handed it a bare list, while real responses were objects.
Five questions we ask of any AI agent that picks a person
- InputsIs every input you hand the agent actually used? If one never gets used, why was it passed?
- MatchDoes the agent check equality of the employer, or only similarity of the name?
- MissDoes a no-match end as a miss for a human, or as a plausible wrong answer?
- CacheIs a saved answer served only if it passes the same rule as a fresh one?
- FixturesDo the tests feed the real shape of the data, not a hand-built one?
How we found it
- Where
- AI SDR, our own sales agent, on a run of 47 small software firms.
- How
- We measured the new rule on the same 47. Three adversarial review rounds each found one Important defect, all fixed, and every fix was run against the code from before it so we saw the new tests fail first.
- What counted
- Wrong picks and right picks, measured on the same 47 firms before and after the rule.
- How we workedWe did not just fix the parsingWhere it showed upA step that guesses email addresses had never worked, because the domain lookup answered with an object and the code accepted only a bare list. Review showed that looking up a domain by company name cannot be made safe, so the search was removed.
What we found, and what it costs
- Finding1 of 3Wrong person: 11 of 47 picks were at a different companyWhat it costs youA caller would have mailed a stranger, mostly with a verified email.
- Finding2 of 3A dead step: no guessed email in any completed enrichment on recordWhat it costs youThe step looked like part of the system and produced nothing, because the tests never used the real response shape.
- Finding3 of 3The price of the strict rule: 3 of 26 right picks lostWhat it costs youThey become a miss for a human to look up, where the database added a word to the company name.
In one case the real chief executive was on the same page, ranked below the look-alike firm's founder. A domain looked up by company name resolved a similar-named firm to a different company's domain.
What changed
A person is now eligible only if the employer the database states equals the requested name, ignoring case, punctuation, accents and legal forms such as Ltd, SRL or GmbH. Old wrong answers are searched again, and the caches serve a saved pick only if it passes the same rule. Guessing emails now happens only at a domain the caller supplies, and a guessed email is never stored where it could be mistaken for a confirmed one.
The longer write-up: our AI sales agent picked the wrong person 11 times in 47.
Putting an AI agent in front of real people?
We test agents on real inputs and count what they get wrong before you do. The first month is a trial.
