Where AI Belongs

There is no Chief Spellchecker Officer. There is no spellcheck budget, no spellcheck strategy deck, no spellcheck center of excellence. Spellcheck is everywhere and nobody thinks about it — which is exactly where AI is going, and exactly why building a chatbot is aiming too low.

Chapter 1 argued that AI is a must and that you’re already qualified to deliver it. This chapter answers the question that comes next and that the rest of the book quietly depends on: deliver it where? Point it at what?

It matters because the technical chapters ahead — patterns, retrieval, agents — are easy to read as instructions for building a chatbot. They aren’t, and if you finish this book having built a very good chatbot, you will have aimed at the wrong target. So before any of the “how,” this chapter fixes the “where,” and it does so on the back of a decade of research rather than anyone’s opinion.

The spellchecker end-state

Start with a thought experiment. Spellcheck is one of the most widely deployed pieces of language technology ever built. It sits in every text field you touch, quietly correcting billions of words a day. And notice what doesn’t exist around it: no executive owns it, no budget line funds it, no consultant is retained to advise on spellcheck maturity. You would notice spellcheck only if it vanished. It is embedded, ambient, invisible — a capability, not a product.

That is not a quirk of spellcheck. It is the fate of a whole category of technology that economists call general-purpose technologies — the class that includes electricity, the internal combustion engine, and the internet. They share three properties: they’re pervasive, used as an input across nearly every sector; they keep improving; and they make everything downstream more productive. The argument, made most prominently by Erik Brynjolfsson and colleagues at Stanford, is that AI now belongs in that category — that it is transitioning from a bespoke set of tools into a general-purpose technology from which applications are assembled. Andrew Ng compressed the same idea into four words that stuck: AI is the new electricity.

Sit with what that implies, because it’s the whole orientation of this chapter. If AI is a general-purpose technology, then intelligence becomes a commodity input, wired into everything, mostly out of sight. The end-state is not a room you visit to talk to an AI. It is AI woven into every system and every workflow and every text field — the spellchecker, generalized. And there’s a second, deeper layer that answers a question you may already be asking: AI doesn’t stop at being embedded in products. As it matures, it starts building and operating them — writing the systems (the AI-assisted development of Chapter 4) and running them (agentic operations). AI moves from feature, to substrate, to the thing that builds and runs the substrate. Intelligence as ambient infrastructure.

If that is the destination, then what you build now should point toward it — toward embedding, not visiting. Which is exactly where most organizations go wrong.

The chatbot trap

Ask almost anyone to picture “using AI” and they picture a chatbot: a box you type into that types back. It’s the most visible form, the first demo everyone builds, the thing that made the technology feel real. It is also, in a corporation, usually the lowest-value form — and mistaking it for the destination is the single most common way to aim too low.

The reason is structural. A separate chatbot sits beside the work. To use it, a person has to stop what they’re doing, switch windows, phrase a question, wait, and copy the answer back into the actual task. Every one of those steps is friction, and the value leaks out in the handoffs. The chatbot makes the human do the integration work by hand, one query at a time.

The grounded alternative is now stated plainly in the enterprise research: value shows up when AI sits where the work already happens — embedded in the system the employee is already in, reducing handoffs, surfacing the right information, drafting, flagging exceptions, and suggesting the next step, all in the flow. The example that makes it concrete is a humble one. Instead of a chatbot you can ask about invoices, you build AI that reads each incoming invoice, checks it against the approved record, and routes only the mismatches to a human. Nobody “uses” it. It’s simply how invoices work now. That is the spellchecker pattern, applied to a business process: embedded, ambient, and valuable precisely because no one has to go to it.

None of this makes the chatbot useless. It is excellent training wheels — the fastest way to prototype, to feel a model’s behavior, to learn the patterns in the chapters ahead. That’s why the early experiments in this book are chatbot-shaped; you start there because it’s the shortest path to touching the technology. The trap is only in mistaking the training wheels for the destination. So carry this reframe through every build chapter that follows: don’t ask “how do I build a chatbot about X.” Ask “how do I embed intelligence into the workflow that does X.”

The jagged frontier: where AI works, and where it hurts

Embed AI into workflows, then — but into which parts? Not all of a workflow should touch a model, and the boundary is emphatically not where intuition places it.

Here the evidence is unusually good. In a large preregistered field experiment, researchers from Harvard Business School and BCG, together with colleagues from MIT Sloan and Wharton, had 758 management consultants perform realistic tasks with and without a frontier model. Inside a certain boundary the effect was transformative: consultants completed more tasks, far faster, at markedly higher quality — on the order of a quarter faster and forty percent better. But step outside that boundary and the sign flipped. On tasks beyond it, consultants using AI were meaningfully less likely to reach the correct answer than those working without it.

The researchers named the boundary the jagged technological frontier: AI excels at some tasks and fails at others of similar apparent difficulty, sometimes within the very same workflow, and the edge juts inward and outward in ways human intuition does not predict. You cannot reason “this task is harder, so it’s harder for AI.” The frontier is jagged, not a smooth gradient — which is precisely what makes it dangerous, because you can’t feel where it is.

And the failure mode at the edge is the worst kind. Outside the frontier, AI does not fail loudly or obviously. It produces confident, polished, well-structured work that happens to be wrong — faster than a human would produce the right answer. This is the silent failure of Chapter 5 wearing a business suit: the output looks like success, which is exactly why it slips through.

So the method — grounded in MIT Sloan’s more recent work on how AI actually lands in organizations — is not “add AI to this job.” It is: decompose the workflow into its component tasks, and assess each task independently against the frontier. Where does AI clearly help? Where does it clearly hurt? Where must a human hold the pen? Then design the sequencing and the handoffs between human and machine deliberately, because the research is clear that AI’s biggest impact comes not from any task in isolation but from how tasks are managed, grouped, and handed off. You are not bolting a tool onto a role. You are redrawing a workflow, task by task.

One more grounded nuance worth carrying: the lift is generally largest for less-experienced people. A Stanford and MIT field study of thousands of customer-support agents found average productivity gains around fourteen percent, rising to roughly thirty-five percent for the newest and lowest-skilled workers. AI tends to raise the floor more than the ceiling — an “inverse skill bias” — and that, too, shapes where in an organization it belongs.

Within the frontier, the same consulting study surfaced two distinct and durable patterns for how a human and an AI actually divide a task. They’re worth naming, because they are the atomic shapes of AI-in-a-workflow.

The Centaur pattern divides and delegates. Like the mythical half-human, half-horse, there’s a clean seam: the person does the parts of the work on their side of the frontier and hands whole tasks across to the AI on the other side. You delegate the literature search and keep the judgment call; the boundary between the two is sharp and deliberate.

The Cyborg pattern interweaves. Here the human and the machine blend within a task, in a tight loop — prompt, correct, extend, redirect, iterate — with no clean dividing line. The work is co-produced move by move.

Neither is correct in the abstract; they suit different tasks. Centaur fits work that decomposes cleanly, where you can name which sub-tasks belong to which party. Cyborg fits work where human and model sharpen each other continuously — drafting, coding, analysis. Most real designs combine them. What the pair gives you is vocabulary for how a human and a system share a task, once the jagged frontier has told you which tasks they should share at all.

The embedding ladder

Put the whole trajectory on a ladder, because it’s how a single workflow matures over time.

Assistant is the first rung: a chatbot beside the work. Fastest to build, easiest to adopt, lowest value, and — as we’ve seen — a trap if you stop there. Genuinely useful for open-ended help and for learning, but not a destination.

Embedded is the second rung: intelligence inside the system where the work happens, surfacing information, drafting, flagging, and suggesting the next step in the flow, so that no one has to “go to” it. This is where most durable enterprise value lives today. It is the spellchecker rung.

Agentic is the third rung: the system runs the workflow itself — planning, calling tools, executing — with a human on the loop rather than in it. This is the frontier that corporations are climbing toward now, and it’s exactly where the reliability discipline of Chapter 6’s “workflows versus agents” and the operating chapters becomes non-negotiable, because an unreliable agent embedded in a core process fails at scale.

The rungs are not better or worse in the absolute; they’re a maturity path. You often start at rung one to learn, but you should always know which rung a given workflow ought to reach, and build toward it deliberately. For spelling, we long ago climbed to rung two-and-a-half — embedded, largely autonomous, with a human able to override. Chapters 6 through 9 are, in large part, how you climb that ladder for your own workflows without falling off it.

Co-invention: what separates the 5% from the 95%

Now the finding that should reframe how you think about every AI project you ever scope, and that closes the loop with Chapter 1’s uncomfortable statistics.

Stanford’s Digital Economy Lab, in its Enterprise AI Playbook, and McKinsey’s enterprise survey data land on the same conclusion, and it is blunt: AI success is not about deploying better models. It is about redesigning how the organization works. The companies actually capturing financial return are roughly three times more likely to have fundamentally redesigned their workflows around AI — on the order of fifty-five percent of high performers versus twenty percent of everyone else — only about a fifth of organizations have done it at all, and workflow redesign is the single factor most correlated with realizing value.

This is Brynjolfsson’s productivity J-curve, and it is the same story electricity told a century ago. Factories did not get more productive by bolting electric motors onto layouts designed around a central steam shaft. Productivity came only once they redesigned the factory around distributed power — a re-invention that took decades. The value of a general-purpose technology is unlocked by the complementary re-invention around it, not by the technology alone. Economists call that complementary work co-invention, and it is the expensive, unglamorous, organizationally hard part. AI’s J-curve is expected to be shorter than electricity’s — partly because AI can help implement itself — but the shape is identical: you invest in the redesign before the returns appear, which is why the returns feel slow and why the impatient give up.

The implication is uncomfortable and clarifying at once. Bolting AI onto an unchanged workflow is the default failure mode — it is the mechanism by which organizations join the ninety-five percent that see no measurable return. Value is not the model. Value is the model plus the complementary investment: the redesigned process, the new handoffs, the retrained people, the better data. And the model, as Chapter 5 showed, is now the cheap part. The co-invention is the work, and the willingness to do it is precisely what the five percent have and the ninety-five percent don’t.

Which turns the whole question inside out. The right question was never “where can I add AI?” It is “which workflow am I willing to redesign around AI?” — because the redesign is where the value lives, and asking it that way filters out the bolt-on projects before they waste a year.

Choosing your first workflow

So here is the on-ramp, and it is deliberately concrete, because it feeds the first real thing you’ll build. Picking where to start is a short, disciplined method:

Choose a workflow, not a task and not a chatbot — you’re redesigning a flow of work, not adding a feature. Decompose it into tasks and map each to the jagged frontier: AI clearly helps, AI clearly hurts, or human-only. Then pick a workflow where the helpful gap is large and measurable, and — this is the part people skip — where you actually have the authority to change the process, because co-invention you’re not allowed to do is co-invention that won’t happen. Decide the collaboration pattern (Centaur or Cyborg) and the rung you’re aiming for (embedded, ideally, not assistant). And define what “delivered” and “better” mean before you build, so you can tell whether the redesign worked.

That is the leapfrog move at the level of the organization. Don’t wait for a corporate AI strategy to descend from on high. Pick one workflow you’re allowed to rebuild, and rebuild it around embedded intelligence. That single redesigned workflow — not a model, not a chatbot, not a strategy deck — is the atomic unit of AI value. Deliver one, and you’ve done the thing the ninety-five percent never manage.