Why do most AI pilots never scale?
Most AI pilots run outside legal and IT, so each one restarts from zero. Notion found 2% of companies let AI run a core process. How the rest scale.
TL;DR
- Notion's 2026 survey puts 88% of organizations at the personal-tool stage.
- Pilots stall when they run outside legal and IT, so nothing carries over.
- Companies that scale connect AI to existing systems far more often (55% vs 37%).
- Start with one weekly workflow, then measure errors and cycle time.
Most AI pilots never scale because nobody built them to. They run off to the side, away from legal and IT, so each new one starts from zero. Notion's survey of 6,118 people found 88% of organizations still use AI as a personal tool. Only 2% let it run a business-critical process end to end. The fix is dull. Pick one weekly workflow and wire it into the systems it touches. Then judge it on error rates and cycle time.
A pilot | A production workflow | |
|---|---|---|
Where it runs | A side tool, outside legal and IT | Inside the systems the work already uses |
Who owns it | Whoever was curious that quarter | A named person whose job depends on it |
How it's judged | Self-reported time saved | Error rates and cycle time |
What happens next | The next pilot starts from zero | The next workflow reuses its plumbing |
Wrong call for | Work you already know you need every week | A tool nobody has tested on real work yet |
How many companies have scaled AI?
Very few. Notion ran its survey from March to May 2026 across 10 markets. It sorted companies into four levels. At level 1, AI is a thought partner in a separate tab. At level 4, agents run business-critical processes end to end.
The split was 57% at level 1, 31% at level 2, 10% at level 3, and 2% at level 4. (Search Engine Journal's breakdown of the report has the detail.)
Leaders read the same company differently. Notion says decision makers are twice as confident as the people who use AI every day. Senior people are also the heaviest users in the data, at levels 3 and 4 six times as often as individual contributors. So the boss sees their own habits and calls it adoption.
Why do AI pilots stall?
Pilots stall because they're built to be thrown away. Futurist Amy Webb told Fortune about a client that ran 14 or 15 agent pilots this year. Each team was small enough to feed on two pizzas. None scaled. "They've gone through a lot of pizza," she said.
The pilots ran without legal or IT, so nothing got embedded. Each one restarted from zero. Webb compares it to a dating app, where the best outcome for the app is that you never settle down. (A vendor that bills per seat feels the same way about your pilots.)
The money keeps flowing anyway. A Bain survey of 951 companies found nearly 40% of the ones measuring AI savings landed under 10%. Their targets had been 11% to 20%. Ninety percent planned to raise their AI budgets regardless.
A pilot nobody owns after the demo is a very expensive screenshot.
What do companies that scale AI do differently?
They do the plumbing before they ask for results. Notion compared the top two levels with the bottom two.
What changes | Levels 1 and 2 | Levels 3 and 4 |
|---|---|---|
AI connected to existing systems | 37% | 55% |
Governance and oversight in place | 26% | 42% |
Impact measured with real metrics | 22% | 37% |
Reason for AI that grows most | Speed and cost | Customer experience (up 8 points) |
The measurement row matters most to me. At the top levels, self-reported time saved is falling out of use; quality metrics like rework rose 19 points. Ask people whether AI makes them feel faster; you'll hear yes. That yes won't survive a budget review.
How does Space decide which AI workflows to keep?
We keep the ones someone can point to. Space, the AI-native brand agency, runs 127 agents and 362 agentic workflows as of October 2026. We hire against a five-level AI fluency ladder, with level 4 as the floor.
The test for level 4 is short. Level 4 people have artifacts, such as chained processes and workflows with names. Level 3 people have anecdotes about prompts.
The same test works on a pilot. Ask who owns it and what it rejected last week. If nobody can answer, you're looking at a demo. (Narrow agents beat general ones, too; more on that in what we learned building 127 AI agents.)
I'd rather fund one boring workflow in production than ten clever pilots.
Where should a company start with scaling AI?
Start with the workflow your busiest operator runs every week. Ask that person which of Notion's levels the company sits at. Compare the answer with the board deck, then fund the gap.
Webb asks CEOs a sharper question. If AI freed 10% of your capacity tomorrow, where would it go? She says nobody has given her an answer yet. Write yours down before you approve the next pilot.
Say hi anytime at sp-ce.co. We're happy to look at a workflow with you.
Facts verified October 18, 2026.
Frequently asked questions
- What percentage of companies have scaled AI?
- Notion's 2026 survey of 6,118 people found 12% of organizations at levels 3 or 4, where AI works as a teammate or runs processes. Only 2% reached level 4, where agents run business-critical work end to end.
- What is AI pilot purgatory?
- It's the phrase executives used with futurist Amy Webb for running pilot after pilot without scaling any. The pilots show productivity, but they sit outside legal and IT, so nothing gets built into how the company runs.
- Why do AI pilots fail to scale?
- Most pilots skip the integration work. They run as side tools, away from the systems and approvals they'd need, so each new pilot restarts from zero. Many also lack a named owner.
- How do you move an AI pilot into production?
- Pick one workflow you run every week. Connect the AI to the systems that workflow already touches, with a named owner. Then measure error rates and cycle time.
- How should companies measure AI ROI?
- Use quality and workflow metrics, such as error rates or cycle time. Notion found self-reported time saved is fading among the most advanced companies, while quality metrics rose 19 points.
.png%3Fprefix%3Dmedia&w=3840&q=95)
.png%3Fprefix%3Dmedia&w=3840&q=95)
.png%3Fprefix%3Dmedia&w=3840&q=95)