11x After the Walk-Back: What Buyers Should Ask Now

A category does not die from one bad quarter. It dies from a gap between what was sold and what shipped, held open long enough that buyers stop extending credit. The autonomous SDR category ran that gap in public, and every AI SDR conversation since happens in its shadow, whichever logo is on the deck.

This is not a hit piece on a company. It is the buyer's version of the story: what is documented, what is contested, and what you have to verify yourself before an AI SDR touches your domain reputation.

The documented timeline

Keep the record and the interpretation separate. Here is the record, with sources.

• Founded 2022. The product, "Alice," was positioned as an autonomous AI SDR and pitched as a direct replacement for sales development headcount.
• March 2025. A TechCrunch investigation documented that ZoomInfo and Airtable had been listed as customers. Both companies publicly denied it. ZoomInfo said it ran a one-month trial and found the product performed significantly worse than its own SDR employees (summary and link).
• May 2025. Founder Hasan Sukkar stepped down. A new chief executive was installed in Q2 2026, per Toarn's competitive analysis.
• Q2 2026. The company still exists. Self-reported retention of 79% is contested, with former employees describing internal messages showing 20 to 30% retention, per CheckThat.ai.

The vendor also publishes customer case studies, including Checkr, Leica Biosystems and Questex. Those are vendor-published and belong in the evidence file as vendor-published, which is a different weight class from a customer speaking to you unprompted.

What actually failed, and why it matters to your shortlist

Three failure modes sit under the story, and none of them are unique to one company.

One: the replacement frame

Selling a headcount replacement forces a comparison the product cannot win in year one. A human SDR who books eight meetings a month is not a benchmark an unsupervised system clears, and framing it that way guarantees the customer measures the wrong thing and then churns. The teams getting value from the same underlying technology bought capacity, not replacement, which is the argument in AI as capacity lift, not headcount cut.

Two: no verification layer

Generation at volume with no check between draft and send means every hallucination reaches a buyer. Published baseline hallucination rates frequently sit above two percent, per LinkedCamp's post-mortem. At list scale that is a brand-damage rate, not an error rate.

Three: the accounting shape

Short-trial contracts counted as full annual value, with a break clause behind them, produce a revenue number that disappears at renewal but stays on dashboards until then. That is a business-model failure rather than a technology failure, and it explains why the survivors in this category price per outcome or per resolution.

Where the same technology is working

The useful contrast is that the company which publicly rejected autonomous outbound reported roughly a sixteen times return on an inbound conversational deployment, per its own operator (1mind case study). Same buyer, same year, opposite result, different problem.

The pattern holds across the deployments we track. Inbound at the moment of intent, signal detection and enrichment, and research work that used to consume most of an SDR's day are all working. Unsupervised outbound at volume is where the reversals cluster. Our compiled evidence sits in the what works research theme and the named reversals ledger.

The seven-question reference script

Use this with two customers from after the leadership change. If the vendor cannot produce two, that is your answer.

1. When did you start, and what were you running before?
2. What is your cost per held meeting, all in, including the licence, enrichment data and the human hours spent reviewing output?
3. What percentage of generated messages does a human read before send today, and what was it at go-live?
4. What happened to your domain reputation and reply rates in the first ninety days?
5. How many meetings booked became qualified opportunities, and how many of those closed?
6. What did you have to build or hire internally to make it work?
7. If you were signing again today, what would you change in the contract?

The fifth question is the one that ends most deals. Meetings booked is a definition the vendor controls. Closed revenue is not.

The contract terms that follow

If you proceed, write the exit before the pilot. Set explicit kill criteria with a named owner and a stated remedy. Cap the initial term. Require the underlying event log, not a dashboard, so your count of sends, replies and meetings exists independently of the vendor's. And write a rollback plan for the domain and the CRM before the first send, using the rollback plan requirements.

The category earned its scepticism. That does not make every product in it unbuyable. It makes evidence from after the event the only evidence that counts.

Related: AI SDR unit economics · Kill criteria before you sign · Reversal ledger · Supplier shortlisting pass