There are two stories about AI doing the rounds. In one, we are months away from machines that do everything and nobody will have a job. In the other, it is all a bubble about to burst. Notion went out to see what is actually happening inside companies, and the picture is more useful than either story.
The editors called the report The Great Renovation. Every renovation costs more than the estimate, takes longer than the contractor promised and, in the end, is usually worth it. Below are the numbers that matter most and what we take from them if you want AI doing real work inside your operation.
What the survey measured
The survey covered two groups: 2,039 people who buy, influence or roll out AI (the leaders) and 4,079 people who used an AI chat tool at work in the past month (the teams). Markets included the United States, the United Kingdom, France, Japan, South Korea, Australia, Singapore, the Nordics, DACH (Germany, Austria and Switzerland) and Benelux.
Each answer was mapped onto a four-level scale: AI as a thought partner, AI as an assistant, AI as teammates and AI as the system. It is the same logic as the 4 levels of AI adoption we use on this blog.
Almost everyone is still early
88% of companies sit at levels 1 and 2, using AI as a personal tool to draft, summarize and brainstorm. Only 12% have reached the point where AI runs recurring workflows or parts of the business on its own.
Two things jump out. The first is reassuring: if you feel behind, you probably are not. Early is the norm. The second is a warning: the hardest step is from level 2 to level 3. It is not about using more AI, it is a change in kind. AI stops helping a person and starts running steps of the work, which takes integration with your systems, clear rules and someone who designs the process.
Only 2% of companies operate at level 4. The keynote demos and case studies you keep seeing are that 2%. Benchmarking against them means benchmarking against the exception.
The profile of the leaders is telling too. Mid-sized companies with 500 to 999 employees lead (17% at levels 3 and 4), while enterprises above 5,000 sit at 7%. Owners and CEOs are more than six times as likely to be at levels 3 or 4 as individual contributors (39% versus 6%). Size does not guarantee maturity; being close to the decisions does.
Leaders and teams describe different companies
Ask the same question and the answer depends on who is answering. Leaders say 26% of their organizations are at level 3 or beyond. Teams in the same markets say 5%. That is a five-fold gap.
This is more than a perception gap; it is an execution risk. An AI strategy built only on the leadership view plans for a company that does not exist. And the pressure rises with maturity: 55% of leaders say they are investing in AI faster than their people can learn it, climbing from 48% at level 1 to 68% at level 4.
Two barriers that never go away: trust and visibility
71% of people would use AI more if they trusted it not to make mistakes on important work. That number does not drop as companies mature. Trust does not show up on its own over time: it has to be designed in, with human review at the right checkpoints, answers that show where they came from and clear rules on what AI may and may not do.
The other barrier is seeing what is going on. Half of all leaders say employees use unapproved AI tools, or admit they do not know which tools are in use. Among teams, 41% say the same. Company data pasted into personal accounts, with no log and no control, is a security problem and a compliance one under GDPR, CCPA and similar rules.
They are two sides of one problem: if people do not trust it, they will not use it; if leaders cannot see it, they cannot help.
As AI matures, the nature of the work changes
At levels 1 and 2, AI makes people faster at what they already did: writing, research, summaries. At levels 3 and 4, the systems work grows, the connective tissue between parts of the business.
Automating repetitive tasks doubles (from 18% to 36%) and routing work between tools climbs 15 points. Writing and research stay near the top but lose ground. Motivation shifts too: at the higher levels people use AI to do work they could not do before (from 10% to 20%) and to make better decisions. Leaders move their focus from productivity and cost cutting to customer experience (+8 points) and new capabilities (+5).
The payoff stops being time saved and becomes new capability. That is where our rule of thumb kicks in: if a task repeats, it can be automated.
The three choices that separate the leaders
The survey listed eight implementation strategies. Three open a gap nearly twice the size of any other:
- Integration (+18 points). AI wired into the ERP, the CRM, email, Slack or Teams, where work already happens, not in a separate tab.
- Governance (+16 points). Who owns each agent, what it may do, where it stops and asks for approval, and how everything is logged.
- Measurement (+15 points). Numbers agreed before you start and tracked after.
Training, written policies and standard tools matter, but nearly everyone already has them. They get a company off the ground and then plateau. What keeps compounding is wiring AI into the systems and rules that already run the business.
The problems do not disappear, they change shape
Companies that move ahead trade old problems for new ones. The human barriers get easier: low trust in AI output falls 8 points and the skills gap falls 3. Training and enablement work.
The structural problems get harder. Tool sprawl more than doubles (from 11% to 25%), real impact gets harder to see and model behavior varies more. Early on, the challenge is helping people use AI. Later, it is controlling the system around it: making it work the same across different models, seeing what each agent is doing and containing the sprawl.
Measuring the right things is the clearest divide
Close to 80% of level 3 and 4 companies measure AI’s impact, compared with 62% at levels 1 and 2. But the bigger difference is what they measure.
Mature companies track quality, flow and money. Early ones lean on “I think it saved me two hours.” Time saved is a number for the AI team; cycle time and revenue are numbers for the business. And measuring well is not just about proving ROI to finance: it is what builds the internal trust that lets AI take on more.
Four questions to take back to your team
The report closes with questions worth a meeting of their own:
- Are we investing in integration or just access? Handing out licenses is the starting line. Wiring AI into your systems is what makes the gap grow.
- Do leaders and teams agree on where we are? Probably not. Ask both sides before you pick the next move.
- Are we measuring the right things? Swap estimated hours for error rates, rework, cycle time and cost.
- Do people trust AI enough to use it, and use it in the open? Trust and visibility get solved together, with rules, logging and human review where it counts.
What this means in practice
The survey matches what we see every day: the jump that matters is moving from everyone’s own chat window to agents running steps of real processes, inside the company’s systems, with rules and numbers. That jump from level 2 to level 3 is exactly where most companies stall.
- The AI Analyst maps the routine, listens to leaders and teams (both sides of the gap), finds what repeats and sets the metrics before anything is built.
- The Forward Deployed Engineer works inside your operation and connects agents to the systems you already use, with governance, logging and human approval at the right checkpoints.
- Training takes care of trust, the one barrier that does not fade on its own.
Finance, customer support, sales, legal, procurement and logistics are just examples. Wherever there is work that repeats, there is a next level.
The renovation will be slow and sometimes messy. The companies that integrate, govern and measure get there first.