I do not yet run an officially registered business. I am building a portfolio, tools and processes that should let me launch one without improvising on every assignment. This is an important distinction, because AI will not decide for someone what their business should be. If the business, its processes and expected outcomes are not described, an agent is not being given freedom — it is being given room to make a mess.

What interests me most is not automating a single action, but increasing one person’s agency. Someone can have an idea, skills and the will to act, yet still be held back by the cost of specialist knowledge, a lack of time or a larger organisation’s advantage. AI can lower some of these barriers, but only when it serves to understand the problem, prepare the work and take verifiable action, rather than produce confident-sounding answers.

First, Level the Playing Field — Not Replace the Expert

Employment law is a typical example. An employer can afford a lawyer; an employee often cannot. When a dispute arises, the employee may not even know which terms to search for or which provisions matter. AI can help narrow the problem, identify possible legal grounds, prepare questions and point to documents to check independently. The conversation with an employer or lawyer then starts from a different position.

That does not make a model’s answer the law. Research on language models in legal research shows that they can generate incorrect facts and sources. The problem also occurs in tools with access to specialist databases, although usually less often than in general-purpose chatbots. AI can therefore lower the cost of an initial investigation, but the user still needs to consult the current legislation, check the jurisdiction and version of the provision, and seek professional advice when the stakes are high. (Dahl et al., 2024; Magesh et al., 2025)

I see preparing a business in a similar way. Instead of asking an accountant ten times about basic charges, I can first use AI to draw up a list of taxes, deadlines and assumptions, then build a spreadsheet calculating different scenarios. The expert does not disappear. Their role changes: instead of doing all the initial investigation from scratch, they check the model, spot omissions and take responsibility where their knowledge and qualifications are genuinely needed.

This change can reduce the information asymmetry between an individual and a larger organisation. It does not provide certainty for free. If someone mindlessly copies an incorrect AI answer into a tax return, the cost of a consultation may simply turn into the cost of a correction, a penalty or lost time.

Agency Grows Within a Specific Process

I built an automation in n8n that prepared blog posts, social media content, and text and visual materials for a client. This did not happen because the model “knew marketing”. I had to describe the stages, prepare prompts, connect tools and decide what made an outcome useful. It may be possible to extend this process to audio and video, but until I test it, that remains a possibility rather than a finished feature.

Research on generative assistants shows a similar mechanism in a narrower, better-measured setting. In a study of more than five thousand customer support workers, the tool increased the number of issues resolved per hour by an average of 14%, with much greater improvements among beginners. This is not evidence for an autonomous one-person business. It does, however, show that AI can pass on some good practices and increase human productivity within a clearly defined process. (Brynjolfsson, Li and Raymond, working paper, 2023)

This is the difference between a tool and a business. A business consists of many processes, dependencies, obligations and value judgements. Automating one of them can provide a real advantage, but it does not define the business’s direction, offering or responsibility for the outcome.

Autonomy Is Not Granted With a Single Instruction

I would not tell a new agent to “write an email and send it to every client”. It could generate rubbish, press the button as instructed and damage the brand’s credibility. A safer process is for the agent to prepare a draft using the client’s message and company knowledge, then for a person to read it, edit it and approve sending it.

I treat spending money the same way. If an agent uses up a weekly token allowance in one day, it should not independently conclude that there is money in the account, so it should buy another package. First, the costs need to be understood, the process optimised and a boundary set at which the system stops.

Repeated human decisions can gradually form a model for delegation. If an agent asks for permission a dozen or so times in similar situations, the answers can be analysed, a rule described, edge cases tested and a specific class of decisions approved for autonomous execution. This is not magical model training “in the background”. It is the creation of explicit operational context: premises, thresholds, exceptions, evidence and an escalation path. (NIST, 2024 — risk management and oversight framework)

A new situation does not automatically inherit old trust. Permission to send meeting confirmations independently is not permission to negotiate prices. A model derived from several other decision models also requires separate approval. Autonomy should grow by classes of action, not through a general slider from zero to a hundred.

Some Limits Cannot Be Removed by Information

AI can help design a house, but the design may still require the signature of someone with the qualifications and responsibility for safety. A robot may someday perform some of the physical work, but it cannot grant itself permission to connect the building to the electricity or water network. That infrastructure is shared, and someone is responsible for keeping it running.

Many regulated professions and high-risk decisions work similarly. Knowledge helps people prepare a conversation, assess an offer and detect nonsense, but it does not remove responsibility, formal approval or relationships with other people. An AI-supported one-person business is not an isolated island.

A One-Person Business Can Be Bigger Than One Person

I hope that easier access to knowledge and automation will make people less dependent on corporations and let them compete more often directly through the value they create. I do not know whether this will lead to the breakup of large organisations. Available research shows productivity gains in selected tasks and potentially lower barriers to entry, but the long-term impact on the structure of businesses remains unresolved. (OECD, 2025)

For me, the practical conclusion is simpler. AI can enable one person to handle more of the investigation, preparation and repetitive work independently. The person still sets the direction, takes responsibility for the consequences and decides when an expert or another organisation’s consent is needed. Agency does not grow when an agent does everything. It grows when a person can decide ever more precisely what they can safely stop doing themselves.

Sources

✦Building systems for greater autonomy.