I've spent my entire career analyzing cold, hard data. I don’t think I’ve seen anything quite like @Larridin's 2025 State of Enterprise AI report. We surveyed 350 senior finance and IT leaders in companies of 1,000 or more employees, and made some startling discoveries. (🧵)

Love to see @CALEBcsw tackling Case after such a great game! This is real support for your teammates! @BarstoolBigCat

"The history of football is defined more than anything by the great quarterback and coach combination.. I love Ben Johnson and Caleb Williams.. I think Ben Johnson is brilliant" ~ @Espngreeny #PMSLive

Wall Street paid Jordan Belfort $750,000 a year to sell like this and now you can watch him do it live for free 22 minutes. Real prospect. Real close. The Straight Line method in action from the man who invented it Nearly 8 million views for a reason. Save it while you can ⭣

There is a layer of enterprise AI measurement that most organizations haven’t yet built, and it may be the most important one. Tool-level visibility tells you what is being used. Spend visibility tells you what it costs. Neither tells you how AI is changing the way work actually moves through an organization. Workflow visibility is the missing piece. Which processes have been restructured around AI. Which ones are running the same way they did two years ago, but with an AI tool bolted on. Where are handoffs getting faster. Where are they getting stuck. Without this layer, you can report adoption, and you can report spend, but you cannot answer the question that matters most: Is AI changing how we operate, or is it just adding new costs to the old status quo?

There is a question that boards are increasingly asking, and most orgs cannot answer it with confidence: What is our AI investment producing? The challenge is that answering this question requires connecting three layers of data that typically live in separate systems. Spend data lives in finance. Adoption data lives in IT. Outcome data lives in business units. Bringing them together into a coherent view requires infrastructure that most organizations haven’t built. The result is that executives walk into board meetings with partial answers. They can report how much was spent. They can report how many people have access. They struggle to report what those people produced with that access, and whether the output justified the cost. This is becoming more visible as AI spend grows. When the AI line item on the budget was small, partial answers were acceptable. As it becomes material, boards expect the same rigor applied to AI investment that they expect for any other significant capital allocation.

Median AI proficiency across enterprise deployments, based on measured output rather than self-assessment, sits at approximately 58.5 out of 100. Just above the midpoint at 50. Organizations are investing in tools, running training programs, and reporting strong adoption, while the typical employee is still operating at a basic level of capability. The gap between adoption and proficiency is where value disappears. A team that has access to AI, but cannot apply it effectively, is not just less productive. Under consumption-based pricing, that team is also more expensive. Every misfired prompt, every vague instruction, every additional, iterative exchange to get a usable output costs money. Self-reported proficiency surveys cannot surface this problem, because people genuinely struggle to assess their own capability. The only way to see it is to measure actual output quality and value.

There is a question that boards are increasingly asking, and most orgs cannot answer it with confidence - What is our AI investment producing? The challenge is that answering this question requires connecting three layers of data that typically live in separate systems. Spend data lives in finance. Adoption data lives in IT. Outcome data lives in business units. Bringing them together into a coherent view requires infrastructure that most organizations haven’t built. The result is that executives walk into board meetings with partial answers. They can report how much was spent. They can report how many people have access. They struggle to report what those people produced with that access, and whether the output justified the cost. This is becoming more visible as AI spend grows. When the AI line item on the budget was small, partial answers were acceptable. As it becomes material, boards expect the same rigor applied to AI investment that they expect for any other significant capital allocation.

Process discovery is the ability to map how work moves through an organization and identify where AI is changing that flow. The value is the ability to find automation candidates that would otherwise remain invisible. Most workflow improvements happen because someone on a team notices an opportunity and raises it. That approach depends on individual initiative and surface-level observation. It misses the patterns that only become visible when you can see work moving across functions and systems. Organizations that implement automated process discovery infrastructure will find opportunities their competitors cannot see. They will be able to answer questions like: Where are the highest-value automation candidates? Which workflows have been restructured around AI? Which are running unchanged? Also: Where are the bottlenecks that AI could address, but has not?

Brandon Veiseh, CEO of MindFort, said his six-person company would have needed twenty employees pre-AI. The difference went to tokens. His framing: "We have to weigh our token-to-people ratio." This is a new workforce metric: the ratio of token spend to human labor required to produce a given output. The implications extend beyond startups. Every company is now implicitly making this tradeoff, whether they have named it or not. Are they making it deliberately or by accident?

Russ was quoted in Business Insider's piece on token budgets. He compared access to frontier models to corporate travel: everyone can book an economy flight, but almost nobody is authorized to charter a jet. The analogy implies a governance structure that most organizations have not yet built, one based on tiered allocation according to demonstrated need rather than blanket restrictions or unlimited access. The practical challenge is determining who needs what and that requires visibility into which employees are using which models, for which tasks, and with what outcomes. Without that infrastructure, allocation decisions become arbitrary, and arbitrary decisions tend to create resentment without producing savings. Link to the full piece in comments.

Larridin research on AI coding assistant spend reveals a pattern worth examining closely. The first third of spending is responsible for 85% of productivity gains. After that inflection point, spending rises faster than productivity. The return on each additional dollar declines significantly. Less than a third of users account for more than half of total spend. These are the heaviest users, often celebrated on internal leaderboards. But leaderboard position correlates with token consumption, not output. The implication is that, without visibility into the relationship between spend and outcomes, organizations cannot connect the two. Several trends converged to create this dynamic. AI providers introduced agentic workflows that burn tokens continuously. They moved to per-token pricing. Companies encouraged adoption without measurement frameworks. Token usage became a proxy for competence. The result was predictable in hindsight. Spending exploded and visibility did not keep pace. Teams are now scrambling to install guardrails around the cost-benefit inflection point. They’re trying to give employees feedback fast enough to adjust behavior, as well as measuring productivity alongside spend, not instead of it.

Asking what AI costs is an incomplete question. Cost in isolation tells you nothing useful. A team spending $100,000 per month on AI might be generating $10 million in value. Another team spending $50,000 might be generating nothing. The return numbers tell the real story. The challenge is that most organizations have built cost visibility without return visibility. They can tell you what was spent. They cannot tell you what was produced - Hours saved. Velocity gained. Outcomes achieved. Without both sides of the equation, you cannot make portfolio decisions. You cannot distinguish between productive spend and waste. You cannot answer the question that matters: is this working?

There is a category of risk accumulating in enterprise AI deployments that we’re referring to as agent debt. This is the accumulation of active agents with no accountable owner. This happens when an employee authorizes an agent, connects it to production systems, and then moves on. In a typical first scan of an enterprise deployment, we find an average of 40+ agents in this state. The parallel to technical debt is instructive. Technical debt accumulates silently, compounding over time. And it becomes exponentially harder to address the longer it goes unexamined. Agent debt follows the same pattern. Each agent is a small liability. Forty-ish agents is a budget risk. Hundreds is a governance failure waiting to surface at the worst possible moment.

There is a category of risk accumulating in enterprise AI deployments that we’re referring to as agent debt. This is the deploying active agents with no accountable owner. This happens when an employee authorizes an agent, connects it to production systems, and then moves on. In a typical first scan of an enterprise deployment, we find an average of 40+ agents in this state. The parallel to technical debt is instructive. Technical debt accumulates silently, compounding over time. And it becomes exponentially harder to address the longer it goes unexamined. Agent debt follows the same pattern. Each agent is a small liability. Forty-ish agents is a budget risk. Hundreds is a governance failure waiting to surface at the worst possible moment.