Every leadership team now hears some version of the same claim: AI will let the company run more of itself. Few can say how much of their own company already does. Ask ten founders how autonomous their operations are and you will get ten adjectives. Ask how many of last week's customer follow-ups closed without anyone chasing them and most will not know.
This guide turns the idea of an autonomous company into numbers you can collect. It defines company autonomy, explains the roles a person can play next to an AI system, sets out a scorecard of company autonomy metrics, and shows how to pair each one with a quality check so the numbers cannot be gamed. It ends with a two-week way to set a baseline and a worked example from a footwear factory.
Key takeaways
- Autonomy is a property of each work stream, so measure quotes, renewals and audits separately before you add them up.
- The core number is the share of work items that close on proof with a person only approving.
- Every autonomy metric needs a paired quality metric, such as reversals, complaints or edits to drafts.
- Set a two-week baseline before changing anything, then track the weekly trend, not a single score.
What is company autonomy?
Company autonomy is the share of a company's recurring work that moves from the first signal to proof of completion while people only set direction, approve consequential steps and handle exceptions. A signal is whatever starts the work: a customer email, a calendar invite, a supplier price change. Proof is evidence that the work happened: the sent quote, the booked meeting, the filed report.
Three points follow from that definition. First, autonomy is measured on work items, the concrete units a company handles every day, never on tools. Owning ten AI agents says nothing about how much work closes without help. Second, autonomy is per work stream. A company can be close to fully autonomous on meeting scheduling and fully manual on pricing, and both facts matter. Third, autonomy is bounded by approval. A sensible company keeps people on the decisions that commit money, customers or reputation, so a perfect score of 100% is neither the goal nor a good sign.
Researchers at the Knight First Amendment Institute make a related point about AI agents. Kevin Feng, David McDonald and Amy Zhang argue that an agent's level of autonomy "can be treated as a deliberate design decision, separate from its capability and operational environment." The same is true of a company. How much work runs on its own is a choice you make per stream, and measuring it is how you make that choice with evidence.
Why measure company autonomy at all?
Without measurement, autonomy is judged by anecdote, and anecdotes push in both directions. One impressive demo makes a team overconfident. One embarrassing email makes it pull back everything. A number per work stream replaces both reactions with a trend.
Klarna's customer service story shows why the number has to include quality. In February 2024 the company reported that its AI assistant had handled 2.3 million conversations in its first month, two-thirds of its customer service chats, and was "doing the equivalent work of 700 full-time agents." Those are volume and cost measures. In May 2025, Entrepreneur reported that chief executive Sebastian Siemiatkowski told Bloomberg the company was hiring human agents again, because the AI chatbots, while cheaper, had resulted in a "lower quality" output. The lesson for measurement is plain: a company that tracks only how much work the machine absorbed will learn about quality from its customers.
The direction of travel also makes measurement urgent. Anthropic's Economic Index report of September 2025 found that "directive" conversations, where users delegate complete tasks to Claude, rose from 27% to 39% of sampled conversations in about eight months. People are handing over whole tasks faster than most companies are building ways to check them.
The five roles a person can play next to AI
The simplest way to describe autonomy for one work stream is to ask what role the person plays. The Knight Institute paper defines five levels by the user's role: operator, collaborator, consultant, approver and observer. Translated to company operations, they look like this.
| Level | Person's role | What it looks like in operations | Where it fits |
|---|---|---|---|
| 1 | Operator | The person does the work. AI drafts or summarises on request. | New or sensitive work with no track record |
| 2 | Collaborator | Person and AI split the steps and hand work back and forth. | Complex work where judgement is spread across steps |
| 3 | Consultant | AI plans and prepares the work. The person gives direction and corrections. | Repeated work with some variation |
| 4 | Approver | AI carries the work to the point of action. The person approves or resolves blockers. | Routine work that still commits the company |
| 5 | Observer | AI acts. The person watches results but does not direct each step. | Low-risk, reversible, internal work only |
For most growing companies, Level 4 is the practical ceiling for anything that leaves the building. A quote, a payment-terms reply or an audit document still gets a human yes. Level 5 suits internal housekeeping, such as filing, tagging or updating a tracker. Map each of your work streams to a level today and you already have a first, rough autonomy profile. The five levels of an autonomous company article applies the same idea to the whole organisation.
Which company autonomy metrics matter?
A level is a label. A metric is a count. The scorecard below uses only things you can count from email, calendar, chat and your business tools. Each row names the question it answers and the warning sign to watch.
| Metric | Question it answers | How to count it | Warning sign |
|---|---|---|---|
| Coverage | How much of our work can the system even see? | Work items detected by the system divided by work items found in a manual sample | Below 80% means every other number flatters you |
| Approval-only close rate | How much work closes with a person only approving? | Items closed on proof where the only human step was an approval, divided by all closed items | Rising fast while reversals also rise |
| Intervention rate | How often does a person have to step in? | Items where a person rewrote, reassigned or rescued the work, divided by all items | Flat for months on a stream you thought was routine |
| Draft edit rate | Are prepared actions usable as they are? | Drafts sent with substantial edits, divided by drafts sent | High edits on the same kind of item, week after week |
| Time from signal to owner | How long does new work wait before someone owns it? | Median hours from the first message to an assigned owner | Long waits on customer-facing items |
| Proof rate by due date | Does work actually finish on time? | Items closed on evidence by their date, divided by items due | Items marked done without evidence |
| Reversal rate | Did the work hold up? | Closed items later reopened, corrected or apologised for | Any rise after you widen autonomy |
| Not-relevant rate | Does the system understand what matters here? | Suggestions owners dismissed as not relevant, divided by suggestions made | Above one in four suggests noise |
Microsoft's 2025 Work Trend Index proposes one more number for the people side: the human-agent ratio. Leaders, it says, must ask "How many agents are needed for which roles and tasks? And how many humans are needed to guide them?" That ratio is useful for planning headcount, but it is an input. The metrics above are outputs. A company with many agents and a low approval-only close rate has bought tools without gaining autonomy.
Why speed and quality must be measured together
Software teams learned this lesson a decade ago. The DORA research programme measures software delivery with five metrics that cover both throughput and instability, and its guide to the metrics states that "speed and stability are not tradeoffs," since for most teams the measures move together. Top performers do well on all five. Company autonomy works the same way. A work stream that closes faster but gets reopened more often has not become more autonomous. It has moved the work to whoever handles the complaint.
Pair every speed or volume metric with a quality metric for the same work stream:
- Approval-only close rate with reversal rate.
- Time from signal to owner with not-relevant rate, so speed does not come from flooding owners.
- Proof rate by due date with draft edit rate, so on-time work is also usable work.
- Human-agent ratio with intervention rate, so fewer people does not mean more rescues.
How long can work run without a person?
A second dimension of autonomy is length: how many steps a work stream completes before a person must touch it. The research group METR measures AI capability in this way. In March 2025 it reported that the length of tasks, measured by how long they take human professionals, that frontier AI agents can complete autonomously with 50% reliability "has been doubling approximately every 7 months for the last 6 years."
Two ideas from that work carry over to a company. The first is to measure length in human time. A follow-up email is a five-minute task. Preparing a vendor audit pack is a two-day task. Track the longest task in each stream that now runs to the approval step without help. The second is that reliability must be stated next to length. Fifty percent is a research threshold, and no operator would accept a quote process that works half the time. Pick your own bar, such as 95% of items closing without reversal, and report the longest task length that meets it.
Common mistakes when measuring autonomy
- Counting tools or agents. The number of AI tools in use is a cost line, not an autonomy measure.
- One company-wide score. A single percentage hides a fully manual pricing process behind a fully automated calendar.
- Measuring only what the system sees. If coverage is low, the approval-only rate describes a small, easy slice of the work.
- Treating "marked done" as done. Without evidence such as the sent email, the autonomy number counts intentions.
- No baseline. Changing tools first and measuring later leaves you nothing to compare against.
- Chasing 100%. Removing approval from consequential work raises the score and the risk together.
How to set an autonomy baseline in two weeks
- Days 1 and 2: list the work streams.
Write down the eight to twelve kinds of recurring work that matter most, such as quotes, sample tracking, renewals, supplier changes and audit documents. Give each one a named owner.
- Days 3 and 4: define proof.
For each stream, agree what evidence closes an item: an email sent, a meeting booked, a file uploaded, a record updated.
- Days 5 to 10: count by hand or with read-only access.
Record every item that starts, who owned it, how long it waited, whether it closed on proof by its date, and whether anyone had to rescue it.
- Days 11 and 12: assign levels.
Place each stream at one of the five roles. Most will sit at operator or collaborator, which is normal.
- Days 13 and 14: choose two streams to move.
Pick the two most frequent, least risky streams and set a target, such as moving from collaborator to approver within a quarter while holding the reversal rate flat.
The first 30 days playbook covers the same start with more detail on choosing sources, and the autonomous company roadmap shows what comes after the baseline.
How GeniOS approaches company autonomy metrics
GeniOS reads the places where a company's work happens, such as email, calendar, chat and the business tools it already uses, read-only to start. It forms the company's goals from its own emails, decks and meetings, and checks every team's pending work against them. That gives the measurement its denominator: the work items the company actually has, not only the ones someone logged.
What reaches a person is one situation, what to do, why, one owner and a date. The owner answers Delegate, Done, Later or Not relevant. Delegate hands work to a person or an AI agent, and nothing is sent until a human approves. Those four answers are the raw material for the scorecard above: approvals, dismissals, rescues and closures.
An item closes only on proof, such as the email that went out, and GeniOS learns from each outcome. Because closure needs evidence, the proof rate and reversal rate rest on records, so a team can see which streams are ready to move up a level and which are not.
Worked example: measuring autonomy at a safety-shoe maker
Names, dates and numbers in this example are made up to explain the idea. Vaultex makes safety footwear in Greater Noida and sells to large brands. Its leadership wants to know how much of its commercial and production follow-through already runs without chasing.
In a two-week baseline, Vaultex lists four work streams and counts every item.
| Work stream | Items | Closed on proof by date | Rescued by a person | Level today |
|---|---|---|---|---|
| Buyer quotes and follow-ups (Harsh, sales) | 18 | 11 | 6 | Collaborator |
| Sample tracking on Line 2 (Anushka, production) | 9 | 7 | 2 | Consultant |
| Audit documents | 6 | 3 | 3 | Operator |
| Supplier price changes | 4 | 1 | 3 | Operator |
The numbers tell a specific story. The Bata quote for 12,000 pairs of S3 safety shoes at $12 a pair, about $144,000, was promised before Friday's vendor review. It was rescued at the last minute because the 8% rise in steel toe cap prices had not reached Harsh, so the price had to be rebuilt. Adidas asked for IS 15298 test reports ahead of its vendor audit, and three of the six documents needed someone to search old email threads. The Reebok samples on Line 2 were the healthiest stream, because Anushka already updates one tracker daily.
Vaultex chooses two streams to move. Sample tracking goes from consultant to approver: the system prepares the status note to Reebok, and Anushka approves it. Supplier price changes go from operator to consultant: every supplier price email is routed to one owner with the open quotes it affects, including any buyer asking for 60-day payment terms, since longer terms on a thinner margin change the decision. The paired quality metric for both is reversals. After six weeks, sample notes close on proof by date in 9 of 10 cases with no reversals, and no quote has gone out on an old cost. That is a measured gain in autonomy, and Vaultex can show it to its partners in one table.
Conclusion
Company autonomy is measurable once you stop treating it as a feeling. Define work items and proof, measure each work stream on its own, and pair every speed or volume number with a quality number. Start with a two-week baseline, move two streams up one level, and watch whether the reversal rate holds. Over a few quarters, that record tells you where people should stay in the loop and where an approval is enough. To see these numbers drawn from your own work, look at the plans on our pricing page or book a call.
Sources
- Knight First Amendment Institute, Levels of Autonomy for AI Agents (Feng, McDonald and Zhang), 28 July 2025
- Klarna, Klarna AI assistant handles two-thirds of customer service chats in its first month, 27 February 2024
- Entrepreneur, Klarna Is Hiring Customer Service Agents After AI Couldn't Cut It on Calls, 9 May 2025
- Anthropic, Anthropic Economic Index report: Uneven geographic and enterprise AI adoption, 15 September 2025
- Microsoft WorkLab, 2025: The year the Frontier Firm is born (Work Trend Index), 23 April 2025
- DORA, DORA's software delivery performance metrics, guide
- METR, Measuring AI Ability to Complete Long Tasks, 19 March 2025