Enterprises bought AI across OpenAI, Anthropic, Google, and Microsoft, but most of it still runs shallow. ATOM makes the work explicit: what the tools can see, what they may do, who reviews the result, and whether the spend paid off.
01
The economics
AI spend is climbing faster than its value.
AI spend rises with every seat and tool added.
Value realized stays flat — it never reaches the work.
The gap is the sticker shock; renewals and expansions stall.
Access went in; adoption didn't follow. People got the tools and were left to work it out — so a frontier model gets used like a faster chatbot, and the return is anyone's guess.
02
The missing middle
The value lives in the middle the enterprise owns.
Layer 1 · rented
AI tools
ModelsCopilotsCoding toolsAgentsModel providers
The customer-owned operating core
Layers 2 & 3 · owned
Layer 2 · the work
The work
PeopleAI-assisted peopleAgentsMeetings · research · docs · sales · decisions
The tools are rented and the systems are fixed; the middle is the part the enterprise owns — and the part that decides whether any of it pays off. ATOM starts with the tools already owned, and every decision that defines success stays with the enterprise.
03
The engagement
Adoption runs on evidence, in three gears.
1
Crawl · foundationMake it reviewableConfigure existing tools to one slice; set the baseline.
→
2
Walk · active useSetup becomes adoptionReal people, real work, real governance.
→
3
Run · scaleScale what's backedMore teams, more work, the software the proof earned.
Native first · the enterprise owns every decision · the pace follows the evidence. The readout, the software, the position →
04
The readout
Six pillars show what is working.
Cost
Is the work worth what it costs — judged against the role and its KPIs, not seats and tokens?
Context
What did the AI have to work with, and what was missing, stale, or off-limits?
Control
What policy, permissions, and review steps shaped the work?
Memory
What turned into reusable knowledge, and what should be retired?
Performance
Did the work get better — faster, sharper, a stronger outcome?
Portability
Does the work survive a change of model, vendor, or team?
In the paper, these six are filled in on one real workflow — a sales account brief, diagnosed pillar by pillar.
05
The software
An operating model with real software under it.
Delivery · sets the baseline
The proof product
Maps a live environment, exposes the gaps and weak evidence, and produces the first six-pillar readout. Same method, every engagement, fast.
→
proof first, installed follows
Customer-owned · persistent
The installed product
Software the enterprise keeps and owns, on its own data — where the custom build happens, earned by what the proof shows.
One set of instrumentation under bothacross the tools already owned and the enterprise's own systems
The partner who turns access into adoption and proves the economics holds the most durable, expandable position in enterprise AI — the one that earns the renewal, and the next build.
Proven today in knowledge work — the work system I run my own business on, and the same pattern I use with customers. Engineering is an honest bet on the same primitives. The first engagement pays for itself and produces a qualified map of what comes next:
Scale the laneThe approach works; take it wider.
Deepen native configMore from the tools already owned.
Build owned softwareWhere a real gap remains.
Commission agentsOnce the work is known well enough to delegate.