Matt Anderson · CincyDeliver 2026
Companion to the talk deck — see also how AI was used to build the talk.
This is a thinking pattern, not a framework. It comes out of years inside transformations — the ones that took, and the ones that didn't — not a proprietary methodology. AI transformation lives in a complex domain (see Part 3.2), and complex domains can't be methodized; a fixed method would repeat the exact mistake that turned Agile into compliance theater. So use what follows as a way of thinking you point at your own context and expand — not a checklist to comply with. It's deliberately paired with one falsifiable claim (the kill-switch, Part 6) so it stays honest: humble in form, rigorous in substance.
Agile transformation was supposed to produce adaptive, customer-centric organizations. For most, it didn't — not because the concepts were wrong, but because organizations kept the old operating model: top-down decisions, annual budgets, waterfall controls, functional silos. We got the vocabulary of Agile without the operating model.
The same failure modes are now showing up in AI adoption, one-to-one. Adoption is near-universal; strategic value isn't. The reason isn't the model — about 10% of AI's value comes from the algorithm, 20% from tech and data, and 70% from changing how people work (BCG). Ninety percent of the prize is everything that isn't the model.
Getting there is a climb with three horizons and two locked gates:
You don't drift between horizons — you pass a gate. Gate 1 is data (it blocks H1→H2). Gate 2 is trust and autonomy (it blocks H2→H3). Most organizations are stuck against one of these and have mistaken it for a technology problem. It isn't. Both gates are locked by four properties of your organization — structure, complexity-handling, culture, and adaptability — and every one of them has a direct Agile precedent.
The bottom line: you can adopt AI without changing your organization. You can't transform with AI without transforming your organization. Both are valid games. Only one creates lasting structural advantage.
"The difference was never the AI model. It was always the organization." — Stanford Digital Economy Lab, 51 deployments across 41 companies. Read it with its selection frame: that's a study of winners, so it tells you what the successes shared, not how often anyone succeeds.
The failure modes that neutered Agile are showing up in AI adoption one-to-one:
| Agile failure mode | AI equivalent |
|---|---|
| Measuring velocity, not value | Measuring AI cost savings, not outcome impact |
| Keeping decisions at the top | Human approval for every agent action |
| Rigid audit & compliance | Book-length AI governance before the first deploy |
| Silos blocking team autonomy | Agents designed around org boundaries |
| "Doing Agile" without changing culture | "Using AI" without rethinking the work |
The talk carries three ordinary processes the whole way up the climb — because the pattern is easier to see in a process you already run than in an abstraction. Find yours in the column you're honestly in:
| What Agile did to it | H1 — AI bolted on | H2 — past the data gate | H3 — past the trust-and-autonomy gate | |
|---|---|---|---|---|
| The Meeting | Renamed it a stand-up, 15-minute timebox — same ten people reporting status to the same manager. The ritual changed; the power dynamic didn't. | AI transcribes the minutes. Waste is documented more efficiently, not eliminated. | A 15-minute decision, not a status round | Exception-only sense-making — it convenes when something is genuinely novel |
| The Budget | Put Jira on the budget and called it planning. Annual cycles and executive horse-trading — untouched. | A faster forecast. Produced faster, not better. | Real-time scenario modelling | Continuous allocation inside guardrails; the budget review becomes an exception, not a quarterly ritual |
| Incident Response | Added a Slack channel to the war room and called it DevOps. The escalation hierarchy stayed intact. | A shiny dashboard on the same war room | Correlated detection across domains | Autonomous playbooks; agents designed around the incident, running in parallel |
The Meeting is relatable, the Budget is high-stakes, the Incident is fast — but they make the identical climb, and they get stuck at the identical two gates.
Borrow Agile's single most useful practice — the retrospective — and point it at your own readiness instead of your last sprint. Score yourselves honestly on four dimensions. Each one predicts a ceiling.
| Dimension | Low readiness | Medium readiness | High readiness |
|---|---|---|---|
| Structural | Functional silos, project funding, centralized decisions | Partial value-stream alignment, mixed funding | Value-stream organized, outcome funding, distributed authority |
| Cultural | Pathological — information hoarded, blame culture | Bureaucratic — information through channels, rule-bound | Generative — information flows, authority distributed |
| Complexity | All work treated as complicated, ROI required for everything | Some experimentation tolerated, innovation labs exist | Complexity literacy embedded, probe-sense-respond is normal |
| AQ | Reverts under pressure, can't unlearn, rigid mental models | Moderate resilience, some unlearning, context-dependent flexibility | High persistence, active unlearning, fluid mode-switching |
Run these like a real retrospective — with the team, with honesty:
H1 AI is forgiving of data debt — it works inside existing boundaries. H2 breaks when data is siloed, inconsistent, or inaccessible, because H2 needs cross-functional synthesis, which needs cross-functional data. Your data architecture mirrors your org chart (Conway's Law); if Sales, Finance, and Operations each have their own "truth," no agent can synthesize across them.
What has to be true to pass Gate 1:
Why it's an org-design problem, not a data-lake problem: you solve it by changing who owns data, who can access it, and what "truth" means across the organization. Organizations that couldn't achieve a shared backlog will struggle to achieve shared data. Same pattern, higher stakes.
On the name. This gate was called Governance through 2026-07-19. It was renamed because "governance" reads as a narrow policy exercise, and this gate is process, culture and decision rights — the whole question of who is allowed to act without asking. "Governance" is still the right word for the mechanism below (the three-model spectrum); it was the wrong word for the gate.
H2 runs fine on traditional governance — humans review AI recommendations and make the call. H3 breaks when governance is approval-based instead of guardrail-based. H3 needs autonomous action within boundaries, and that's a different governance model.
The governance spectrum:
| Model | Description | AI ceiling | Agile parallel |
|---|---|---|---|
| Gate-based | Every action requires explicit approval before proceeding | H1 | Waterfall with Agile vocabulary |
| Review-based | Actions proceed but are reviewed at checkpoints | H2 | Sprint reviews with stakeholder sign-off |
| Guardrail-based | Actions proceed autonomously within defined boundaries; exceptions escalate | H3 | Truly self-organizing teams with clear mission and constraints |
What has to be true to pass Gate 2:
Conway's Law. The 1968 paper concludes that organizations "are constrained to produce designs which are copies of the communication structures of these organizations." (The more familiar phrasing — "any organization that designs a system … will inevitably produce a design whose structure is a copy of the organization's communication structure" — is Conway's own restatement, added as an author's note in 2001, not the 1968 text. Both are his; only one is from the paper.) Your AI agents will mirror your org chart unless you deliberately design otherwise. This is the root cause behind both gates — data architecture (Gate 1) and decision rights (Gate 2) both mirror the hierarchy.
Team Topologies applied to agent design:
| Team topology | Human team role | AI agent pattern |
|---|---|---|
| Stream-aligned | Delivers value to a specific customer segment | Primary agent aligned to a value stream or customer outcome |
| Platform | Self-service capabilities to stream-aligned teams | Shared AI infrastructure (data access, model serving, guardrail enforcement) |
| Enabling | Helps teams adopt new capabilities | AI coaching/optimization agents that improve other agents |
| Complicated-subsystem | Manages technically complex components | Specialized agents for deep-expertise domains (regulatory compliance, risk modeling) |
Cognitive load is the design constraint — a single "budget AI" doing forecasting + scenario modeling + allocation + compliance + comms is overloaded. Split it the way Team Topologies splits human teams. (The Team Topologies 2nd ed., 2025, explicitly extends the model to "humans and AI.")
The Inverse Conway Maneuver (Skelton & Pais): deliberately design the organization to produce the architecture you want, instead of letting the existing org chart dictate it. For AI: (1) define the desired agent architecture, (2) align teams, data access, and decision authority to match, (3) let the agents become the connective tissue across the boundaries humans find hardest to cross.
Most organizations operate as if all work is Complicated (analyzable, expert-solvable). But agentic AI is most powerful in the Complex and Chaotic (unordered) domains — probe-sense-respond, act-sense-respond. An agent that can only execute predefined workflows in predictable environments is an automation script with better marketing. The Agile parallel is almost word-for-word: organizations that couldn't tolerate "we'll discover the solution through iteration" won't tolerate "the agents will discover the optimal response through experimentation."
Ron Westrum's typology (backed by years of DORA research — ~23,000 individual survey responses over four years, not "thousands of organizations") sorts cultures by how they handle information and authority:
The same property that lets a generative culture trust a team to deploy without a change board is the one that lets it trust an agent to act within guardrails. Culture isn't a soft variable here. It's the gate.
The Laloux ↔ Westrum ↔ Horizon mapping (for those who think in Laloux's Reinventing Organizations colors — same territory, different vocabulary; Westrum carries the evidence base):
Read this as an interpretive mapping, not a set of source-defined equivalences. Westrum's three types classify how an organization behaves with information and authority (cooperation, failure response, how novelty is handled) — not its developmental worldview. So they can't separate Amber from Orange from Green as cleanly as Laloux's stages do. The distinction that carries the load: Amber is rule-bound; Orange is performance-varnished but still bureaucratic in how information actually moves. That's why classic Orange enterprises talk performance yet stall at the H2 trust-and-autonomy gate.
| Laloux stage | Metaphor | Westrum culture | Information & authority | AI ceiling |
|---|---|---|---|---|
| Red (Impulsive) | Wolf pack | Pathological | Power through fear; information hoarded as a weapon | H1 |
| Amber (Conformist) | Army | Bureaucratic | Formal roles, rule channels, centralized escalation | H1; bounded procedural H2 only |
| Orange (Achievement) | Machine | Usually Bureaucratic at enterprise scale; can hold Generative islands | Performance targets, MBOs, silos — innovation under managerial control | H2 — stalls at cross-functional governance unless information flow turns genuinely generative |
| Green (Pluralistic) | Family | Generative-leaning, not cleanly Generative | Trust, empowerment, values; formal authority often still pyramidal | H2; early H3 only if authority/governance are structurally redesigned |
| Teal (Evolutionary) | Living organism | Generative | Self-management; information and authority flow to the work within guardrails | H3 |
We've run the self-management experiment before — mostly it didn't take:
The honest counter: agents aren't people — no egos, they don't quit when confused. So maybe agents succeed where human self-management failed. The ceiling still holds, because it was never about the workers — it's about whether the surrounding organization can stomach distributed authority, fuzzy accountability, and a real loss of control. That's a property of the organization, not the agent.
The Adaptability Quotient adds the capability to adapt, not just the willingness. Natalie Fratto's three markers are below; the organizational tells are mine — her AQ measures individuals, and the extension to organizations is my argument, not her finding.
The talk landed Step 1, the IC on-ramp, and Start Monday. Here's the complete sequence.
Step 1 — Make the choice explicit. Before any AI strategy, answer honestly: are we trying to optimize our current operating model, or transform it? Both are valid; they need different investments. The worst outcome is pursuing H3 ambitions with an H1 organization — or capping yourself at H1 when you're capable of H3.
If you carry nothing else out of this, carry these. The whole argument compresses to three reflex-corrections — each one a sentence you'll hear in a meeting, and the correction that goes with it.
| When you hear… | The reflex |
|---|---|
| "Which model should we use?" | The model is 10%. Your organization is the ceiling. When AI underdelivers, audit the org — not the algorithm. |
| "We've adopted AI." | Adoption is the on-ramp, not the destination. Bank the quick wins and the learning — then choose: optimize, or transform? |
| "AI will get us past our dysfunction." | AI mirrors your organization and amplifies it. Silos, hoarded information, decision bottlenecks: encoded and sped up. Fix the org for real transformation. |
That's the pattern applied once — not a new framework. Everything below is the same three reflexes worked out in detail.
Step 2 — Build an AI portfolio across horizons. Most of your effort into H1 (it pays the bills and builds literacy), a meaningful slice into H2 (decision quality, cycle time), and a deliberate seed into H3. The H3 seed is small but critical — it's your learning investment, the probe in probe-sense-respond. Without it, you never build the muscle for transformation. (The exact split is a heuristic, not a prescription — argue with it.)
Step 3 — Redesign governance from approvals to guardrails. The single highest-leverage organizational change for AI transformation.
| From | To |
|---|---|
| "Can I do this?" (permission-seeking) | "Is this within my boundaries?" (guardrail-checking) |
| Approval before action | Audit after action |
| Risk avoidance | Risk management within boundaries |
| Compliance as gate | Compliance as continuous monitoring |
Start with one process. Define the guardrails. Let the team (or agent) operate inside them. Review outcomes, not approvals. Expand on evidence.
Step 4 — Do the Inverse Conway move. Pick one value stream. Redesign the team structure around the customer outcome, not the org chart. Then design the AI agents to match. That's your proof point for structural change.
Step 5 — Make adaptability a first-class deliverable. Track the AQ tells as real metrics: reversion rate under pressure (grit), "what did we stop doing?" (unlearning), and whether leaders can name optimize-vs-experiment by context (flexibility).
Step 6 — Start Monday. Replace one recurring meeting with an outcome-oriented agent loop. Not a pilot. Not a proof of concept. A real experiment:
Then do the same with one budget-allocation decision and one incident-response scenario. Three experiments, three horizons of organizational complexity, real data on your readiness — lived experience, not a scorecard.
If you don't own the budget: everything above is written for someone with a budget line and an org chart to redraw. Most of us aren't that someone — we're the engineers, QA leads, and scrum masters who run the processes. You don't own the budget. You do own a recurring meeting. That's not a consolation prize — it's the leverage point. Every locked gate shows up first as a small local process someone already runs. Start with the one you control. That's Step 6.
The Optimizer. Bureaucratic culture — efficiency and control, hierarchical decisions. Strategy: H1 focus, automate everything automatable. Outcome: significant cost savings and competitive parity, but no structural advantage — every competitor has the same H1 tools inside 18 months.
The Augmenter. A culture in transition — bureaucratic loosening toward generative. Strategy: H1 + H2, human-in-the-loop for every significant decision. Outcome: a real advantage in decision quality, limited by approval chains. The risk is the "frozen middle" — recommendations the org can't act on fast enough.
The Transformer. Generative culture — distributed authority, outcome-based governance. Strategy: H1 + H2 + H3, continuous allocation inside guardrails. Outcome: compounding structural advantage — agents extend the organization's adaptive capacity. The risk is real transformation investment and results that take longer to show.
"I'll just buy it." Agentic AI isn't a process change like Agile — it's a new kind of worker, so why transform at all? Sequoia's 2026 thesis says exactly this: "Replacing an outsourcing contract with an AI-native services provider is a vendor swap. Replacing headcount is a reorg." Cursor hit ~$4B revenue with ~300 people. Three answers: (1) Buying tools is adoption, not transformation — Anthropic asked 81,000 users who benefits from the gains; of those who named anyone, only ~1 in 10 said their employer. The gain defaults to the individual; value leaks around the edges of the org chart. (2) The AI-natives aren't the exception — they're the proof. Cursor wins because it was built around agents with no legacy structure to fight. Structure is decisive. (3) If your answer really is "I'll rent it," you become a reseller of someone else's intelligence while the AI-native keeps the margin. Surviving as more than a passthrough means building the capability inside — which is the transformation.
"We'll fix our data first." The waterfall trap reborn. You discover your data problems by trying to use AI across silos, not by auditing data in the abstract. Data quality is never "done" — it's a practice, not a project. Start AI experiments that expose data problems, then fix them in the context of real value delivery.
The kill-switch (the falsifiable claim). A thesis you can't disprove isn't worth much. Here's the condition that would prove this one wrong: if a rigidly hierarchical, command-and-control organization lets AI agents autonomously set and reallocate its discretionary funding across functions — moving money between R&D, marketing, hiring, and capex — inside guardrails, without per-decision human approval, and sustains it for a full year, then the thesis is wrong. Not within-envelope optimizers (ad-bidding bots, cloud-cost agents, trading algos all reallocate inside a single human-sized budget toward a single number — real, autonomous, and beside the point). The claim is about the agent deciding what the organization funds — the cross-domain trade-off with no single metric to maximize. Base rate as of 2026: Bain surveyed 951 companies and found only 7% running any fully autonomous agent in production.
The cost of waiting — transformation debt. AI capability advances on a monthly cadence; organizational change takes three to five years. That gap is transformation debt, and it compounds. On the hardest agentic coding benchmark, the field went from ~33% to ~77% in about a year (SWE-bench Verified; the 77.2% top mark is Anthropic's Claude Sonnet 4.5, Sept 2025 — OpenAI retired its own reporting in Feb 2026 citing contamination, and the measuring stick wore out before the capability did). On expert-judged professional deliverables across 44 occupations (GDPval), GPT-5.2 beat or tied top human specialists on ~70% of comparisons as of Dec 2025 — note that's one model at one date, not frontier models as a class; the Sept 2025 GDPval field topped out at 47.6%. Ethan Mollick: "remarkably little has changed in most organizations" — and his next line is the warning: "most" isn't "every." The capability isn't the bottleneck. The organization is.
Companion to the talk "You Can't AI Your Way Out of an Organization That Couldn't Agile" — CincyDeliver 2026. The frameworks here are meant to provoke organizational reflection, not prescribe a universal solution. Argue with it.
Matt Anderson · linkedin.com/in/mattanderson