HubSpot's new AI SDR guide: one team uninstalled theirs after duplicate CRM records sent prospect emails to customers
Before an AI SDR sends a single email, your CRM has to be clean enough to trust.
What happened
HubSpot published two guides this week on automating sales development. The AI SDR guide cites the Bridge Group's 2025 benchmark of 351 B2B companies, where only 1% reported AI SDRs as a distinct staffing model. It tells the story of Kyle Asay, SVP of sales at LaunchDarkly, whose team uninstalled its AI SDR a couple of months in. Duplicate accounts in the CRM meant customers got emails meant for prospects, personalization landed on the wrong details, and response rates sat near zero. The guide says AI SDRs are safer on inbound than cold outbound, and cites practitioners interviewed by SignalFire who earmark four to eight hours a week from RevOps and enablement to maintain the agent.
The automated prospecting guide covers the build side. Standardize job title, company size and region fields first. Pair enrollment triggers with suppression rules that remove a contact who has an active support ticket, an open deal or an opt out. Use score thresholds, with an example of 15 points for a demo video, 10 for a webinar, 20 for an enterprise pricing page visit and a trigger at 60. Keep a rep reviewing AI drafted emails before they send.
My take
Every AI SDR story I have seen go wrong went wrong in the data, not the model. The LaunchDarkly case is the clearest version: a duplicate record is a small problem for a human rep and a large one for an agent that emails at scale.
So I build these in a fixed order:
- Dedupe and lock the fields that routing depends on, with dropdowns instead of free text.
- Write suppression rules before enrollment rules. Know who must never get the email.
- Start on inbound only, where the buyer asked to hear from you.
- Keep human approval on drafts until the reply review backlog is boring.
- Track response time and handoff speed, not emails sent.
And budget the upkeep. HubSpot's guide cites four to eight hours a week to maintain one. If nobody owns the agent after launch, it will drift, and the first sign will be a confused customer.
More posts
- Censys found about 815 AI hacking consoles open to the web. The attacks they ran mostly used weak passwords and missing MFANov 7, 2026
- Thinking Machines' Inkling 2 Small scores 31 on Artificial Analysis at $1.18 per task, but uses about 4x the output tokensNov 7, 2026
- Zapier watched Claude Code read its site. The agent pulled the Markdown version, and a second model summarized it firstNov 7, 2026
