Aaron Levie Enterprise AI Diffusion Framework

// When should you use the Enterprise AI Diffusion Framework?

Use this skill whenever planning, auditing, or advising on an agentic AI rollout inside a large or mid-market enterprise. Also apply it when evaluating whether a startup opportunity exists in the bridge layer between AI capability and real-world enterprise workflow.

// What information do you need before applying the framework?

• Organisation typerequired
  Size and sector of the enterprise (e.g. Global 2000 manufacturer, mid-market financial services firm)
• Current AI maturity stagerequired
  Where the org sits today: Chat Stage, early Agent Stage, or no deployment yet
• Target use case(s)required
  The specific knowledge-work workflow(s) targeted for agentic deployment (e.g. client onboarding, contract review, marketing campaign generation)
• Existing data environmentrequired
  Description of where enterprise data lives, how structured/unstructured it is, and the current state of access controls and entitlements
• Available internal talent
  Whether the org has internal FDEs, repositioned IT engineers, or must hire from outside
• Budget model
  Whether AI spend currently sits in IT budget, line-of-business budgets, or is undecided

// What core principles drive enterprise AI diffusion?

// How do you apply the Enterprise AI Diffusion Framework step by step?

1. 1

   Locate the organisation on the Diffusion Stages map

   Determine which stage applies: (A) Chat Stage — chat assistant deployed or rolling out, productivity rate-limited by human conversation pace; (B) Early Agent Stage — pilots of task-executing agents underway, architecture decisions unmade; (C) Agent Scale Stage — agents running in stateful or workflow-triggered modes across multiple functions. Most Global 2000 firms in 2025 sit at A-to-B transition. Do not prescribe Agent Stage solutions to a Chat Stage organisation.

2. 2

   Run the Coding vs. Knowledge Work Checklist on the target use case

   Score the use case against five criteria: (1) Are users technical? (2) Are models well-trained on this domain? (3) Is output verifiably testable? (4) Does context live in a single accessible repository? (5) Are access controls clean and correctly scoped? The more criteria that fail, the longer the diffusion timeline and the more internal FDE investment is required before deployment.

3. 3

   Audit the Data Environment for the three agent-blocking failure modes

   Check for: (a) Too Little Access — agent cannot reach the information it needs due to entitlement gaps; (b) Too Much Access — agent has broader access than appropriate, risking data leakage or wrong answers; (c) Data Integrity Gaps — inconsistent definitions, FX adjustments, unmapped ontologies, or multiple redundant stores that a centralised data-science team previously mediated manually. Flag all three. Resolve (a) and (b) via access-control remediation before any agent goes live. Flag (c) as a semantic layer / ontology project — this is a known 20-year-old problem now critical because agents democratise data access to every employee, not just the analytics team.

4. 4

   Map the Token Cost Exposure and design the Mosaic of Models

   Identify which tasks require Frontier model capability (complex reasoning, novel contract language, advanced coding) versus which tasks are now saturated and repeatable (standard customer service responses, document classification). Route saturated tasks to lower-cost or OSS models with a capped token budget. For Frontier tasks, establish dedicated capacity arrangements where available to lock in pricing. Flag that a single poorly-structured agent prompt can cost the equivalent of a staff benefit in one execution — employee training on prompt cost awareness is mandatory, not optional.

5. 5

   Resolve the IT Budget Escape question with finance and line-of-business owners

   Determine whether AI compute spend will remain inside IT budget (constraining to 3–7% of revenue ceiling) or migrate to line-of-business opex. If migrating, assign budget ownership to the CMO, COO, or relevant business owner for their domain. Build a lightweight FinOps function for each major line of business — they do not currently have one. Acknowledge this is an unsolved tooling gap; a centralised procurement and governance layer with decentralised spend-decision authority is the interim best practice.

6. 6

   Staff the Internal FDE function

   The Internal FDE (Field Deployment Engineer) role is a highly technical person embedded in or adjacent to a business function whose job is to: understand how that function works day-to-day, wire up the correct data sources and access controls, configure agent instructions and human-in-the-loop checkpoints, and maintain the deployment as models upgrade. This is not a one-time implementation role — model upgrades continuously create new work. Source from: repositioned IT engineers, new CS-graduate hires, or pivoted software engineers. Do not conflate with the External FDE role that vendor/startup field teams play, though both are required.

7. 7

   Choose the Reference Architecture and hold the line for 12 months

   At any given moment there are 10–15 viable reference architectures for a given agentic use case (managed agents, MCP-connected SaaS, workflow-native agents, lab-direct APIs, etc.). The proliferation of choices extends sales cycles and creates decision paralysis. Pick one architecture that is lab-neutral enough to swap underlying models without full rearchitecture, get the data layer right inside it, and commit for a 12-month horizon. Avoid multi-year lab contracts given the pace of innovation; favour one-year terms.

8. 8

   Design the Headless + Seated business model for any software being built or procured

   If building a product or negotiating vendor terms: per-seat pricing covers end-user interfaced access; consumption pricing covers agent-driven headless volume. Evaluate whether agents in the specific use case need a stateful identity in the system (warranting a cheaper agent-seat) or are purely on-demand operations (pure consumption). By raw operation volume, headless will dominate — price and architect accordingly.

9. 9

   Apply the Jevons Paradox lens to workforce planning

   Do not model AI deployment as headcount reduction first. Model it as: (a) which functions previously unaffordable can now be unlocked? (b) which projects can now be scoped larger because existing engineers/analysts have 3–10x capacity? (c) which new roles (agent managers, internal FDEs, workflow designers) does the value surfaced by agents create demand for? Only after mapping expansion scenarios should reduction scenarios be assessed, and reductions are most credible in functions already at demand saturation with humans.

10. 10

    Build the employee AI-proofing plan

    Companies owe employees a real shot at upskilling as part of the social contract. The plan has two tracks: Company track — funded enablement, training, internal tooling access, and embedding technical AI capacity in each major function. Employee track — spend 5–10% of work time using tools hands-on (Codeex, co-worker, Perplexity Computer, Cursor for semi-technical users); connect tools to personal workflows; develop the 'unlimited chief of staff' mental model to identify where agents create the most leverage in your specific role.

// What do real enterprise AI diffusion scenarios look like?

// What mistakes should you avoid when deploying enterprise AI agents?

• Extrapolating coding-agent productivity gains to non-engineering knowledge work — the five structural conditions that make coding agents work (technical users, verifiable output, clean codebase context, clean access controls, model training depth) do not transfer.
• Deploying agents before fixing the data layer — agents with messy access controls either bounce off entitlement walls immediately or leak data they shouldn't touch; agents pointed at inconsistently-defined data give confidently wrong answers at scale.
• Treating the Internal FDE role as a one-time implementation project — model upgrades continuously create new work; this is a permanent embedded function, not a deployment consultant engagement.
• Locking into multi-year architecture or lab contracts — the Capability Overhang Paradox means today's reference architecture is likely obsolete within 12–18 months; one-year terms are the current best practice.
• Assuming AI spend stays within the IT budget ceiling — the 3–7% of revenue IT budget cap will be broken as agents deliver value in line-of-business functions; failing to plan for IT Budget Escape leaves organisations with no governance model when bills arrive in the CMO's budget.
• Letting employees use agents without token cost awareness — a single poorly-structured prompt to an agent with broad MCP access can cost as much as a full employee benefit in one execution; cost literacy is a mandatory training component, not a nice-to-have.
• Treating the absence of a fully headless future as a failure — the Headless + Seated Dual Model is the correct end state; trying to go fully headless or refusing to go headless at all are both wrong; GUI interfaces retain genuine leverage for complex human tasks.
• Dismissing job-protection concerns as purely political — ignoring the social contract dimension creates real downstream risk to talent pipelines, employee adoption, and political capital needed to drive change management; companies owe employees funded upskilling as part of the deployment plan.
• Picking one of 10–15 reference architectures based on the most exciting recent lab announcement — architecture decisions made under Capability Overhang pressure should prioritise lab-neutrality and data-layer correctness over frontier-model novelty.

// What key terms do you need to understand enterprise AI diffusion?

Tokenmaxxing

The Silicon Valley engineering culture of optimising for maximum token consumption — using the most capable, highest-context Frontier models at the highest volume possible to extract maximum output. Contrasted with enterprise cost management instincts.

Capability Overhang Paradox

The counter-intuitive dynamic where the faster AI breakthroughs arrive, the slower enterprise diffusion becomes, because each breakthrough renders the previously implemented standard architecture obsolete before it finishes rolling out, removing the stable deployment environment enterprises require.

Diffusion Stages

The three-stage model of enterprise AI maturity: Chat Stage (question-and-answer assistant, productivity rate-limited by human conversation pace), Early Agent Stage (task-executing pilots, architecture undecided), and Agent Scale Stage (stateful or workflow-triggered agents running across multiple functions).

Bridge Layer

The required layer of integration, security, change management, data preparation, and workflow configuration that sits between raw AI model capability and an enterprise's real-world workflows. It is not a temporary gap — it is a durable product and commercial opportunity.

Headless Software

Enterprise software accessed entirely via API or agentic interface, with no human-facing graphical user interface involved in the transaction. In the Headless + Seated Dual Model, headless agent-driven volume will dwarf human-seated usage by raw operation count.

Headless + Seated Dual Model

The future business model for enterprise software: per-seat pricing for end-user-interfaced work by humans, plus consumption pricing for agent-driven headless operations. Both models coexist; neither fully displaces the other.

Internal FDE

Internal Field Deployment Engineer — a highly technical role embedded within a business function whose job is to understand that function's workflows, wire up data sources and access controls, configure agent instructions and human-in-the-loop checkpoints, and maintain agent deployments through model upgrades. A permanent function, not a project role.

External FDE

Field Deployment Engineer at a vendor or startup who goes on-premise or works closely with enterprise customers to make AI products work in their specific environment. A necessary and durable commercial motion, not evidence that the technology is failing.

Mosaic of Models

The enterprise practice of routing different tasks to different models based on required capability and cost: Frontier models for complex, novel, high-stakes tasks; lower-cost or OSS models for saturated, repeatable tasks where capability requirements are well-understood.

IT Budget Escape

The structural shift in which AI compute spend migrates out of the traditional IT budget (historically 3–7% of corporate revenue) and into line-of-business opex as agents deliver value in marketing, sales, legal, manufacturing, and other functions — creating a governance and FinOps gap in those business units.

Jevons Paradox (applied to AI jobs)

The principle that productivity gains from agents do not eliminate net jobs because they cause organisations to sign up for larger projects, unlock previously unaffordable functions, and generate new roles to capture agent-surfaced value. Named after the 19th-century economist William Stanley Jevons, applied here to knowledge-work employment.

The Data Problem

The root cause of most agentic failures, expressed as three failure modes: Too Little Access (agent can't reach needed information), Too Much Access (agent roams into inappropriate data), and Data Integrity Gaps (inconsistent definitions that were previously mediated by a centralised data-science team and now scale incorrectly to every employee via agent access).

AI Psychosis Period

The phase a user or organisation goes through when first encountering agents — characterised by extreme enthusiasm, weekend-long build sessions, and belief that everything will be transformed immediately — followed by a grounding phase when the maintenance burden, error-catching overhead, and model-upgrade disruption become apparent.

Unlimited Chief of Staff Mental Model

A heuristic for employees learning to identify high-value agent use cases: ask 'what would I give an unlimited chief of staff to work on?' to surface tasks where agents create the most personal or organisational leverage.

// FREQUENTLY ASKED QUESTIONS