The Implementation Era Begins

When I was at the Kennedy School about seven years ago, "the future of work" was everywhere. Study groups, projects, programming. A lot of smart people were thinking carefully about how AI was going to reshape things. The hard part was timing.

Yesterday, OpenAI and Anthropic announced parallel joint ventures with some of the biggest private equity firms on Wall Street. OpenAI's vehicle is called The Deployment Company, $10 billion, anchored by TPG. Anthropic's is unnamed, $1.5 billion, with Blackstone, Hellman & Friedman, and Goldman Sachs. Both ventures are going to embed engineers directly inside client companies and do the actual implementation work that those companies haven't been able to do for themselves. They'll start with the portfolio companies of the PE firms backing them.

This morning, Brian Armstrong announced 14% layoffs at Coinbase, citing AI and following the Jack Dorsey playbook we've covered before. Fewer layers, one-person teams with agents, leaders running 15+ direct reports. The org-chart redesign Dorsey laid out at Block is now showing up at the next public company.

Most of the conversation right now is fear. AI is going to take the jobs. The Coinbase headline is going to drive that story all week, and the next layoff announcement is going to drive it the week after. The fear isn't unreasonable. Some jobs will go away, and some already have.

What fewer people are saying is that AI works middle-to-middle, not end-to-end. That phrase comes from a Balaji Srinivasan post last summer that I think about almost every week.

"AI doesn't do it end-to-end. It does it middle-to-middle. The new bottlenecks are prompting and verifying."

A useful framework I've come back to is 10/80/10. The first ten percent of any meaningful task is human, where you set the goal, frame the problem, and decide what the deliverable should look like. The middle eighty percent is where the model earns its keep, drafting and structuring and analyzing and generating options. The last ten percent is human again, the part where you bring judgment, taste, and the conversation with the colleague who doesn't trust the output yet. The first mile and last mile are still human, and they probably always will be.

The fear narrative collapses both ends into the middle and concludes the whole job is gone. That's wrong. The middle is getting compressed, the ends are getting more valuable, and the people who can operate well at both ends are about to be in short supply. That shortage is exactly what Anthropic and OpenAI just put $11.5 billion behind.

The other thing not being said clearly enough is that the consulting pie is getting much bigger. McKinsey, Bain, and BCG aren't going out of business. They're going to end up with a smaller piece of a substantially larger market.

Those firms dominate high-end strategy consulting, which is where the work is concentrated and the margins are best. But strategy is the smaller pie. The much larger pie, by roughly an order of magnitude, is implementation work, the kind currently done by Accenture, Deloitte, the Big Four, and the Indian IT services giants. That's the pie the labs are going after, and it's the part of the market most exposed to what they're building.

Below the labs there's room for boutique AI implementation firms to service SMBs, which is the lane I'm in with Shannon Advisory. Bootstrapping a firm takes time. Take a big PE check and you can move faster, but you give up control. Speed is part of the calculus right now.

So in three years the market probably looks something like this. The labs and their PE-backed ventures take a meaningful share of enterprise implementation work. MBB keeps the high-end strategy work and a narrower slice of implementation that depends on deep industry expertise. Boutiques work the small-to-mid market. In-house teams at large companies handle whatever is core enough to keep internal.

The other reason this is going to take a long time is that implementation is human adoption, and human adoption is messy and slow. There's resistance, retraining, and people who simply don't want to change how they work. The model improvement curve and the workflow change curve have never been the same curve, and they aren't about to start.

The closest pattern is software implementation. Forty years in, consulting firms are still doing software implementations. AI implementation is going to look the same. Long-term, lumpy, full of failed pilots and partial rollouts and second attempts that finally take. The labs raising $11.5 billion this week is the early phase. The deployments that actually move the numbers will land in 2028 and 2029 and 2030.

For most knowledge workers, Monday morning didn't change because of yesterday's news. Most people didn't even notice. But the executives who are paying attention are starting to ask a different question. Not whether to adopt AI, but how much of the implementation expertise to build in-house and how much to outsource. That question is going to define the next decade of work.

Pick one piece of work you'd normally hire a consultant for. A market sizing, a competitive analysis, a board memo, a strategic options review. Before you call anyone, spend two hours doing it with Claude as a partner, not as a search tool. Set up the brief, push back on the first draft, verify the sources. You won't get a polished deliverable, but you'll get a much clearer sense of where the model ends and your judgment begins.

A subway commute map for New York. Anthony Castrio built a website where you can see everywhere reachable in under an hour on the NYC subway. It's both a heat map that recolors as you hover and a cartogram that warps the geography to reflect actual travel time, so distant neighborhoods on a fast express line pull closer and physically adjacent ones on slow lines push away. The animation in his tweet is the best part. The map breathes as you move the cursor.

One in ten AMC showings sells zero tickets. Which means you can effectively rent out a movie theater for the price of one seat. Riley Walz built a site that finds them. As he puts it: "Go enjoy your private theater." Useful? Maybe. Kind of silly? Yes. This is the kind of app that gets built when AI tools make it possible to ship something interesting in a weekend. I’m working on similar projects, although I’m nowhere near Riley’s skill level.