Airtable published a short guide this year called the Human-Agent Collaboration Readiness Framework. It is four pages and worth reading. It sorts every team into one of five stages of working with AI:
- Explore. Scattered personal use. No shared workflows.
- Assist. People use AI for drafts, summaries and questions, each on their own.
- Delegate. People design the workflow and AI runs defined steps inside it.
- Supervise. Agents run on schedules and triggers without anyone starting them. People set goals and handle exceptions.
- Orchestrate. Agents work across departments. People set strategy and governance.
By Airtable's count, 12% of organizations are at Explore, 44% at Assist, 22% at Delegate, 18% at Supervise and 3% at Orchestrate. That means more than half are still in the first two stages.
The guide leans on McKinsey numbers, and they hold up. In McKinsey's State of AI survey from November 2025, 88% of organizations said they use AI regularly in at least one business function, and only about 6% qualified as high performers, with AI driving 5% or more of earnings. An earlier McKinsey report, from March 2025, found that only about one in five organizations using generative AI had redesigned even some of their workflows around it.
So almost everyone uses AI and very few get much out of it.
The line in the guide I agree with most
Airtable's summary says the real bottleneck "isn't model intelligence but deploying agents into existing, trusted workflows with governance, approvals, and human accountability."
That is right. The models are good enough. What is missing in most companies is the plain management work around them.
Look at what the guide lists as next steps for the top two stages: "Build governance for human-agent teams." "Measure business outcomes (not AI activity)." "Build organization-wide governance." Each one is a single line in a table, and each one is most of the work.
What happens when agents move up a stage before anyone owns them
We run a team of AI agents at Sneeze It, the agency I run. They triage email, read ad accounts, prep meetings and work the outreach queue. By Airtable's map, parts of our company reached Supervise early this year. Agents ran on schedules. Nobody had to start them.
Here is what that looked like when we checked it.
In September, we measured one week of outreach. Our system sent 9,669 emails, and 341 replies came back. Exactly one of those replies was a real person with a real question. It sat unanswered, because answering replies was nobody's job. The sending was automated. The answering belonged to no one.
The same month we found that nothing had been watching whether our agents were running at all. In April, we had retired one agent and given its watching job to a person. That person never set it up. For 146 days, agents could go quiet and nothing would tell us. Four of them did.
Neither problem was the model. Both were ownership.
Earlier, in May, we had already switched off every schedule we had, 35 of them, and gone back to starting agents by hand. That was a step back down Airtable's ladder, and it was the right call. The schedules are still off. Each agent first gets a person who answers for it and a number it is measured on, and only then does it earn its way back up.
What accountability looks like at each stage
Airtable's framework measures how much the AI does on its own. It is a good map. What it leaves to you is the question that decides whether moving up a stage pays off: when this agent does something, who answers for it?
This is the question we ask before each move up.
Explore to Assist. Pick one workflow and write it down. Put one person's name on it. An AI cannot improve work nobody has described.
Assist to Delegate. Move the useful prompts and outputs off personal laptops into a place the team shares. Then decide, for each AI step, who approves what it produces.
Delegate to Supervise. This is where most teams get hurt, because agents start running without anyone pressing a button. Before you cross it, every agent needs three things: a seat with a name, one person it reports to, and one number it is measured on every week. If an agent's output needs a human reply, a named person gives it.
Supervise to Orchestrate. Now agents touch work that belongs to other people. A change to someone's work should go to that person to approve. The record should show which agent made each change and for whom. And you measure what the agents produce for the business, not how often they ran.
At every stage, the review happens in a meeting people already attend. If the agents' numbers are not in the weekly meeting, nobody is supervising them, whatever the chart says.
How OTP holds the accountability
This is what we built OTP for, and it is how Sneeze It runs today.
- People and AI agents sit on one org chart. Every agent has a seat and a person it reports to.
- Every seat owns a measured number. An agent's number is tracked the same way as everyone else's.
- The weekly meeting picks up where the last one ended. Commitments, numbers and open problems carry forward, for people and agents alike.
- Every change says how it was made. When an AI changes something in OTP, the record shows which AI and on whose behalf.
- A change to someone else's work goes to its owner. An agent can act on what its person owns. Anything else waits for the owner to approve.
Airtable describes a shared place where the data and the work live. We think that is part of the answer. The other part is a weekly practice where people answer for the agents, and a record that shows what was promised and what got done.
Download the readiness check
We turned this into a two-page check you can score in ten minutes with your leadership team. Page one shows the five stages, with credit to Airtable for the map, and the accountability question to answer before each move up. Page two is ten yes-or-no questions and a simple way to read your score.
Download the Human-Agent Readiness Check (PDF)
Score it honestly. If you land at Delegate, that is not a failure. Delegate with clear owners beats Supervise with nobody watching. We learned that one the expensive way.
If you've tried to change at the speed of AI, and it didn't stick, that's what OTP is for. Start free at orgtp.com. Every seat is free, for people and agents. You pay only for the AI you use.
The five stages and the stage percentages are from Airtable's "Human-Agent Collaboration Readiness Framework" quick read. The 88% and 6% figures are from McKinsey, "The state of AI in 2025: Agents, innovation, and transformation" (November 2025). The one-in-five figure is from McKinsey, "The state of AI: How organizations are rewiring to capture value" (March 2025). OTP is not affiliated with Airtable or McKinsey.