The Second Station

You, as an enterprising AI robotics founder, want to reindustrialize America! So you sell a factory a smart robot for its slowest station, and the station goes from about a hundred parts an hour to five hundred. Great. The station after it does 125, max. That 25% gain might be worth enough to wedge your foot in the door. But the plant manager now wants a faster second station, probably somebody else’s machine with totally different tech. If you want to be measured on the line and paid like it, you may have to help him buy it, or rethink the flow. Alas.

Massive AI hopes ride on the pharma industry’s disruption, but it has the longest line of all. The time from a new target to a clinical candidate has perhaps accelerated, but then comes the clinic, which takes most of the decade and most of the money. I think a better molecule can raise the odds of success there! But site activation, enrollment, and regulatory review are the stations after discovery, and if the real goal is “approved drugs in patient hands,” those stations have to get going too.

A lot of point-solution SaaS of the 2010’s fit itself into whatever process the customer already had. Make one person better at their piece of it, and sell that person a seat. Few products changed the collaboration and process. Systems of record interred the results of that work in databases, and the databases were sticky enough. The conventional Silicon Valley wisdom of “do one thing well” naturally led vendors to conclude “the rest is your problem.”

Much angsty ink has been spilled about the “diffusion gap” between AI capability and business results. A trillion-dollar buildout assumes someone will pay for more capability. But companies will pay, at scale, only for returns. Ryan and Sean at Watney see this with big datacenter customers. Make one task faster and the customer immediately wants the next ones faster too, because the number they are watching is days until the site is serving tokens. The more of the line Watney moves, the bigger the pie for the customer, and the bigger the slice it can share.

AI is also expensive. A seat was cheap enough to buy on faith (or inertia). An agent’s bill is metered, and its big enough that the CFO reviews it. The mania of tokenmaxxing is already over.

A company is full of work waiting for other work. Make one team faster and the next team gets the extra work, which they did not ask for, and at the management meeting one team shows its productivity gains and the other gets asked why it is behind. These people are more than a capacity constraint. They also have to agree to change.

I underrated the agreeing bit for years.

Back at Greylock we once did a website review. A traditional venture partnership is a particular kind of many-headed CEO-less beast, and we spent more partner hours on a few paragraphs of homepage messaging than I would like to admit. Any one of us could have been more “productive” by sending whatever we had come up with that day. It looks like waste, but most people and organizations discover their true and shared preferences only through iteration and negotiation, and we had never thought hard enough about ours before. The second station was a dozen other partners. If that is true of a homepage, it is true of a bank’s loan process.

For AI companies, someone has to run that negotiation inside the customer. “Forward-deployed engineer” now seems to mean about five different jobs. Maybe the important one is an outcomes owner: a project manager who happens to be good at AI, coaxing other humans along with the P&L owner who can actually authorize the changes, and the security owner who can give you access.

Some labs think the agents will eventually do the coaxing themselves. Maybe. 1,000 toks per second will create a lot of drafts but it’s really hard to help a business that doesn’t know what it wants. Models have the IQ for this work and, judging by how good they are at persuading me, a fair amount of EQ now. They’ll need an OQ too, an organizational quotient: how the company works, who gets the rewards vs. the extra work, and why a manager can agree with you in the meeting and then not take action. Agents will have to win over Sally from finance, and that still feels a little beyond the jagged edges of general model intelligence today.

Until then, changing an organization takes authority. The least a vendor can do is align it in the contract. All pricing models have to break through inertia, but I think we will tend toward outcomes-based pricing in the end. I admit this stresses me out some. Companies that believe in outcomes pricing tend to have complicated pricing, with baselines, attribution rules, true-ups and caps, and as an investor I have an allergic reaction to complicated pricing. But seats were simple because the product was small. A contract that pays on how a bank closes loans will be about as complicated as closing loans. Some companies are already there. Sierra, which increasingly charges for longer-horizon, business-level outcomes (for example, customer conversion or retention), passed $400 million in ARR (opens in new tab) in October. Cognition guarantees (opens in new tab) its value: if Devin’s measured output is worth less than the customer paid, it funds the gap, up to $10 million.

I think this (along with software itself becoming cheap to build) will lead to a more consolidated market structure. You can buy a hundred software subscriptions, but you probably cannot reorganize the business with a hundred strategic partners, which could push AI companies to consolidate even if all hundred have good products.

The most extreme form of this consolidation is less reliance on vendors. Incumbents with the institutional will to rebuild are using AI in-house and keeping the gains, with fewer vendors. Citi (opens in new tab), with 14,000 engineers on Devin, said on its Q3 earnings call that bringing work in-house has cut its technology vendor spend 30% in a year, while accelerating its releases.

Or (given enough capital) you buy the whole business. That is Sequence, Ithaca, Thrive Holdings, Long Lake: the right to change the operation and the right to keep the gain, and the burden of running it. There are signs of life – Sequence (opens in new tab) took BankSouth’s loan origination from nine stages to three and roughly halved application-to-closing. There’s not much value in being the “SaaS category leader” on one of the steps that disappeared.

If the return you’re promising depends on the second station, then make the second station your business, too.