Everyone wants to be AI-pilled. Most Companies Are Still Level 1

Ann Miura-Ko · tweet · 2026-05-01 source ↗ Reader 🖥️

Agent Governance (Internal)AI Confidence Theater Citizen Developer Enterprise AI Adoption & Readiness Organizational Legibility Risk Tiering Team Autonomy

‘AI-pilled’ is not a binary state but comes in six observable levels, defined by what AI can see, what it can do, who can extend it, and how the organization's structure has actually changed — and most companies, despite confident AI rhetoric, remain stuck at the lowest levels because they skip the hard, unglamorous work of making their own operations legible and structurally rewiring who does what before real autonomy becomes possible.

mixed · expository — A VC's field-observation essay (office visits to startups and a 1,500-person scaled company, Ramp) that builds a theoretical classification framework (six levels of organizational AI autonomy, by analogy to AV autonomy levels) and uses it practically as a diagnostic tool, with named 'hard tests' and 'common false positives' per level.

Structure

Part	Role
Office-visit observations contrasting tiny AI-native startups with a 1,500-person scaled company (Ramp)	evidence / motivation
The autonomous-vehicle levels-of-autonomy analogy, arguing against treating 'AI-pilled' as binary	framing device
The four diagnostic questions (what can AI see / do / who can extend it / how has the org changed)	method
The six-level framework (L0–L5), each with a hard test and a common false positive	central classification
Closing claim that AI-native companies are rebuilt around a new operating model, not an AI-assisted version of the old one (Steve Blank analogy)	conclusion

Key terms

AI-pilled
A loosely used, effectively binary popular label the author argues should be replaced by asking what level of organizational AI autonomy has actually been achieved, since companies vary independently along intensity and technical-capability axes.
legible (to a machine) Organizational Legibility
Whether an organization's work — knowledge, processes, systems of record — exists in a form AI can actually read and act on, versus living undocumented in people's heads, meetings, and unintegrated SaaS tools. The first of the four diagnostic questions ('what can AI see?').
AI as theater (L0)
The lowest level: a company where AI rhetoric (transformation speeches, announcements) is decoupled from any actual end-to-end process running on AI — 'Announcements ≠ adoption.'
fake autonomy (L5 false positive)
A system that claims self-driving, generative behavior but is actually only executing preconfigured rules or surfacing threshold-based alerts — glorified observability mistaken for genuine noticing/deciding/acting.

Propositions

explicit‘AI-pilled’ is commonly treated as binary, but companies actually differ along two independent axes — intensity (how deeply AI is embedded into daily work across the organization) and technical capability (what AI is actually allowed to see, do, and change) — so the diagnostic question should be 'what level of autonomy has the organization achieved,' not 'is this company AI-pilled.'conf: high
supports#2: The Unsexy Truth of AI Adoption · p1 — Both sources independently converge on the same empirical claim — that despite loud AI rhetoric, most real companies remain at the lowest maturity levels (L0/L1) — using parallel but independently constructed leveling frameworks.
explicitApparent AI-nativeness at a tiny startup is really a property of the founders/people, not the organization's structure; at a scaled company the bar is categorically higher because AI must become part of the company's systems and DNA rather than a few individuals' personality trait.conf: high
explicitOrganizational AI maturity decomposes into six levels (L0 theater, L1 personal productivity, L2 team workflow, L3 organizational infrastructure, L4 compounding operating system, L5 virtually self-driving), each defined consistently across four questions: what AI can see, what it can do, who can extend the system, and how the organization's structure has actually changed.conf: high
extends#2: The Unsexy Truth of AI Adoption · p9 — The other source identifies 'the company cannot describe its own workflows' as the real L1→L2 blocker; this thread's 'what can AI see?' question operationalizes exactly that same legibility blocker as the first of its four diagnostic axes, independently.
explicitEach level has a named, falsifiable 'hard test' and a named 'common false positive' — e.g. L0's false positive is a CEO's confident transformation speech with no process actually running end-to-end on AI; L1's is a meaningless usage statistic ('80% of employees use AI weekly'); L3's is 'a landfill of meeting transcripts and dashboards with no synthesis' (mistaking capture for legibility); L5's is 'fake autonomy' (preconfigured rules or threshold alerts mistaken for genuine noticing/deciding).conf: high
explicitcontestedNon-engineers authoring and shipping their own tools without filing an engineering ticket (not merely consuming AI-assisted workflows) becomes common only from L3/L4 onward, and the author treats this as a load-bearing signal of genuine organizational change rather than merely of tool adoption.conf: high
contradictsLooking back at the Future of Software Engineering unconference · p8 — This thread frames non-engineers authoring and shipping production tools (L4) as an unambiguous maturity signal; the other source warns the same pattern risks organizations drifting back into a 'they build it, you run it' model that dumps unsupported operational ownership onto citizen developers — a risk this thread does not acknowledge.
explicitcontestedFrom L4 onward, agents hold policy-driven decision authority within scoped domains (e.g. a security agent that detects, validates, and fixes an issue, opening a PR with human review only at the merge step), shifting human oversight from reviewing every action to reviewing scoped categories of action.conf: high
contradictsDesigning the AI-native engineering organization · p2 — That source argues organizations should NOT restructure their org charts in response to AI — changes should be operational, not structural; this thread's L3/L4 levels explicitly require a materially different org chart and an explicit structural choice about who does what as the defining signal of real change.
explicitL4's defining failure mode is 'agent sprawl' — a large number of brittle, uncoordinated automations that superficially resemble a compounding system but lack managed lifecycle, observability, and evaluation ('compaction discipline') — so sheer automation count is a false positive for compounding capability.conf: high
explicitAt L5 (not yet observed to exist, by the author's own explicit caveat), the defining boundary crossed is that the system acts on novel decisions because it notices it should, not because a human configured a policy for that specific situation — and institutional knowledge survives personnel transitions because it lives in the system rather than in any individual.conf: medium
explicitA company rarely sits at the same level across all four dimensions (see/do/extend/org-change) simultaneously; the resulting asymmetry across dimensions is itself diagnostic of where an organization's next intervention should focus.conf: high
explicitAn AI-native company is not an 'AI-assisted version' of an old company but an organization rebuilt around a new operating model, paralleling Steve Blank's claim that a startup is not a small version of a large company.conf: high

Arguments

Most companies' self-reported or publicly projected AI maturity systematically overstates their actual level, because the easy-to-produce signals of adoption (speeches, usage statistics, tool purchases, automation counts, dashboards) are decoupled from the hard-to-produce signals (an end-to-end autonomous process, non-engineer authorship, real structural change) that the framework's 'hard tests' actually require. strength: medium

Explicit premises:

Every level has a named common false positive — a way of convincingly appearing to have reached that level without having passed its hard test.
The author's own office visits found most companies, including a 1,500-person scaled company, still operating near the lower levels despite general AI enthusiasm.

Implicit premises:

Cheap, visible signals of AI adoption (talk, spend, usage percentages) are not merely imperfect but systematically biased upward relative to the harder, costlier signals that would demonstrate genuine capability.
An external, test-based assessment (rather than self-report) is necessary to calibrate real organizational AI maturity.

Inferences:

This implies most publicly circulated 'AI transformation' claims and survey statistics (e.g. usage-percentage figures) should be treated with structural skepticism unless paired with evidence against the framework's specific hard tests — a methodological critique of the broader AI-adoption discourse, not just a classification scheme.

Narratives

The autonomous-vehicle industry's multi-year debate over levels of autonomy — insisting that cruise control, lane-keeping, and driver assistance were each explicitly not full self-driving — is offered as a precedent for why precision in naming levels matters and prevents premature claims of a destination state.
→ Grounds the whole six-level framework in a recognizable prior case where graduated, enforced precision (rather than a feel-good binary) disciplined an industry's claims about its own progress.
Concrete per-level examples: a CSM handling 200 accounts instead of 50 with AI assistance (L2); a finance person building an automated contract reviewer and an account executive shipping a sales tool in under an hour, neither of them engineers (L4).
→ Anchors otherwise abstract level definitions in specific, vivid, single-case illustrations — though each is reported secondhand from office visits rather than independently verified.

Questions addressed

✅ Is 'AI-pilled' a meaningful binary category for describing a company?
No — it collapses two independent axes (intensity, technical capability) into a false binary; six levels, each with falsifiable tests, are more diagnostic.
✅ Where do most real companies actually sit on this framework?
Mostly at low levels (L0–L2), per the author's own office-visit observations, despite confident public AI rhetoric — the thread's title claim.
✅ Does a fully self-driving organization (L5) exist yet?
No, by the author's own explicit admission — L5 is a speculative extrapolation of the pattern, not an observed case.
⭕ How do you reliably distinguish real emergent autonomy from 'fake autonomy' (glorified observability dressed up as noticing and deciding)?
Named explicitly as 'the open challenge at this level' rather than solved.

Gaps

incompleteThe entire six-level framework rests on the author's informal office-visit method (a handful of companies, most recently Ramp) with no systematic sample, no inter-rater reliability on how a visitor assigns a level, and no account of how the named 'hard tests' would actually be administered or verified in practice. This is the same self-report/signal problem the framework itself diagnoses in the companies it criticizes — the author's own method is not held to its own standard of evidence.
Read next: Look for a more systematic, externally-audited version of a similar AI-maturity-level framework (e.g. with a defined assessment protocol) to compare against this one's informal, anecdotal basis.
incompleteL4 and L5 carry the most argumentative weight (the dramatic future-state claims) but are also the least evidenced: L4's examples (a finance person's contract reviewer, an AE's one-hour sales tool) are single, unverified anecdotes, and L5 is explicitly speculative ('I realize L5 does not exist yet'). No discussion of what observation would falsify an L4 or L5 claim, or how to distinguish a genuinely higher level from a well-marketed L3 capability.
Read next: Compare against corpus accounts of specific companies' agentic-SDLC or agent-governance rollouts that report failure rates or walked-back claims, to test whether the L4/L5 success anecdotes generalize.
uninformedThe framework treats 'non-engineers shipping production tools' (L4) as an unambiguous maturity signal without engaging the operational-ownership and support-burden risk this creates — who is on call when a finance person's contract-reviewer tool breaks, and who maintains it after that person moves on. This exact risk (citizen developers' work becoming unsupported production dependencies) is raised elsewhere in this corpus but is absent here.
Read next: See the corpus's treatment of citizen-developer operational-ownership risk (e.g. the 'they build it, you run it' dynamic) for the missing counterweight to this framework's unambiguously positive framing of non-engineer authorship.

Closes gaps elsewhere

#2: The Unsexy Truth of AI Adoption · g1 — That article names L2–L5 but only analyzes them behind a paywall, leaving L3 (trusting data), L4 (acting on signals), and L5 (self-improvement) unanalyzed for free readers. This thread supplies a free, fully worked-out six-level framework covering exactly that territory — with explicit 'what can AI see/do/who extends it/how has the org changed' breakdowns, hard tests, and named false positives for L3, L4, and L5.