We Analyzed 160 of Our Own AI Consulting Sales Calls. Here Is Why AI Projects Really Stall.

Between March and July 2026 we recorded 398 internal and external meetings. We isolated the 160 that were sales or discovery conversations with real prospects, 66 distinct companies in total, spanning e-commerce, real estate, construction, recruiting, professional services and more.

Then we did something most firms never do. We stopped trusting our own memory of the calls and checked every deal against the email record: signed agreements, invoices, explicit declines, or silence.

The outcomes across those 66 companies: 11 closed. 2 said no directly. 17 went silent. 21 were still in active conversation at the time of analysis. The remaining 15 had not reached a clear outcome by then. Everything below comes from comparing the calls that closed with the calls that died.

Across 160 conversations we could not find a single deal that died on price level. Not one prospect said it was too expensive and walked.

What killed deals was ambiguity. Prospects who heard too many options in one call stopped replying. One buyer told us directly that our offer had a lot of moving parts. Another, five weeks into evaluating a proposal, said the honest thing out loud: it is like, a lot. She went silent two weeks later.

When price objections did surface, they were really shape objections. Small-business owners did not want a smaller invoice. They wanted a different thing: a built tool instead of a report, one number instead of a menu.

This one embarrassed us. 9 of our 11 closed deals looked undecided at the end of the call. Meanwhile the single most enthusiastic prospect in the entire dataset, a founder who said on the call that she was ready and needed this to be her life, declined by email nine days later.

Why? On that call, the close was deferred. Instead of naming a price and booking the decision, we promised to send a pricing document. The document arrived. The heat died. A hot buyer plus a homework assignment equals a cold buyer.

If you sell services and you grade your pipeline by how calls felt, your forecast is fiction. Grade it by what is booked on a calendar.

Of our 17 silent deals, nearly every one ended its last call the same way: send it over and we will discuss internally. A passive exit. No date. No decision point.

Worse: in about half of those deals, the materials we promised were never actually sent. The prospect did not ghost us. We ghosted them first.

The volume difference is stark. Closed deals averaged 16 email touches from first call to signature. Ghosted deals averaged 6. The market did not reward cleverness. It rewarded showing up eleven more times.

The deals that closed shared three behaviors, and none of them were scripts.

First, a structured diagnosis of the buyer's own operations came before any signature. Our cleanest closes followed a deep, methodical mapping of the client's actual workflows. One client said the walkthrough of their own operations was what sealed the deal. Another told us to stop explaining and deliver something. We did, and they signed.

Second, a booked next step, every time. No closed deal ever ended a call with a vague follow-up. There was always a date.

Third, depth about the buyer, not the catalog. Going deep on the prospect's operations built trust. Going deep on our own service menu destroyed it.

This was the single most useful discovery in the dataset. The behavior that closed deals was diagnosis-shaped: understand the company's real operations first, prove it back to them concretely, then recommend. We have since productized that exact behavior as our front door. It is called the Pathway Diagnosis.

Our data matches what the research community keeps finding. Failure is now the default experience of buying AI, which means every serious AI vendor should assume the buyer across the table has already been burned once.

The Pathway Diagnosis is Zero-to-Agent's paid diagnostic: a fixed-price decision process that determines which AI operating model fits a company, what to build first, what not to automate, and realistic investment bands. The fee starts at $7,500 and is credited toward the first approved implementation within 60 days.

It exists because of the data above. The behavior that preceded our strongest engagements was a structured diagnosis of the client's own operations. The behavior that preceded ghosting was a menu of options and a passive exit. So the diagnosis became the product: one offer, one number, visible pricing, and a booked next step at the end of every call.

For a company evaluating AI adoption, the practical takeaway is the sequence, not the vendor. Diagnose the operation before funding the build. RAND's failure analysis names wrong-problem selection as a leading cause of AI project failure, which is precisely what a diagnosis step exists to prevent.

If you are evaluating any AI partner, ask them three questions. What will I see working before I sign anything? What exactly does it cost, said out loud? And when is our next dated step? If a vendor cannot answer those in one sentence each, you have found the ambiguity that kills most of these projects.

Why do most AI projects fail? Research from RAND (2024) finds over 80% of AI projects fail, about double the rate of ordinary IT projects. Our analysis of 160 sales conversations adds a commercial reason: deals and projects stall on unclear offers and unowned follow-through far more often than on technology or price.

What is a Pathway Diagnosis? A Pathway Diagnosis is Zero-to-Agent's paid diagnostic. It determines which AI operating model fits a company, what to build first, what not to automate, and realistic investment bands. The fee starts at $7,500 and is credited toward the first approved implementation within 60 days.

Is a paid AI diagnostic worth it before an implementation? In our 160-call dataset, the strongest client engagements were preceded by a structured diagnosis of the client's own operations, while deals presented with a menu of options tended to stall. A fixed-price diagnostic that is credited toward implementation removes the risk of paying twice.

How should a company restart after a failed AI project? Start by naming why the last attempt failed: surprise costs, tools nobody adopted, or automation that acted without approval. Each failure type needs a different safeguard. A diagnostic that identifies the failure mode before recommending anything new prevents a second stall.

Methodology: 398 recorded meetings (March to July 2026), 160 qualifying sales conversations, 66 companies. Outcomes verified against email records, not call impressions. Client identities anonymized; aggregate figures only. Questions about the data: hello@zerotoagent.com.