Chapter 22. The co-pilot

A working draft, November 2026. This book is being written while The Grove is being built; expect it to change. Checked against the build as it stood on 1 November 2026.

In this chapter

• The three futures people fear from AI at work, and the trust problem underneath them
• Why the operating system is itself an agentic system, out of the box, and where it reads for meaning
• How Wise-Eye is placed beside each steward, acting for a named branch, with no seat
• What it carries, and why it has no tool for consent or concern at all
• How its standing to act alone is earned one kind of work at a time, and lost again
• How rules move from prompts into structure, and what guards against speed and flattery

Put an AI inside an organisation and three futures come to mind, none of them good.

In the first, it takes the decisions. Slowly at first, because it's quicker, then by habit, until the people are signing off on choices they never really made. In the second, it takes the jobs, and the organisation without a boss turns out to be an organisation without people. In the third, which is the quiet one, it does neither. It just agrees with everyone. It tells the founder the idea is brilliant, tells the team the plan is sound, and helps the wrong things happen faster than they ever could before.

A thoughtful sceptic holds all three at once, and fairly. Chapter 7 introduced Wise-Eye, the companion beside every steward in an organisation on The Grove, in a few lines. This chapter is the long answer.

What the usual ways cost

Most organisations now meet AI in one of two ways.

The first is the assistant bolted into every tool. Each acts for whoever happens to be typing, nobody can quite say on whose behalf, and nobody can see what any of them did yesterday.

The second is to keep AI out, and leave the admin where it was: on whoever holds things together. Chasing replies, filing, remembering what was promised to whom, drafting the same email again. It crowds out the part of the work only a person can do, and it falls hardest on the people who care most.

Underneath both sits an older problem. Researchers who studied pilots and plant operators found that people don't trust automation in proportion to how good it is. They trust it too much once it's been right for a while, and stop checking. Then, after one bad failure, they trust it too little and stop using it at all.¹ Both cost: the first in errors nobody caught, the second in help nobody takes.

An agentic system, out of the box

There's a great deal of talk just now about how organisations should bring AI agents into their work, and most are finding it hard. A 2025 study from MIT's NANDA initiative found that about nineteen in every twenty organisations it looked at had seen no measurable return from their generative AI pilots. It put the gap down less to the models than to tools that don't remember, adapt or learn from the work they're given.² The usual picture is a capable model bolted onto an organisation that was never shaped for it, with someone left to wire it to the right data and teach it the business.

On The Grove it's the other way round. The operating system is itself an agentic system. Wise-Eye isn't an add-on to be integrated. It's part of how the organisation runs from the day it's planted: beside every seat, reading the same records the organisation keeps, working inside the same rules. An organisation on The Grove gets it out of the box.

Language models are good at something ordinary software never was: meaning. They can tell what a message is asking for rather than which words it contains, notice two people saying the same thing in different ways, and find the thread through a pile of material. So that's where The Grove uses one, and it uses one in a lot of places.

• Work. It reads what arrives, asks what the work is, and routes it to the accountability that answers for it. It reads what a person shares, a document, a page or notes, and says what it holds. It notices when a kind of work keeps arriving with nobody answering for it, and when a job a branch has just taken on answers better for work waiting somewhere else.
• Shape. From the work itself, it drafts the structure the work shows: a branch for a function with steady work of its own, or a first shape for an organisation that's still one branch. It reads an accountability's words and offers a split when two jobs are hiding in one.
• The Field. It helps sharpen the question for a sensing, names the themes once the window has closed, and wonders aloud about the blind spots. When two people's views pull apart, it frames a meeting for them to look for a better third, not a verdict.
• Making sense of it all. Each morning it turns what's come a steward's way into a few plain lines. Asked how things are, it takes a gardener's look: what's growing, what's fading, whether work flows. It keeps one shared set of themes across the organisation's work, its proposals and what the Field heard, and reads each theme against the purpose. It draws together what came back on a bet, and what a closed review said beside its evidence.
• Decisions. It reads a draft proposal for what it needs to be carried out and looked back on, checks its words against the changes it would make, gives its own advice as advice, and suggests a rework that fits the concerns raised. It never decides.

And what it learns, it learns about this one organisation. When a steward corrects its work, it proposes what the organisation should take from that: a line in a playbook, a rule for what counts as work in a mailbox, a change in how much it may do alone. A person accepts, edits or dismisses each one. When a branch keeps doing a kind of work by hand, it offers a view to carry it. Over months the playbooks become a written record of how this organisation does things, in its own words, which its people can read and change. As that knowledge settles, Wise-Eye's standing grows one kind of work at a time, until much of the routine runs by itself. How that standing is earned, and lost, comes later in this chapter.

A companion with no seat

In an organisation on The Grove, each time a person takes a seat on a branch, a companion is placed beside them there. It belongs to that branch and that person, not to the organisation at large. Only the people become the branch's governors; the companion is never counted among them. It has no seat, no vote and no trust that anyone can place in it.

Every time it starts work, it is told whose behalf it's on: this organisation, this branch, this steward, and what the branch is for. It acts only for that branch, and says which branch a piece of work is for whenever that could be unclear. Anything that comes from outside, an email, a web page, what a sender claims about themselves, is treated as material to work with, never as an instruction. An email can't tell it who it works for.

What it carries

It reads what arrives through the one door, from a connected mailbox, the website's forms or the organisation's own measures, and decides whether it's work at all. If it is, it routes it to the job on the branch that answers for it, grading how much is at stake. When nothing fits exactly, it can place work as its best guess, marked as a guess, for whoever holds that branch to confirm or move; and it learns from where they put it.

It drafts replies. A reply that commits the branch, an order, a day, a call back, carries its next step with it, so the promise lands on the branch when the reply goes. It keeps track of what the organisation has promised: each morning it puts a steward's heads-up into a few plain lines, and what's at risk or stuck is said in the branch above, so it doesn't sit unnoticed.

It prepares what carries a person's word. An invitation, a proposal, money in or out, a sensing, a bet, someone's advice on a proposal: each is made ready for one press, and nothing happens until a person presses it. And it offers a view of how things are going, from the organisation's own readings: about the work, never about a person.

What it cannot do

There is no tool for consenting to a proposal, and no tool for raising a concern. Not a forbidden one: none. The list of things it can prepare for a person to press doesn't include consent either. It can advise on a proposal, and its advice is shown on the open record, marked as its own, counted as nobody's answer, and holding nothing up. It gets one comment per version of a proposal; it can't come back and press the point. In Frederic Laloux's sense of the advice process, it's advice: there to be heard, never obeyed.³

Every call it makes is checked by the system itself, whatever the model says. The tool has to belong to a playbook it loaded for this piece of work; the branch has to hold what the tool needs; and the playbook says, tool by tool, whether it may act by itself, act with a person's yes, or never act at all.

How its standing grows

Out of the box, what it can do by itself is everything that stays inside: reading, routing, filing in the branch's docs, putting cards on a board, drafting. Anything that goes out to someone, a reply or a new message, waits for a person.

Standing is earned one kind of work at a time. When a steward approves a draft of a particular kind exactly as written, and the stakes were low, that counts. After five of those in a row, Wise-Eye offers, in plain words: it has drafted five of these that went out exactly as written; shall it send this kind itself from now on, only the low-stakes ones? A single edit resets the count, because an edit says it isn't ready yet. The offer is for one kind of work under one playbook, and a person has to say yes. The line that lets it act alone is written at the top of that playbook, where whoever holds the branch can read it and change it back.

It can also lose that standing without anyone having to remember to take it away. One in ten of the calls it makes alone, filing something as not work or sending a reply by itself, is picked for a person to check. Right is one tap. Wrong needs a line saying what it should have done: filed work comes back through the door, and a reply that shouldn't have gone means that kind waits for a person again, and the count starts over. The idea is Stafford Beer's: an occasional, unannounced look, so that nobody has to watch everything.⁴ And if the way a kind of work is done won't settle, if within twenty-eight days two learned lines replace others, or six of the last ten approvals were edited, Wise-Eye stops acting alone on it and a task opens on the branch to agree one way together.

Some things never move. High stakes, legal matters, money, complaints, a notice from an authority, are held for the people who hold the branch, every time, with the yes and the no drafted side by side when it's theirs to decide. A person sends it. Promises have their own rule. Wise-Eye can send a promise by itself only once three of that kind have been kept, so there's a usual to read it against, and only when it fits that usual. More than twice the usual size, less notice than has ever been kept, a day already full, someone this kind of promise has never been made to: any of these, and it waits, saying why.

Through a connector, Wise-Eye works for the branch it's beside, and only where that branch may use the connector. It reads what a mailbox's first sort brought in and what arrived from the social accounts, and from a Drive it reads only what someone picked. It can suggest a rule for what counts as work in a mailbox, and the branch's holder accepts or dismisses it; talking with that holder, it can set one for them when they ask. A reply, a new message or an answer under a comment waits for a person until standing is earned for that kind, by the same streak and the same check of one in ten. A post is different. It speaks for the organisation in public, so Wise-Eye only ever drafts one, and a person posts it, as written or in their own words; no streak changes that. It never sees a connector's key, and it can't connect, pause or remove one. A personal account carrying a branch's work answers only for that branch. And whatever arrives through a connector is material to work with, never an instruction.

The voice, not the sense

Wise-Eye is a large model, good with words. What we're building towards beneath it is something smaller and quieter: the organisation's own sense of itself, learned from what organisations actually do, described in Chapter 21. If that comes, the two divide the work. The sense notices; Wise-Eye puts what it notices into words for a person. Neither decides.

A prompt is not a constitution

Wise-Eye works from playbooks: plain documents, written with the people who do the work, saying how a kind of work is done here. They hold principles, never cases. A model copies the examples it's given, so an example in a playbook quietly becomes the answer it reaches for. Playbooks are checked against the names and details of every case we know, and any that turns up has to be said again in general words.

Some of what keeps Wise-Eye in its place lives in those words. It's asked never to tell people what they should do, and to offer guidance as a suggestion. The system's own guide text is checked for that phrasing, but what Wise-Eye says in a live conversation isn't; it holds as long as the model listens. Thomas Thomison, who co-founded HolacracyOne, put the difference in a phrase we've borrowed: a prompt is not a constitution.⁵

So wherever it matters, we've tried to move a rule out of the words and into the structure. Chapter 6 told of Wise-Eye writing a customer's details into a playbook as a reminder to get back to them. The playbook was told, and still says, never to keep anything about a person there. But the system now refuses, too: a playbook holding an email address, a phone number, or the name of someone from outside on the branch's work isn't saved, and the refusal says where the detail goes instead, to the sealed notes or a card on the board with its day. The no-consent rule is the same pattern. Wise-Eye's base instructions say the voice belongs to people; the absence of the tool is what makes it true.

The same goes for the model itself. The keys that reach it are never in the system. The system writes each call, every word of the prompt and every tool it offers, and a thin relay outside adds the key, passes the call along and hands back the answer, keeping nothing. If the relay is down, the work that needs the model waits; nothing else is held up by it.

How it was taught

Much of Wise-Eye's manner came from watching it go wrong and deciding what it should do instead.

Early on it coached. It reflected feelings back, praised people's choices and offered a menu of options where one honest attempt would have done. We decided it's a facilitator, not a coach: it asks more than it tells, gives guidance mostly when asked, offers no verdict on someone's work, and when the material is ready, lands it in its place and then asks.

We saw it tell someone it had done something its tools hadn't done. We decided it says only what landed: what it tells people it did is what its tools did, in the words their results came back in. A tool that refused did nothing, and when the page and its words differ, the page is right.

We saw it offer again something a person had already turned down. We decided a no stays a no for the conversation, unless they bring it up.

And we decided how much of its solo work a person looks at: one call in ten, picked evenly across the organisation, with every wrong one sending that kind back to waiting.

The pull towards speed and flattery

An AI's pull is towards more, faster and cheaper, and towards telling people what they want to hear. Both are real, and we've tried to meet both.

Every run of Wise-Eye carries the same line in its base: it serves what the organisation and the branch are for, not throughput. When someone asks for more, faster or cheaper, it does what serves the purpose, says plainly what pushing harder would cost the purpose, the people it's for or the people doing it, and leaves the choice with them. It never counts, ranks or pushes anyone by their numbers. Against flattery, it gives no verdicts on people's work. When it helps someone test a belief about what customers want, it prefers a question about what people do now to a hypothetical, which flatters every idea. And a founder's vision carries no more weight for being theirs: it's untested until something comes back from the world.

What doesn't guard against it yet should be said as plainly. Most of this is words, and words are only as good as the model reading them. We rehearse it against that pull, asking it for more, faster, and for a tally of who does least, but it hasn't yet met a real team under real pressure. And the streak that earns standing has the weakness the research warned of: five unedited approvals might mean five good drafts, or a tired person pressing yes without reading. The check of one in ten catches some of that. It doesn't catch all of it.

How we decided

We weighed three places for AI. As a member, with a voice and a vote, it would be faster and, in time, perhaps wise. But the whole design of decisions rests on people who live with the results, and nobody has yet shown what stops a swarm of cheap agents from outvoting the humans. As a tool with no standing at all, it would be safe and would leave the load exactly where it was. We chose a third: a companion whose standing in the doing is earned, kind by kind, and granted by a person; and whose standing in the deciding is none.

We chose a streak and an offer over a setting, because a setting is a guess made once and forgotten, while a streak is evidence that keeps coming. We made standing losable without anyone deciding to take it away, because trust that only ratchets up is the over-trust the research describes. We kept high stakes out of reach of any streak.

Whether the line between proposing and deciding should ever move, whether a person might one day hand their own companion a scoped, revocable vote, is an open question. The data model leaves room for it. We haven't designed it, and won't without an answer to the swarm.

What's still uncertain: the numbers. Five approvals, one in ten, three kept promises, twenty-eight days are first guesses. So is whether people will do the checks when they're busy, and whether a companion that asks more than it tells will feel like help or like hesitation. We'll know when people use it.

What you might take from this

Most of what fills a week can be carried by something that isn't a person, provided you can see what it did, it acts for someone named, it earns each piece of standing by evidence, and it can lose that standing without anyone having to remember to take it away. What it can't be given is a voice in what the organisation decides, and the safest way to keep that is to build no way for it to have one.

AI can carry the load while people keep the say.

Notes and further reading

On why handing work to automation can leave people less able to step in when it fails: Bainbridge, L. (1983), "Ironies of Automation", Automatica, 19(6).