Working With the Machines

What actually happens after your organization adopts AI — and the human plan that makes it work.

Working With the Machines

What actually happens after your organization adopts AI — and the human plan that makes it work.

Somewhere in your organization this week, someone did something like this: they opened an AI assistant, pasted in a chunk of a client proposal, and asked for a summary. It took two minutes. The result was good. Nobody stopped them, because nobody knew whether they should.

And someone else did the opposite. They wanted to use AI for a task everyone agrees it would be great at, but they had no idea whether it was allowed, or who to ask. So they didn’t. That lost hour is harder to see, and it repeats silently every day.

If some version of this happened in your company this week, you are not behind. You are normal. But you should understand what it means: this is not a failure of your people, and it is not a failure of your policy. The problem is the space between them. A policy sitting in a document is not the same thing as an operating system people run. What is missing is not more rules. It is the machinery that turns rules into daily practice — and the shared language that lets people use it without asking permission for every task.

The capable new colleague

The clearest way I know to explain AI to a boardroom is this: think of it as a remarkably capable new colleague. This colleague works at extraordinary speed, never gets tired, is available to everyone, and is improving month by month. But this colleague has two properties every organization must manage.

First, they do not know your rules, your clients, or your red lines until someone teaches them what applies where. Second — and this is the one people skip — they cannot be held accountable. An AI cannot answer to your board, your customers, or a regulator. Only a person can.

Everything I have built for the last few years follows from those two facts. Because AI is capable and everywhere, you cannot ban it and you cannot ignore it. Because it cannot be accountable, every consequential use of AI needs a named human standing behind it. That is the whole game.

Delegation is a contract, not a badge

• Execute — the machine does exactly what it was told, and a person checks the work.

• Recommend — the machine proposes; a named person decides.

• Decide within bounds — the machine may act inside limits someone set, and the trail shows it stayed there.

• Decide and act, reviewed after — the machine acts, and its work is inspected on a stated cadence.

• Set the bounds — humans only. Deciding what may ever be delegated, and answering for that decision, is never itself delegated.

The industry treats delegation to AI like a maturity score: as the model gets smarter, you hand it more. I think that framing is exactly wrong, and it is where most AI governance quietly fails.

Delegation is not a measure of how smart the machine has become. It is a contract for a consequential decision. The question is never "is the AI good enough to do this?" The question is: who answers for this if it goes wrong, what exactly did they agree to, and can they stop it?

That is the core of the framework I call Decision-Scoped Delegation, and it is laid out in my book, The Constitution of the Frontier. The contract has a structure you can actually write down — the Delegation Ladder:

Read that last rung twice, because it is the one that makes the whole ladder safe. The authority to decide what machines may decide stays with people. Permanently. That is not a limitation on AI — it is the condition for trusting it.

The Constitution itself is fifteen principles across five articles, governed by two meta-rules: every principle must cost something (if following it is free, it is decoration), and no principle may rest on what machines can’t do (because that list shrinks every year). Principles built any other way expire quietly, and everyone pretends they didn’t.

Five questions every organization must answer

1. What is AI for here? A stated purpose people can apply without asking permission for every task.

1. Which decisions belong to people? Clear lines, situation by situation, not blanket slogans.

1. How does AI-assisted work stay visible? A record of what runs, what it produced, who reviewed it.

1. Who answers for each use? A named human owner — never "the committee" in the abstract.

1. How does this keep working next quarter? A repeatable path from new idea to decision to release, run by your own people.

Strip away the frameworks, and an organization working well with AI has simply answered five plain questions:

Notice what is not on that list: which model to buy, how to prompt better, how many agents to deploy. Those are the questions the industry loves because they are easy to demo. The five above are the ones that decide whether your AI use survives contact with your customers, your lawyers, and your own employees’ judgment.

You cannot automate a process you have not decided how to run

Here is the pattern I keep walking into: an organization tries to jump straight to AI automation, and it fails in the same predictable way every time — not because the technology was weak, but because nobody had decided what the process was supposed to be, who owned it, or what "good" meant. The machines worked fine. The decisions were missing.

So the work happens in stages, and the order is not optional. First the foundation: the rules people follow, the inventory of what is actually running, the review process your own committee runs — about six weeks of deliberate work, not a six-month program. Then capability: training your people in the shared decision language, because an operating system is only as good as the people running it. Then activation: the first ninety days of approvals closing, owners acting, the cadence beginning. Only then automation, one proven workflow at a time, inside the boundaries someone already set. And finally operations: AI governed like any other part of the business — on a schedule, with records.

The stage order is the whole trick. Foundation before capability, capability before automation, decisions before speed.

Governance your machines can read

There is one more piece, and it is the one people find surprising. The rules we build are not just documents for people. Every deliverable ships in a form your AI assistants can read and follow too: step-by-step instruction files an agent can execute — how to take in a new AI request, how to check it against the policy, how to record the decision — and structured, versioned records behind them. When an organization chooses automation later, its agents already know the rules.

And baked into every one of those instruction files is the rule that makes me comfortable writing them at all: the agents prepare and record; the people decide. No agent approves its own work. No agent approves an AI use. Anything that touches a red line escalates to a named human, immediately.

The machines do the work. People answer for it.

That sentence is the design principle behind everything above — the book, the engagement, the course, the instruction files. It is also, I would argue, the only stable position for the next decade. Not because people are better at the work than the machines — we mostly aren’t anymore. Because answering is the one part of work the machines cannot do, and the one part we cannot give away.

Released capacity is not realized value

There is one more idea worth putting in plain sight, because it is where every AI program quietly goes to die: the economics. When an enterprise promises that AI will deliver a thirty percent productivity improvement, here is what is actually being said. To the organization: more output from the same people, for the same money. To the worker: accomplish more, without necessarily sharing in the additional value you create. Both statements are true at once, and the distance between them is the deepest risk in AI adoption — deeper than any ROI calculation.

Here is the accounting problem in one line: released capacity is not realized value. If AI saves an organization ten thousand employee hours, those hours are not money yet. They become money only when someone converts them — into additional production, increased revenue, avoided hiring, reduced costs. Or into something the workers keep: returned time, better pay, better jobs. Until someone decides the conversion, the savings exist on a productivity dashboard, and a dashboard is not a strategy. Worse, an unconverted surplus creates a quiet incentive to simply expect more from the same people. That is how a productivity program becomes an extraction program without anyone voting for it.

So when we measure the return on AI work, we keep two ledgers. Enterprise ROI: returns created, costs avoided, revenue gained — counted only when converted, with evidence. Human ROI: hours returned and where they actually went, changes in pay and job quality, mobility and economic security. Neither number substitutes for the other. An organization that tracks only the first ledger is not measuring productivity; it is measuring how much more it can ask of the same people for the same money, and calling that a return.

There is a bigger question underneath, and it deserves honesty: if machines gradually take over the production of economic value, can wages tied to human labor remain the primary way prosperity gets distributed? That question belongs to society, and it is considerably bigger than any enterprise ROI model. But the measurement belongs to every organization that adopts these tools. You cannot answer who benefited if nobody kept the ledger.

Where this is real

All of this exists as more than writing. The Constitution of the Frontier and the Decision-Scoped Delegation paper are published and citable. The AI Governance and Enablement Foundation turns the framework into a six-week engagement organizations can actually buy — fixed scope, editable deliverables, their own people operating the result. The Delegation Ladder course teaches the language to a whole workforce. A value review measures what the work returned on both ledgers — the enterprise’s and the people’s. And the whole system ships machine-readable, versioned, and owned by the client, not rented from a vendor.

If you take one thing from all of it, take this: the future of work is not humans versus machines, and it is not humans handing the keys to machines. It is organizations that finally answered three questions, in writing, with evidence: who decides, who can work, and who benefits. Three questions — a whole collaboration in one breath. I have come to think of them as the collaboration trinity. Answer those, and you can let the machines do the rest.

Ed Fassio is the founder of ByteBrain and the author of The Constitution of the Frontier. This article draws on the Decision-Scoped Delegation paper and the Agentic Futures body of work.

Take the Readiness Scorecard Explore the offerings

Working With The Machines — the slide deck

The Human Playbook for Intentional AI: fourteen slides that turn the article into something you can present. A practical guide to the Agentic Frontier.

Get the deck (PDF)