A hosting vendor the buyer had literally never heard of just won the first test slot. When SaaStr founder Jason Lemkin asked his company's AI agent where to move hosting for SaaStr AI Connect, the agent returned a full evaluation topped with one line: "For hosting the app, I'd pick Render." Render — a company Lemkin admits he couldn't have described — raised at a $1.5B valuation earlier this year, yet his brand awareness of it was zero. It still won.

The workflow shift is the story. The agent gathered requirements, compared five vendor categories, eliminated four, selected one, and queued up a test. Lemkin was involved in exactly two steps: asking the question and approving the test. Count the stages a human seller used to own. Then count the ones left.

What did the agent do before naming a vendor?

It wrote the requirements first, unprompted — because, as Lemkin puts it, "it knows the codebase better than I do." The list:

• An always-on web server for the site and API
• An always-on background worker running schedulers, refills, warm-ups and email builds
• Long-running jobs — email builds that take minutes
• Logs kept for days, a current pain point
• Zero-downtime deploys and a choice of machine sizes

"Most human buyers start from the brands they already know," Lemkin writes. The agent started from the workload. That inversion is the whole mechanism: the vendor no longer needs to reach the buyer; it needs to be legible to the agent doing the evaluation.

How did the five contenders score?

The agent's own verdicts, in its own words:

• Render (my pick): "Built for exactly this shape" — web service plus background worker plus scheduled jobs, health checks, zero-downtime deploys, longer log retention, machine sizes.
• Railway: Easy, pay-per-use, handles web plus worker — but "a bit less mature for heavy production."
• Fly.io: Fast and flexible, "more hands-on."
• AWS / Google Cloud: Maximum control, but "overkill for now."
• Vercel / Netlify: Struck through. "Not suitable [for this app]" — built for short web requests, not minutes-long jobs.

The Vercel elimination deserves scrutiny. Lemkin's team has used Vercel many times and he assumed it would be the default choice. The agent crossed it out on a single criterion: long tasks. Lemkin hasn't verified whether that's the full picture of Vercel's capabilities today. He notes it doesn't matter for this purchase — the agent believed it, and the agent made the shortlist. Measured fact: one criterion. Asserted belief: that the criterion disqualifies. The buyer accepted the assertion without audit.

What does this mean for the invisible lost deal?

Here is the sharpest consequence for vendors: when an agent filters you out on a stale or incomplete belief, you never find out. As Lemkin writes, "There's no lost deal in the CRM and no closed-lost reason to review." The funnel doesn't leak. It simply never renders your entry — pun fully operational at Render's valuation.

What should vendors do this week?

Lemkin's prescription, grounded in his own agent's output:

• Ask many agents your buyer's actual question. For hosting: "Where should I host an app with a web server, a background worker, and jobs that run for minutes?" Run it through Claude, ChatGPT, Gemini and coding agents. Record whether you appear and the exact sentence used to describe you.
• Find your strikethrough. Each rejected vendor lost on one short phrase: "less mature," "more hands-on," "overkill," "not suitable." If yours is wrong, fix the docs and public pages agents read. If it's right, it's a roadmap item.
• State plainly what shape of customer you're built for. Render won on "built for exactly this shape." The agent matched a workload description to a product description — a sentence a machine can parse.

What happens next?

Render gets tested against the two pain points the agent flagged: the worker process and log retention. If it fails, Railway is next on the agent's list. Lemkin says he'll report back — which makes this less a vendendor anecdote than a running experiment in who controls the shortlist when the buyer's agent, not the buyer, writes it.