The Last Mile of AI: Why the Future Is Bespoke, Not One-Size-Fits-All

For most of the last fifty years, software has been built for people sitting in offices. Traditional platforms are essentially static databases with an interface wrapped around them. They sit there, immovable, and they require humans to work around them: to feed them information, to enter data into them, to navigate their rigid, pre-defined logic. In this arrangement, the human being has quietly become the manual glue, spending the working day bridging the gaps between fragmented tools.

This operational tax is most painful in our mobile lives. Picture booking a simple meeting while you are out and about. You open a calendar app, tap, type, tap. Then you jump to email to recover the vague memory of a few keywords, search for the right address, copy it, and jump back to the calendar. It is a strangely dehumanizing experience. The user is forced to follow the software’s logic instead of the software following the user’s intent, holding broken context in their head and stitching it back together by hand.

The human as manual glue

We saw this clearly while building for mobile at Google. The world has only made the problem sharper since. More and more high-value work now happens in the real world rather than in a static office setup. It happens in coffee chats, in external introductions, in partnerships. And the unit of business is steadily shrinking, from the corporation to the team to the individual.

The data backs this up. According to the U.S. Census Bureau, the number of one-person businesses in the United States grew at an average of 2.7 percent a year between 2012 and 2023, more than double the 1.1 percent rate of firms with employees, and they now make up the large majority of all American businesses. We are entering the era of what I think of as the super-individual: a single operator who can run a global operation with a small fleet of agents doing the work. When the unit of work shrinks to the individual, the design of our tools has to flip. Instead of humans following software, software has to follow humans.

The last mile of AI

General assistants like ChatGPT and Gemini are a genuine leap. They let people achieve more, faster, across an extraordinary range of tasks. But there is a gap between a model that is generally capable and a tool that actually fits your life, and in AI that gap is the last mile. It is where most of the real value sits, and it is the part almost nobody has solved.

A general model arrives like a brilliant new hire on day one. It is fluent, fast, and widely read, but it does not know your clients, your preferences, the shorthand you use, or the way you like things done. Powerful is not the same as yours. A general model is, by design, the same for everyone who opens it. What people actually need is the opposite: something shaped to one person, that grows more theirs the longer they use it.

Historically, that last mile has been expensive. Being tailored to one person meant hiring an executive assistant or a chief of staff, or commissioning custom software, options reserved for large companies and the wealthy. So the rest of us settled for general, mass-produced tools and quietly trained ourselves to fit them. The cost of bespoke is the reason most people have never had it. What is changing now is that cost. As it falls, the thing people have always quietly wanted becomes possible at the scale of one. People are not really craving more generic capability. They are craving something that is theirs.

This is not only a matter of taste, and the evidence is sharper than it looks. The value of AI is not spread evenly across tasks. In a large field experiment run by Harvard Business School with Boston Consulting Group, 758 consultants using GPT-4 completed about 12 percent more tasks, roughly 25 percent faster, with higher quality, but only for work that sat inside what the researchers called the jagged technological frontier. On a task deliberately chosen to sit outside that frontier, the people using AI were about 19 percent less likely to reach the correct answer. Gains appear when the tool is fitted closely to the real contours of the work, and they fade, or even reverse, when it is not. Fit is not a finishing touch. It is where the value lives.

There are also two kinds of empowerment worth separating. One helps you do something you could not do before. The other frees you from the things your time has simply become too valuable to spend on. General models are very good at the first. The second is a question of fit, and fit has to be bespoke. This is the difference between a tool that makes you more capable and a tool that becomes an extension of you.

From destination to sidecar

If bespoke is the property, what is the form? Here it helps to borrow a term from software architecture. In distributed systems, the sidecar pattern attaches a helper process to a primary service to handle the cross-cutting work: logging, configuration, networking, observability. The core service gets to focus on its actual job while the sidecar quietly absorbs the operational complexity beside it.

I believe the same pattern is coming for people. Rather than a tool you visit, the next generation of personal software will behave like a sidecar bound to you, working proactively inside the apps you already live in. Not a destination. A companion that rides alongside, and one that is bespoke by its very nature, because it is attached to a single person and learns only them. A useful sidecar, as opposed to a chatbot you prompt from scratch every morning, would rest on three pillars.

Three pillars of a personal sidecar

Teachability and persistent memory. A sidecar is not a tool you re-explain every day. It is something you teach once. When you describe a preference or a workflow, it remembers, building a private store of individual memory that travels with you and captures the nuance of how you specifically work. This is the last mile, accumulating.

Observation and agency. A sidecar observes, thinks, and acts. It moves from knowing facts to doing tasks: filtering the noise, connecting the dots, drafting the reply, holding the thread of a project so you do not have to.

Exploration and transaction. Over time, this companion carries your context and your relationships across different pieces of software. It helps you find the right introduction, surface the right opportunity, and eventually even transact on your behalf.

From knowing to doing

The deeper technical shift underneath all of this is a move from retrieving to acting. Retrieval-augmented generation, the technique of grounding a model’s output in relevant documents it looks up on the fly, made today’s assistants good at knowing things. The next step is persistence and agency: companions designed not just to answer, but to execute.

The industry is already turning in this direction. Gartner expects that by the end of 2026, 40 percent of enterprise applications will include task-specific AI agents, up from less than 5 percent in 2025, a shift it describes as moving from assistants that respond to agents that act.

There is a quieter benefit here too, one that organizations tend to underrate: institutional memory. In a world of high turnover and fluid partnerships, an enormous amount of context walks out of the door every time someone leaves a role. A companion bound to the individual captures the expertise and the relationships that would otherwise be lost, and keeps that context intact and usable.

But this same abundance creates a trap. As agents let people take on more, people quietly do take on more: more to remember, more to act on, more to switch between, more to hold in their heads at once. There is a biological ceiling to how much of that a person can carry. The point of a sidecar is to sit beneath that ceiling and absorb the load, not to pile more on top of it.

Rehumanizing the work

The real aim is not to automate people away. It is to get them out of the software and back into the room. McKinsey’s research on knowledge work found that the average interaction worker spends around 28 percent of the week on email and nearly another 20 percent hunting for internal information or tracking down the right colleague, before you even add internal meetings and coordination. That is most of a week spent on operational tax rather than on the work itself.

If a companion can absorb that static admin layer, the time it returns does not have to flow back into the machine. It can go back into the coffee chats, the partnerships, and the judgment calls where the actual value gets created. This is not about replacing the human in the loop. It is about taking the human out of the parts of the loop that were never a good use of a human in the first place.

The future of AI probably is not a box you talk to, and it is not one model that serves everyone the same way. It is more likely a companion that is unmistakably yours: one that walks beside you, learns how you work, and quietly handles the parts you should not have to.

So the question worth sitting with is not how clever your AI can sound. It is simpler than that: how much of your week is spent being the glue, and what would you do with that time if you got it back?