Not subscribed? Sign up to get it in your inbox every week.

Last week, for the first time, an AI agent onboarded me to a product.

I was setting up a third-party ML training platform for a project. Anyone who has evaluated developer tooling knows the 2020 version of that chore: read the docs, watch the setup video, book a call with a solutions engineer, then burn an afternoon wiring credentials and config files until something runs.

This time I downloaded the platform's plugin, told my agent which ML workload I wanted to focus on, and watched it install itself into my repository. It read my codebase, picked sensible defaults, wired up the configuration, and asked me clarifying questions along the way. An hour later I was running training jobs.

That experience is the whole post: the next buyer, and the next user, of your product is going to be an AI agent.

We have seen lightweight versions of this already. OpenClaw normalized the single prompt that gets something installed, and half the skills on my machine arrived that way. But last week felt like a different category: a conversation instead of a script, with software that configured itself around my project rather than asking me to conform to it. Software is getting malleable. It can reshape itself to meet the user where they are, which means the effort of adoption, the most expensive friction in all of B2B, is heading toward zero. Three parts of the business sit on top of that friction, and all three are starting to move.

The earliest version of the new marketing channel already has a name: AEO (AI Engine Optimization), the practice of getting your product cited when someone asks an agent what tool to use. Discovery matters, and plenty is being written about it. The underpriced part comes after discovery, because the agent runs the evaluation too. It reads your docs, your pricing page, and your API reference, then decides in minutes whether your product can actually be stood up for its human's use case.

The pixel perfect landing page gets parsed for facts by a reader with no eyes. Therefore the assets that close this buyer look like clean documentation, working quickstarts/easy installation, and machine-readable pricing, and very little like a brand campaign.

Think about what a B2B proof of concept costs today. A solutions engineer, a kickoff call, a shared Slack channel, and three to six weeks of back and forth while the prospect's team hunts for integration time on their sprint calendar. Companies ration POCs precisely because each one burns scarce sales-engineering hours, which is why only "qualified" prospects get one.

An agent-led install collapses that cost. The prospect's agent can stand up a working pilot in an afternoon, on their data, in their environment, with nobody from your company in the loop. Therefore the gate flips: rather than rationing pilots to the prospects you qualify up front, you hand a working pilot to anyone curious and qualify on what the usage shows you. This allows much more complex software to offer a PLG motion.

Anthropic took this further than anyone with Project Deal, an experiment where Claude agents ran an entire marketplace on behalf of their humans. Sixty-nine employees each handed an agent a $100 budget and an intake interview covering what they wanted to buy and sell. The agents posted listings, made offers, countered, and closed 186 real deals worth over $4,000, with no human approval required inside the negotiations.

One detail from the writeup has stayed with me. Participants whose agents ran on a smaller model got measurably worse outcomes, and they didn't notice: the same broken folding bike sold for $38 when a Haiku agent negotiated the sale and $65 when an Opus agent did.

That was a Craigslist-style experiment with snowboards and ping pong balls. But the mechanics (intake interview, agent negotiates, human gets the outcome) port directly to a software purchase, and an agent that can already install and evaluate your product is most of the way to buying it.

A surprising number of B2B assumptions are quietly assumptions about human effort.

Pricing pages stay vague to force a sales conversation, demos sit behind forms because attention was the scarce thing worth capturing, trials get time-boxed because a human evaluator loses interest, and seats get sold because a human at a keyboard was the unit of usage. Every one of those made sense when a person did the evaluating, the installing, and the negotiating. Each reads differently when the counterparty is an agent with perfect patience, total recall of your docs, and can work 24/7.

Here's a test to run: Take a clean machine or an empty repository, point an agent at your own product, and tell it to get a working setup running, the way a prospect's agent would. Watch where it stalls. Every stall marks a place where your funnel still assumes a human. That list is your roadmap.

Till next time

Chris