The Rung the Machine Cannot Climb

There is a ladder that everybody who uses personal AI climbs, and almost nobody describes honestly. Crawl, Walk, Run, Fly. Four stages, three transitions. The interesting thing about it is not the stages. It is that two of the three transitions were handed to me by the software, and the one that was mine took longer than both of the others combined.

I want to be precise about what each rung is, because the usual version of this ladder is a maturity model built to sell something, and the rungs in it are all capability. That is not what moved me. What moved me, twice, was a product decision made by somebody I have never met. And what stalled me, once, was a sentence.

Figure 01 Four stages, three transitions, one of them mine

1. Crawl — you try it Initiated by curiosity Occasional, novelty-driven, impressive and useless. Every session begins from zero. You show it a thing, it says something clever about the thing, you close the tab.
2. Walk — you use it constantly, as a better search engine Transition initiated by the machine · memory The tool stops asking you who you are. The cost of starting drops to nothing, so you start ten times a day. Almost every request is still phrased as a question.
3. Run — you delegate Transition initiated by you · nothing ships this The request stops being “help me with X” and becomes “X.” Same model, same context, same tools. The only thing that changed is the verb, and the only person who can change it is you.
4. Fly — it works without being asked Transition initiated by the machine · scheduling The tool proposes standing work. A brief every weekday at six. A watch on a thing you care about. You stop issuing the instruction and start approving a cadence.

01Crawl to Walk was a feature release

Memory did it. Not reasoning, not context length, not tool use. Memory.

Before it, every session opened with a tax. I had to re-establish who I am, what I run, how I write, what I already decided, what I do not want to be asked. The tax was maybe ninety seconds, which sounds like nothing, and was in fact the whole thing — because ninety seconds of self-introduction is enough friction to make you not bother for a small task, and small tasks are where the volume lives.

When the tax went away, the volume arrived. And the part that surprised me is that it was not the answers that improved. It was the questions it stopped asking. An assistant that knows you does not open with three clarifying questions about your situation. It opens with the work. Every clarifying question is an admission that the arrangement has no continuity, and every one of them is a small invitation to give up and do it yourself.

The feature that changed my usage was not one that made the machine smarter. It was one that made it stop asking me things it should already know.

Note what I did not do here. I did not decide to use it more. I did not adopt a practice or read a guide or set an intention. A company shipped something, the friction fell below a threshold, and my behaviour changed underneath me. I noticed the change afterward.

02Walk to Run is a verb

Walk is a comfortable place to live, and most people using AI heavily are living there right now. It looks like adoption. The usage graph goes up and to the right. You are in the tool all day.

And nearly every request you make is a query.

Twenty years of search trained a reflex: when you need something, you phrase it as a lookup. How do I. What is the best way to. Can you help me with. The reflex is so deep that it survives contact with a system that does not need it. You get an excellent answer to the query, which confirms the reflex, which is exactly the trap — the reward for staying at Walk is high enough that you never feel the ceiling you are standing under.

The move to Run is one substitution. Not a workflow, not a tool, not a subscription tier. You stop saying help me and you say do it.

• Help me draft the follow-up emailDraft the follow-up and send me the draft
• Can you look into what changed in that policy?Find out what changed and report back
• What are my options here?Pick the best option, start it, tell me what you picked
• How would I clean up this spreadsheet?Clean up the spreadsheet

Read those pairs and notice how little is happening on the right-hand side. No new capability is being invoked. The model was always able to do the thing on the right. In most cases it was already doing the thing on the right, internally, and then throwing the result away to produce instructions for me instead.

03Why the verb is not cosmetic

It would be easy to file this under tone. It is not tone. The two sentences ask for different things, and they land on opposite sides of the only equation I care about.

Reality is Actual over Expectation. Help me is a request for guidance — and guidance is forecast, and forecast is Expectation. When I ask the machine how I should approach something, the deliverable is a better model of what is coming. That is the denominator. A perfect answer to a help me question raises my expectation and produces no Actual at all.

Do it asks for the numerator. Something happens. A file exists that did not exist. A message is drafted. A thing is found.

Help me is a request for a better expectation. Do it is a request for an actual. The model does not change between those two sentences. The equation does.

Which means you can spend a year at Walk, in the tool constantly, delighted by it, and never once move the number. High usage is not the metric. It never was. You can raise your expectation all day long and go home with nothing.

04The honest reason it is hard

I do not think people stay at Walk because they are lazy or unimaginative. I think they stay there because help me keeps the undo.

An advisory request is a delegation with the risk stripped out. You get the benefit of the machine’s work and you retain every decision, every keystroke, every chance to back out. Nothing it produces can embarrass you, because nothing it produces leaves the room. That is not irrational. It is just capped — and it is capped at precisely the level where the tool remains a search engine.

The evidence that this is the real gate, rather than trust in quality, is that it shows up everywhere the question has been asked, and it shows up as a preference for the weaker arrangement.

Figure 02 Four findings about the middle rung, and what each one actually measures

Figure	Finding	Source and sample	What it does and does not say
44 / 19	Offered the same model as Advisor, Coach or Delegate, 44% chose the Advisor and 19% the Delegate — and only the Delegate condition significantly raised collective surplus	Zhu, Thain, Tsai, Wexler and Qian, Choose Your Agent, arXiv 2602.12089, n=243, submitted Feb 2026, revised Jun 2026	The authors call it preference-performance misalignment. The model was superhuman within that negotiation setting, not in general. One study, few citations, no replication — it is the only direct evidence on this choice I could find, not a body of it.
32%	Share of AI users who have ever let AI act on their behalf at least once without a final sign-off	Menlo Ventures, 2026: The State of Consumer AI, fielded with Morning Consult Jul 2026, published Sep 16 2026; survey of 5,067 US adults, this figure based on the AI-user subset	Self-reported posture, not task-weighted. It measures whether the line has ever been crossed, not how often.
0.8%	Share of agent actions that appear to be irreversible; roughly 73% of tool calls appear to have a human in the loop in some way	Anthropic, Measuring AI agent autonomy in practice, Feb 18 2026, from its own Claude Code and API traffic	Anthropic hedges both figures with “appear to,” and says plainly that the latitude granted to models in practice lags behind what they can handle. It is one vendor’s own traffic, not agents in general.
79%	Share of enterprise IT leaders who have had to reverse an action taken by an AI agent	Propeller Insights for Kore.ai, Agent Productivity Index 2026, fielded May 2026, 400+ IT leaders at US firms of 2,000+ staff	Vendor-commissioned by a company that sells agent governance, and the sample was screened to ages 34–50 — an odd restriction for a claim about IT leadership. Directionally useful, not authoritative.

Put those together and the shape is the same in every one: people will take the machine’s output and will not authorise its acts. The gate is not is it good enough. The gate is can I take it back.

Which is useful, because it tells you how to climb the rung. You do not climb it by trusting more. You climb it by delegating the reversible things first — the draft, the search, the cleanup, the file that can be deleted — and noticing that the ceiling you were standing under was never made of competence.

05Run to Fly, and the machine takes over again

I am watching the next transition happen to me right now, and it is happening the same way the first one did.

The tool has started asking to work on a schedule. Not “what would you like me to do,” but a proposal: I can run this every weekday at six and have it waiting for you — want me to? That is not a delegation I issue. It is a delegation the software drafts and hands me to sign.

And look at the form it takes. It asks permission. It proposes a cadence rather than an act. It offers a thing I can cancel at any time, which means the vendor has handled the reversibility problem on my behalf — the standing task is the most undoable form of autonomy there is. The product is doing exactly what it did with memory: removing my reason not to, without requiring me to become a different kind of person first.

So for the second time, the rung gets climbed by somebody in a product meeting rather than by me.

06What the pattern predicts

A vendor can ship anything that lives inside the software. Memory, scheduling, tools, context, permissions, undo. They have shipped two of my three transitions and they will keep shipping.

They cannot ship a verb.

So the adoption curve should keep showing a plateau in the same place — enormous usage, overwhelmingly advisory — and the industry should keep responding to that plateau by shipping more capability, which is the one intervention that has no effect on it whatsoever. A better model does not change what you asked for. It makes the answer to help me better, which deepens the groove.

If that is right, the highest-leverage change available to anyone reading this costs nothing and requires no upgrade. It is a substitution in your own grammar, made maybe thirty times a day. It is the cheapest thing on this page and the only one nobody can do for you.

07The ledger

Claim 1 — Memory, not capability, was what moved ordinary users from occasional to constant use.

Already true

Persistent memory and project context shipped across the major assistants through 2025 and 2026 and are now default rather than opt-in. My own usage step-change followed the feature, not any model release.

What has to happen

Someone publishes per-user request volume before and after memory activation, separated from the model upgrades that arrived alongside it.

Where I am probably wrong

I cannot separate my own memory effect from the fact that the models genuinely improved in the same window. This is one person’s introspection about causation, which is the least reliable kind of evidence there is. If the volume jump was mostly quality and I have attributed it to continuity, the whole first rung is misdescribed.

Claim 2 — The Walk-to-Run transition is behavioural, not technical, and the gate is reversibility rather than quality.

Already true

People offered the same model in advisory and delegate form prefer the advisor by more than two to one, while the delegate is the condition that actually improves outcomes. Roughly two-thirds of AI users have never once allowed an action without sign-off.

What has to happen

Someone measures requests by grammar — imperative versus interrogative — against outcomes, on a real population. Until then this is mechanism plus circumstantial evidence.

Where I am probably wrong

Reversibility may be a rationalisation for ordinary risk aversion about quality, and the two are hard to separate when the same person holds both. It is also possible that most people’s work genuinely does not contain many delegable tasks, in which case the plateau is correct behaviour and I have mistaken good judgement for a stuck rung.

Claim 3 — The Run-to-Fly transition will be made by vendors, the same way Crawl-to-Walk was.

Already true

Scheduled and recurring task features are shipping now and arrive as a permission request from the tool rather than an instruction from the user. The form is a cadence you can cancel, which keeps the undo intact.

What has to happen

Standing tasks become common enough that a meaningful share of a user’s AI output arrives unrequested, and survives past the first month rather than being cancelled.

Where I am probably wrong

The strongest counterargument is to my whole thesis, not just this claim: if a vendor can propose a schedule, a vendor can propose a delegation — want me to just do this one? — and if that lands, the middle rung gets climbed by the software too and the human-behaviour argument collapses entirely. I think it lands less well, because a one-off act is less cancellable than a standing cadence and the proposal arrives at the exact moment the stakes are concrete. But I am not confident, and this is the way I expect to be proved wrong.

08What I left out

• Any measurement of how requests are phrased. The central claim of this essay — that most AI requests are interrogative rather than imperative, and that the ratio is the real adoption metric — has no published measurement behind it that I could find. Nobody counts verbs. I am arguing from mechanism and from my own logs, and I have not dressed that up as data.
• The usual enterprise pilot-failure statistics. The widely circulated figures about AI pilots never reaching production trace back to sources I could not verify against a primary methodology, so none of them are in here even though they would have supported the argument.
• Cross-vendor memory numbers. No assistant publishes usage before and after memory activation, so the first rung rests on my own behaviour rather than anyone’s telemetry.
• The institutional version of the problem. There is a much larger argument about what happens when the delegating party is an employee and the reversibility sits with an employer. That is a separate piece and it changes the answer.

One more thing worth saying plainly. The reason I find this ladder interesting is not that it explains a product category. It is that on three transitions, my behaviour was engineered twice and chosen once — and the one I chose is the one that produced anything. That ratio is worth sitting with regardless of what you think about AI.