On this page · 7 sections
TL;DR
AI is paying off only when it moves revenue, performance or EBITDA, the numbers that matter, and not just your team’s activity and output.
- AI raises activity and output: more documents, drafts and decks. Output is not a return.
- Every document, meeting and tool has to move one KPI connected to your P&L, with one owner and a weekly review.
- A draft is not a decision. AI can’t tell which market your team can win first, what to stop selling, or which price your buyers will accept.
- AI is a tool to optimize a process, not the base of a system. It goes in after the base system runs.
Someone wants budget for one more AI tool. You already pay for several, and your teams produce more drafts, decks and reports than ever. What nobody can show you is whether any of it moved the P&L. Here’s how I tell, and what I change when the answer is no. My simple check:
For each AI tool, name the decision or customer result it should change. Measure that before you start and for two quarters after. If it doesn’t move, fix the process before buying more AI.
How can you tell if AI is paying off?
You can tell AI is paying off by the decisions and customer results it changes, not by how much your team produces with it.
Most of what AI produces is the work around a decision: research, summaries, drafts. That work prepares a decision, or puts it to work. The decision itself is something else: a choice your team commits to, and everything after it depends on.
Take a go-to-market strategy. AI can draft one in minutes, and I use it for research. But a draft is not a decision. AI can’t tell which market your team can win first, what to stop selling, or which price your buyers will accept. Those calls need your buyers’ voices, your numbers and the people who will run the plan in the room. The same goes for the decisions a GTM playbook starts with: who to sell to, the problem you own, how you’re positioned and what everyone says.
The research points the same way. In MIT NANDA’s 2025 study of AI in business, over 80% of organizations had explored or piloted tools like ChatGPT and Copilot, and nearly 40% reported deployment1. Yet the authors found that these tools “primarily enhance individual productivity, not P&L performance.” After $30 to 40 billion of enterprise investment in generative AI, 95% of organizations were getting zero return1. The buyers who did get a return judged tools by business outcomes, not software benchmarks. Some of them cite customer results: better retention and sales conversion1.
It’s a preliminary study, built on interviews with 52 organizations, answers from 153 senior leaders and a review of over 300 public AI initiatives1. So read its numbers as a direction, not a precise count. The direction is clear: count what changes in the business, not what the tool produces.
Why does AI raise output but not revenue?
AI raises output but not revenue because it speeds up a process, and when no one can say what a document is for or which KPI it moves, the extra output is noise. That’s AI workslop, and it’s a system problem.
Everyone wants to get into AI as soon as possible. When I get into a company, what I find is noise: more documents, more drafts, more decks. AI raised the activity and the output. It didn’t raise revenue, performance or EBITDA: the numbers that matter. A Harvard Business Review article of September 2025 called output like this AI-generated “workslop”2.
A faster process makes more of what it already makes. If a team writes documents nobody reads, AI helps it write more of them. If nobody owns the number, nobody notices that it didn’t move. That’s the AI productivity paradox inside one company: more output, the same P&L.
MIT NANDA’s researchers reached a similar view. In their study, the gap between the few organizations that get a return and the rest doesn’t seem to come from model quality or regulation, but from approach1. To me, approach means the system around the tool: people, process and tools, with one owner and one KPI.
What should you measure AI against?
Measure AI against the decision or customer result each tool should change, on a KPI connected to your P&L, with one owner and a weekly review.
People move toward what they’re paid for. So every system has one owner and one KPI, connected to your P&L, and the people in it are paid toward that KPI. Your team reviews it every week on one dashboard. Every document, meeting and tool has to move that number, and an AI tool is no exception. My simple check applies that rule one tool at a time.
The simple check, step by step
| The check | What to write down |
|---|---|
| For each AI tool, name the decision or customer result it should change. | The decision or customer result, in one sentence. The KPI it moves, connected to your P&L. The one person who owns that KPI. |
| Measure that before you start and for two quarters after. | The number before the tool starts. A weekly review of that number on one dashboard. Measurable results tied to business goals for two consecutive quarters. |
| If it doesn’t move, fix the process before buying more AI. | The root cause, in people, process or tools. Whether the people in the process are paid toward that KPI. Whether the base system passes my 12-point done checklist before more AI goes in. |
If you can’t write the decision in one sentence, you haven’t made it yet. And because the check runs tool by tool, each tool answers for its own number. Nobody has to argue about AI in general.
Two quarters is close to how MIT NANDA judged success: a tool counted only when it went past the pilot with measurable KPIs, and the researchers measured the return six months after the pilot1.
What do you change when AI isn’t paying off?
When AI isn’t paying off, change the system before the tool: find the root cause in people, process and tools, then give the result one owner and one KPI tied to your P&L.
Most of the trouble sits around the tool. In BCG’s 2024 survey of 1,000 executives, 74% of companies had yet to show tangible value from AI3. Around 70% of their AI challenges came from people and process, 20% from technology and only 10% from the AI algorithms3.
I change things in this order:
People first, because AI is just a tool: the right people in the right roles, with job descriptions they know.
Then the incentives. Who each person reports to, and what you pay them for. Hold people accountable, and follow up every week.
Then one dashboard. A single KPI, connected to your P&L, reviewed every week and changed quarterly if needed.
Then the process that helps the people get there. Inside it, look for what AI tools or software can improve.
The AI tool and the software serve the process that serves the person, and not the other way around. AI goes in after the base system runs. My guide on how to build business systems shows how to build that base and when to call it done.
This order isn’t only for AI. I used it at Digital Smile Design, where I led a new business unit as a fractional executive, with a new service offer, a new team, positioning, pitch and sales process, built as systems of people, process and tools. It went from $0 to $200K in a quarter (client cases).
Questions people ask
How long should you measure an AI tool before you judge it?
Two quarters, against the number from before the tool started. It’s the same bar I use to call any business system done: measurable results tied to business goals for two consecutive quarters. MIT NANDA’s 2025 study used a similar window: it measured the return six months after the pilot.
Why do most AI pilots fail to pay off?
Most AI pilots fail to pay off because the tools don’t fit the daily work. In MIT NANDA’s 2025 study, 60% of organizations evaluated custom or vendor-sold enterprise AI tools, 20% reached a pilot and 5% reached production. Most failed because of brittle workflows, tools that don’t learn from context, and a poor fit with day-to-day operations. That’s why AI goes in after the base system runs: the tool has to fit a process that already works.
Does AI increase productivity?
For individuals, yes: MIT NANDA found the common tools mostly help individual productivity, not P&L performance. For whole companies, the gain is small so far. In an NBER working paper from 2026, nearly 9,000 executives in seven countries said AI had raised annual labor productivity growth by 0.2% over the last three years, and expect 0.6% a year over the next three. So far, they see essentially no net effect on employment.
Who should own an AI tool in a company of 20 to 300 people?
The owner of the system the tool serves. Every system has one owner and one KPI, and tools are one of its three pillars, next to people and process. So the tool answers to that owner’s number. MIT NANDA saw something similar among the buyers who got a return: frontline managers brought the AI projects, not a central lab.
The bottom line
AI pays off inside a system: one owner, one KPI tied to your P&L, people paid toward it, and a weekly review. Count what moves that number, not what AI produces. When the number stays flat, fix the system first, and bring AI in after it runs.
AI is a tool. It speeds up a process. It can’t be one.
Sources
- Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari, “The GenAI Divide: State of AI in Business 2025”, MIT NANDA, July 2025 (preliminary findings). mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf
- Kate Niederhoffer, Gabriella Rosen Kellerman, Angela Lee, Alex Liebscher, Kristina Rapuano and Jeffrey T. Hancock, “AI-Generated ‘Workslop’ Is Destroying Productivity”, Harvard Business Review, September 22, 2025. hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity
- Boston Consulting Group, “AI Adoption in 2024: 74% of Companies Struggle to Achieve and Scale Value”, press release, October 24, 2024. bcg.com/press/24october2024-ai-adoption-in-2024-74-of-companies-struggle-to-achieve-and-scale-value
- Ivan Yotzov, Jose Maria Barrero, Nicholas Bloom and others, “Firm Data on AI”, NBER Working Paper 34836, February 2026, revised October 2026. nber.org/papers/w34836
