An AI account-research agent built in sixty minutes: every step, and where it broke

There is a promise doing the rounds in every sales leadership meeting right now. Point an AI agent at your target accounts, and it will research each one, write the opener and book the meeting. No more junior reps spending their mornings on LinkedIn.

It sounds like the end of a chore. Here is the catch: the agent does not get tired, and it does not get embarrassed either. It will send a confident, wrong sentence to your best prospect at three in the morning, and nobody will know until the prospect replies, or does not.

What happened when companies tried it

In March 2025, TechCrunch reported that 11x, a venture-backed start-up selling an “AI SDR”, had listed companies as customers that were not customers. ZoomInfo said it had run a one-month trial and that the product “performed significantly worse” than its own sales development staff. Airtable said it ran a short trial, found the product was not a fit, and never used it in production. Former employees described hallucinations and results far below what customers expected. 11x disputed parts of the reporting and said its retention was 79%.

A few months later, Jason Lemkin, the founder of SaaStr, wrote about a well-crafted cold pitch he received from the chief executive of a marketing platform. It told him SaaStr Annual was eleven weeks away. The event had already happened. One wrong fact, and the deal was gone. When he pointed it out, the reply was that the AI is “really good 90% of the time”.

That is the whole problem in one sentence. In outbound, 90% right is not a pass mark. The 10% is what the buyer remembers.

The analysts see the same pattern at scale. In June 2025, Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, because of rising costs, unclear business value or weak risk controls. It also warned about “agent washing”: of the thousands of vendors calling their products agents, Gartner estimates only about 130 are genuine.

The build: sixty minutes, five steps

In this episode we build an account-research agent on screen, with the failures left in. The build is deliberately simple, because simple is what most teams can run.

1. Write the brief first. What does a rep actually need before a first call? We settle on five things: what the company does, who buys from it, one recent change, the likely owner of the problem, and one reason to talk now. Everything else is noise.
2. Choose the sources. The company’s own website, its newsroom and recent press coverage. Nothing that needs a login, nothing scraped against a site’s terms.
3. Extract, with a source for every line. The agent fills the five fields, and every field carries the link it came from and the date of that page.
4. Check it against your ideal customer. The agent scores each account against the profile your sales and marketing teams agreed, and says why.
5. Draft the opener, for a person to send. The agent writes a first line from the brief. A rep reads it, edits it and sends it.

Where it broke

These are the places an agent like this breaks, and the fix for each one.

• Old news, presented as new. Pages do not always carry dates, so the agent treats a two-year-old announcement as this month’s. The fix: no date, no use. If the agent cannot date a fact, it does not go in the brief.
• Confident facts with no source. When a field is hard to fill, the model fills it anyway. The fix: an empty field is allowed, an unsourced one is not. Blank beats wrong.
• The wrong company. Two firms share a name, and the agent blends them into one. The fix: anchor every account on its website domain, never its name alone.
• Openers that sound like everyone else’s. “I noticed you are growing fast” is not research. The fix: the opener must use the one recent change from the brief, or it is not sent.
• No human in the loop. This is the one that sinks deals. The fix: the agent prepares, a person sends. Every time, until the error rate is measured and acceptable.

Before you buy an AI SDR, ask three questions

• Can it show the source and date for every fact it uses?
• What happens when it does not know: does it leave a blank, or guess?
• Who reads the message before the prospect does?

If the answer to the last one is “nobody”, you are not buying a researcher. You are buying a risk to your brand, sent at scale.

Sources

• TechCrunch: 11x has been claiming customers it does not have, 24 March 2025
• SaaStr: AI in sales, incredible when it works, catastrophic when you do not review it, 25 July 2025
• Gartner press release, 25 June 2025