Frequently Asked Questions About Aaron Levie Enterprise AI Diffusion Framework
22 answers covering everything from basics to advanced usage.
// Basics
What are the three enterprise AI diffusion stages?
The Chat Stage (a question-and-answer assistant deployed, productivity rate-limited by human conversation pace), the Early Agent Stage (task-executing agent pilots underway, architecture decisions unmade), and the Agent Scale Stage (stateful or workflow-triggered agents running across multiple functions). Most Global 2000 firms in 2025 sit at the Chat-to-Agent transition. Never prescribe Agent Stage solutions to a Chat Stage organization.
What is the Capability Overhang Paradox?
It's the counter-intuitive dynamic where faster AI breakthroughs make enterprise diffusion slower, because each breakthrough renders the previously implemented architecture obsolete before rollout finishes. Enterprises need a stable, temporarily frozen capability baseline to deploy — and the market won't give them one. The practical takeaway: prioritize lab-neutrality and data-layer correctness over frontier-model novelty, and sign one-year contracts.
What is the Mosaic of Models approach?
It's the practice of routing different tasks to different models by required capability and cost: Frontier models for complex, novel, high-stakes work like advanced contract reasoning, and lower-cost or OSS models for saturated, repeatable tasks like document classification or standard customer service. For Frontier tasks, establish dedicated capacity arrangements to lock in pricing; for saturated tasks, cap token budgets.
What inputs do I need to run this framework on my organization?
Four required inputs: organization type (size and sector), current AI maturity stage (Chat, Early Agent, or none), target use cases (specific knowledge-work workflows), and existing data environment (where data lives, how structured it is, access-control state). Two optional inputs strengthen the analysis: available internal talent (FDEs, repositioned IT, or hiring need) and budget model (IT budget, line-of-business, or undecided).
// How To
How do I run the Coding vs. Knowledge Work checklist on a use case?
Score the use case against five criteria: Are users technical? Are models well-trained on the domain? Is output verifiably testable? Does context live in one accessible repository? Are access controls clean and correctly scoped? Count how many pass. A score of 1-2 out of 5 means a multi-month data-layer project and heavy Internal FDE investment before any agent should go live.
How do I audit my data environment for agent-blocking failures?
Check three modes: Too Little Access (entitlement gaps block the agent from needed data), Too Much Access (overly broad access risks leakage and hallucination), and Data Integrity Gaps (inconsistent definitions, FX adjustments, unmapped ontologies, redundant stores). Resolve access-control issues before deployment. Flag integrity gaps as a semantic-layer or ontology project — a 20-year-old problem now critical because agents give every employee data access.
How do I resolve the IT Budget Escape question?
Determine whether AI compute will stay inside IT (constrained to the 3-7% of revenue ceiling) or migrate to line-of-business opex. If it migrates, assign budget ownership to the CMO, COO, or relevant business owner, and build a lightweight FinOps function per major line of business — they don't have one. Interim best practice: centralized procurement and governance with decentralized spend-decision authority.
How do I choose a reference architecture without getting stuck?
At any moment there are 10-15 viable architectures (managed agents, MCP-connected SaaS, workflow-native agents, lab-direct APIs). Pick one that's lab-neutral enough to swap underlying models without rearchitecting, get the data layer right inside it, and commit for a 12-month horizon. Avoid multi-year lab contracts given the pace of innovation; favor one-year terms. Don't pick based on the most exciting recent lab announcement.
// Troubleshooting
My agent pilot gives confidently wrong answers — what's going wrong?
Almost certainly a data problem. The agent is likely pointed at inconsistently-defined data that a centralized data-science team previously mediated by hand — different definitions, unmapped ontologies, redundant stores. At scale, the agent produces wrong answers for every employee, not just analysts. Stop deployment, run the data-environment audit, and treat integrity gaps as a semantic-layer project before resuming.
My token bills are unpredictable and spiking — how do I control them?
Build the Mosaic of Models: route saturated, repeatable tasks to lower-cost or OSS models with capped token budgets, and reserve Frontier models for genuinely complex tasks. Make employee prompt-cost awareness mandatory training — a single poorly-structured prompt to an agent with broad MCP access can cost the equivalent of a full employee benefit in one execution. Add a lightweight FinOps owner per business line.
My agent keeps bouncing off access walls or touching data it shouldn't — how do I fix it?
You have a Too Little Access or Too Much Access problem. Remediate access controls and entitlements before the agent goes live — don't treat it as a tuning issue post-deployment. Too Little Access means the agent can't reach information it needs; Too Much Access means data leakage and roaming hallucination risk. Both are entitlement-scoping failures that must be resolved in the data layer.
Our AI deployment stalled after a great pilot — why?
You may be experiencing the AI Psychosis Period aftermath: extreme early enthusiasm followed by grounding when maintenance burden, error-catching overhead, and model-upgrade disruption become apparent. Structurally, stalls usually trace to an unfixed data layer, a use case that failed the Coding vs. Knowledge Work checklist, or no permanent Internal FDE to maintain the deployment through upgrades.
// Comparisons
How does this framework compare to generic AI readiness assessments?
Generic assessments score tooling, strategy, and culture in the abstract. This framework is diagnostic and prescriptive: it locates you on a specific diffusion-stage map, scores your actual use case against five transferability criteria, audits three concrete data-failure modes, and prescribes exact sequencing for data, talent, cost, and architecture. It also warns against the Capability Overhang Paradox most readiness models ignore entirely.
How does the Internal FDE differ from the External FDE?
The Internal FDE is embedded within the enterprise's own business function, permanently maintaining agent deployments through model upgrades. The External FDE works for a vendor or startup, going on-premise to make AI products work in customer environments. Both are durable and necessary — the External FDE is a legitimate commercial motion, not evidence the technology is failing. Don't conflate the two roles.
How does the Jevons Paradox lens compare to standard headcount-reduction planning?
Standard planning models AI as headcount reduction first. The Jevons Paradox lens reverses the order: first map which previously unaffordable functions can be unlocked, which projects can be scoped larger with 3-10x engineer capacity, and which new roles the surfaced value demands. Only after mapping expansion do you assess reductions — and reductions are only credible where humans are already at demand saturation.
// Advanced
Should I build a vertical AI app or buy from a startup?
Assess the Bridge Layer Imperative. If the startup provides deep integration with your data model, domain-specific ontology, and change-management support, it occupies a durable bridge-layer position labs are unlikely to replicate without hundreds of vertical go-to-market teams. If it's a thin wrapper with no proprietary data integration, the Capability Overhang Paradox means the next model release may render it obsolete.
Where is the startup opportunity in enterprise AI according to this framework?
The biggest flagged opportunity is AI compute ERP — FinOps tooling for the IT Budget Escape. As spend migrates from IT to line-of-business opex, every function (marketing, sales, legal, manufacturing) needs FinOps capability it lacks. No existing tooling measures ROI per token, enforces per-team budget caps, or attributes compute cost to business outcomes. The wedge: triage token spend across decentralized owners while giving IT centralized governance.
How do I design the Headless + Seated business model for software I'm building?
Use per-seat pricing for end-user-interfaced access and consumption pricing for agent-driven headless volume. Evaluate whether agents 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 — so price and architect for that volume while retaining GUI leverage for complex, nuanced human tasks.
What is the Unlimited Chief of Staff mental model?
It's a heuristic for employees learning to find high-value agent use cases: ask 'what would I give an unlimited chief of staff to work on?' This surfaces tasks where agents create the most personal or organizational leverage. It's part of the employee AI-proofing track — spend 5-10% of work time hands-on with tools like Cursor, Perplexity, and co-worker agents, connecting them to your real workflows.
Why is the bridge layer a product and not a temporary workaround?
Because advanced AI capability never self-deploys into enterprises — security, change management, data integrity, and access controls always stand between the breakthrough and the workflow. Each model generation still requires this bridge. That persistence makes it a durable commercial position, not a gap labs will close. Startups owning deep bridge-layer integration are defensible against both labs and thin-wrapper competitors.
What's the right contract term length for AI vendors and labs?
One year. The Capability Overhang Paradox means today's reference architecture and chosen models are likely obsolete within 12-18 months, so multi-year lab contracts lock you into stale capability. Favor one-year terms and lab-neutral architectures that let you swap underlying models without full rearchitecture. Commit to your chosen architecture for a 12-month horizon, then reassess.
Do I need to address job-protection concerns, or are they just political?
You must address them. Dismissing the social-contract dimension creates real downstream risk to talent pipelines, employee adoption, and the political capital needed for change management. The plan has two tracks: a Company track (funded enablement, training, embedded technical capacity per function) and an Employee track (5-10% of time hands-on with tools, connecting them to personal workflows). Companies owe employees funded upskilling.