The Last Mile Is the Whole Game

August 20, 2026

In the NFL, 57% of the entire 2016 season was decided by one score. That is 146 games where the difference between winning and losing was eight points or fewer. A team can spend all year getting roughly as strong, fast, and prepared as everyone else. Then the season turns on one block, one route, one decision made half a second faster. (NFL)

The NBA, the MLB, the Olympics… it's the same.

At the highest level, almost everyone has already traveled the first 99 yards. The whole game is the inch that remains.

Manufacturing is the same, except the inches are defects.

TSMC does not describe a new process as finished when it can produce a chip once. Its annual reports separate baseline process setup, yield learning, yield enhancement, reliability, equipment matching, defect detection, and volume production. In 2025, it spent 6.5% of revenue on R&D. Its 2nm process only became commercially meaningful when it entered high-volume manufacturing with good yield. (TSMC 2025 annual report)

The breakthrough is not "we made a 2nm chip." The breakthrough is making millions of them predictably, at a cost customers can bear.

The nines hide this difference. At 99.9% accuracy, a system makes 1,000 errors in one million decisions. At 99.99999%, it makes 0.1. The percentages look almost identical. Operationally, they are 10,000 times apart.

The last mile is the whole game.

AI collapses the first 99 yards of research, coding, design, and analysis. It can give everyone a plausible first answer in seconds.

That does not make the last yard less important.

It makes the demo less informative.

You can now describe a workflow and get something that mostly works in an afternoon. But the last few digits do not get cheap at the same rate. Most people can get to 95%. Some teams can get to 99%. Getting to 99.999% requires a disproportionate amount of testing, operational discipline, and judgment.

The exact number depends on the stakes. A 95% answer may be fine for a throwaway internal tool. It is not fine for a payment system, a medical workflow, or software that quietly writes to a company's institutional memory.

Agents are exploding the 95% layer. They can generate more code, features, and integrations than most teams could have produced before. But every feature adds states, dependencies, edge cases, and paths that can fail. The software works in the happy path while becoming harder to reason about underneath.

So the experienced human is needed less to type the first version and more to reduce the complexity it accumulated. They choose what not to build. They define the invariants. They identify the dangerous edge cases. They decide when the system is trustworthy enough to put in the path of a real person.

The last mile is not just customer-specific fit. It is the conversion of abundant, plausible functionality into something simple enough to trust.

There are two last miles. Only one is a moat.

The first is reach.

Be where the user is. Mobile, browser, Slack, email. Show up at the moment of need.

This matters. It is also where the agent already lives. If your strategy is "we will be on every surface," you are competing with the agent on its home court.

Reach is necessary. Reach is not a moat.

The second is fit.

Fit is the wiring into a customer's reality. Their brand, data, workflow, internal language, approval process, and weird historical decisions that nobody remembers but everyone still respects.

Fit is expensive per customer. It also accumulates.

A forward-deployed engineer spending eighteen months inside an enterprise builds fit. So does a design system that learns how my decks should look every time I make one.

This became concrete when Taiwan Mobile announced a tender offer to take control of Systex, one of Taiwan's largest IT-services and systems-integration companies.

Taiwan Mobile is not buying a foundation model. It is buying the ability to get technology into enterprise organizations: the relationships, implementation teams, account coverage, integration knowledge, and institutional trust that turn a capability into a working system.

This is the difference between reach and fit. Reach is having a network, channel, or customer list. Fit is knowing how to make a new capability survive inside the customer's actual workflow.

But the last mile is the whole game only when it compounds.

There is a nasty problem here.

The last mile often does not scale. Forward-deployed engineers serve one customer at a time. The part that defends is exactly the part that refuses normal software economics.

Meanwhile, the generic middle scales beautifully. And the generic middle is getting compressed.

What scales doesn't defend. What defends doesn't scale.

The way out is a self-accumulating last mile.

The customer uses the product, and the product quietly learns their taste, assets, data, language, and workflow. No army of people doing bespoke setup. No giant integration project every time.

Palantir's forward-deployed engineer model is the linear version of fit. A brand system that becomes more "yours" every time you make a deck is the scalable version. Leaving means abandoning the accumulated representation of how your work is supposed to look.

Gamma recently hired its first forward-deployed designer. That makes the choice feel very live.

Do they make the last mile manual, defensible but linear?

Or do they use people to discover what the product should learn on its own?

That is the question I would watch.

I almost stopped there. Then I realized I had missed the human problem.

I assumed gen AI would make software easier to use. In one sense it has. You can ask for things in natural language instead of learning menus.

In another sense, it made software harder.

A normal person can open Claude Code or Codex and have no idea what to do next. The tool can do anything. That is the problem. "Anything" is not an interface. It is a panic attack with a cursor.

The blank terminal is the new blank page.

The product between raw capability and a human who wants the work done is worth more now, not less. It narrows the choices and provides an opinionated path to a good outcome.

Call that product the wedge.

Gamma was never just a rendering tool. It was the wedge for presentations. An LLM could generate slide content, but Gamma removed a thousand degrees of freedom and gave people a path to something that looked good quickly.

The wedge and the last mile solve different problems. The wedge makes overwhelming intelligence usable. The last mile makes the result mine.

Put them together and you get the durable product: a productized consultant fed by compounding context.

I am the wrong person to judge these products.

My home base is a terminal. Most of my work happens in Claude Code and Codex, inside a bundle of files that has become a second brain. I like raw power. I like being close to the machine.

So when I see a product that wraps that power in a friendly, constrained experience, my reaction is predictable:

I don't need this. I would just do it raw.

That reaction is correct for me. It is also poison for my judgment as an investor.

The wedge serves the person who opens the tool, sees too much freedom, closes the laptop, gets coffee, and never comes back.

That person is the market. Not me.

The startup market will fill with products that look finished in a demo. The demo is now the first 99%.

I care about what happens after it meets a real company: the ugly data, the exceptions, the compliance review, the person who refuses to change their workflow, and the tenth time the system has to be right without supervision.

For investors, polish is becoming a weaker signal. Services revenue may actually be a positive one.

The question is what happens to the customer-specific work. If every deployment needs another army of people, the company has built a consultancy. But if each deployment teaches the product what to remember, automate, and refuse, the ugly work is creating the moat.

That is the line I care about now: does the last mile repeat, or does it compound?

Do not bet on raw power. It is becoming abundant.

Do not bet on reach. The agent already has it.

Bet on the wedge, fed by context that compounds. The winning product will narrow overwhelming intelligence into a usable path, then get more specific to the customer every time they use it.

The demo is now the first 99%.