Models and tools are not the constraint. The operating model underneath them is, and it's where the savings, the growth, and the compounding stay locked. We name where it blocks progress, redesign the layers that have to change first, and build the shared brain that both humans and agents work from.
Twelve stages. Four tiers. The hinge sits at six.
Each stage creates the conditions for the next, and opens a different layer of the return. Most companies try to skip ahead. Most companies discover the skip did not take.
We score departments and teams across all six vectors, surface the workflow gaps, and name the binding constraints that keep work inaccessible to agents. This is where the baselines get set: the hard savings available now, the KPIs with a dollar chain attached, and the cross-functional potential waiting downstream. The output is a map your leadership team can act on.
Roles get decomposed into mandates. Mandates get sorted: delegate, assist, keep human, retire. The canon gets written down and approved. Policies, pricing logic, customer history, and review criteria move out of people's heads and into the intelligence layer. The structure takes a shape agents can read and action, and the hard savings start to unlock.
Agents enter the redesigned world model with named mandates, permissions, and supervision protocols. Structured work routes to agents; judgment, relationships, and growth route to people. Every loop deposits learning back into the canon. The intelligence layer compounds. The next agent inherits what the last one learned, and the revenue and compounding layers of the return begin to open.
The stages below name the destination at each rung. Where your company actually sits is the heatmap, further down this page.
Everything runs on memory, relationships, and individual habit. Tools exist; nothing connects. The CEO holds the only complete picture, and even that has gaps.
Tools exist department by department, adopted individually. Each team has its own way of working. Information travels through hallway conversations and forwarded emails.
Core systems integrate. The company can see itself in data, at least at the dashboard layer. Individual AI experimentation becomes possible, though fragile.
Real workflows are written down with named owners, steps, and outputs. The handbook starts to resemble the work. Onboarding shortens. Exceptions still live in heads.
Structure is rebuilt around the work the company actually needs done, given what AI can already absorb. Roles get sharper. Coordinator-only roles thin out. The team gets smaller and more capable in the same move.
The organization has a machine-readable representation of how it actually operates: processes, roles, mandates, canon, permissions, decisions, structured as context agents can act on. Everything below is preparation. Everything above is enabled.
AI sits inside every meaningful workflow as an assistant. Humans review and decide; agents draft, classify, and route. Hours recovered begin showing up in real numbers.
Agents complete multi-step workflows autonomously inside named mandates, escalating at defined checkpoints. The redesigned org from Stage 5 is now operational. Cost base looks different.
Multi-agent coordination with a supervisory layer. The intelligence layer is the operational backbone. Humans do the work that requires being human; the routing happens underneath.
Company-wide intelligence replaces most human coordination. Decisions get made at the right altitude because the context is shared.
Intelligence predicts needs, composes capabilities proactively. The organization moves before the prompt arrives.
The organization itself is a self-improving system. No company has fully achieved this. It has been named as a vision and theorized at length.
Stage 6 is the hinge of the entire model. Everything below it is preparation. Everything above it is enabled by it.
A world model is a continuously updated, machine-readable map of how the organization actually works, not how the wiki says it works, but how it actually moves, today, with the decisions and exceptions and carve-outs that make work happen. Without one, agents reference the wrong tools, miss undocumented steps, and hallucinate about processes they cannot see. With one, they act with the understanding of a senior employee.
That is the gap a world model closes. Four things hold it.
How work gets done end-to-end, with exceptions and handoffs named: not the documented flow, the actual flow.
Who is accountable for what, including the relationships invisible on the org chart. Decomposed into mandates. Who knows what. Who works with whom. Where tribal knowledge lives.
How information moves across the organization, with ownership, access, and gaps named. Where it pools, where it breaks, where agents lose the thread.
What the stack can do and how much of that capacity the organization actually accesses day to day. The gap between installed and operational.
Each vector scores 1–12, grouped into four bands: Foundational (1–3), Structured (4–6), AI-Enabled (7–9), Anticipatory (10–12). Your effective capability is constrained by the weakest critical vector, not the strongest. The weakest vector is also where the return stays locked.
How much of the organization's real work, including the exceptions and tribal knowledge, is documented, structured, and accessible. Our working bet: most companies overestimate themselves on this vector by a full tier.
How connected, clean, and accessible the company's data is across systems. From scattered silos through integrated cores to a real-time substrate agents can read and write against.
The gradient from no awareness to organizational intelligence. Individual experimentation can begin at any stage. It will just be fragile until the substrate underneath catches up.
How well the company's structure matches the work it actually needs done. If you were building this team from zero today, knowing what AI can handle, what handful of humans would you hire?
The willingness and ability to change how work gets done. Kills more transformations than any technical gap. A company can have Stage 6 documentation, Stage 7 data, and still fail here.
Guardrails for AI decision-making, data access, autonomous action. The companies that move fastest with AI are the ones that decided in advance what AI may and may not do.
Companies don't sit at one stage. They have peaks and valleys across departments and vectors. The heatmap below shows one example, a fictional 200-person consumer brand we use as a specimen, scored across eight departments and all six vectors. This shape, ambition outpacing infrastructure, is the pattern the Compression Scale is built to detect.
Reading this heatmap: Engineering is operating at Stage 6–7, but Data Architecture is the binding constraint across the rest of the company: three departments are at Stage 1–2 on it. That valley is where the audit's redesign work concentrates. Closing it unblocks the rest.
This is the load-bearing principle behind the diagnostic. Most companies invest in their strengths because their strengths are visible. The weakest vector, the one quietly dragging everything down, is harder to see and harder to fund. The audit's job is to find it and name it.
Our working bet is that the binding constraint usually lives in one of three places: Data Architecture, Governance, or Cultural Readiness. That bet shapes the sequencing rule we bring to every diagnosis: close data architecture gaps first (they block everything), formalize governance second (it becomes urgent the moment AI touches real work), then invest in process legibility and organizational redesign in parallel.
Foundation before Legibility. Legibility before Intelligence. Within a tier, address binding constraints first.
A month-long sprint. Six vectors. The binding constraint holding your company below its potential, named and sequenced. A three-layer return estimate your leadership team can take to the board, and an offer structure that puts us at risk against it.