One entrepreneur, a whole office: how AI changes small business

A business may have just one employee: its owner. Yet that person has many jobs—sales, marketing, purchasing, project management and the evening paperwork. The constraint is often not a shortage of ideas. It is having the same person carry every idea through to completion.

AI changes that arithmetic. It can draft a proposal, find a clause in a document, prepare questions about a contract and help organise a working day. The owner gains additional capacity. That capacity needs organising, however: a random collection of chatbots can create more checking than useful work.

News from the US and China offers a reason to discuss one-person businesses. Here is what the figures actually show, which tasks AI can support, and how AI Office can help an entrepreneur build a manageable workflow.

On September 1, 2026, US Treasury Secretary Scott Bessent linked AI to growth in small-business startups. That is a prominent public assessment of the technology's potential, rather than statistical proof of what caused registrations to rise. Scott Bessent: published September 1, 2026 video clip on AI and small-business startups ↗

Census Bureau data distributed through FRED show 5,671,836 US business applications in 2025, compared with 3,498,990 in 2019 and 4,356,498 in 2020. These are sums of monthly observations without seasonal adjustment. The surge began during the pandemic, before widespread adoption of generative chatbots. FRED / US Census Bureau: Business Applications, not seasonally adjusted ↗

The revised seasonally adjusted series records 576,512 applications in July 2026—roughly 577,000—and 531,728 in August. These figures were checked on October 4. Unadjusted annual totals and seasonally adjusted monthly observations use different comparison bases. FRED / US Census Bureau: Business Applications, seasonally adjusted ↗ US Census Bureau: August 2026 Business Formation Statistics release ↗

An application is not yet an operating company, profit or a job. The US series counts filtered applications for an Employer Identification Number, or EIN. Its total cannot establish that most new businesses consist of exactly one person or were created because of AI. US Census Bureau: Business Formation Statistics definitions ↗

China registered 16.194 million new 个体工商户 in 2025, according to the State Administration for Market Regulation. This individual-business category also permits family operations. The figure measures new registrations rather than net growth in operating businesses; it is not directly comparable with US applications. SAMR: 2025 individual-business development, March 26, 2026 ↗ SAMR: individual-business registration rules, including family operations ↗

These figures show the scale of interest in starting a business. They do not mean millions of people have already replaced their staff with bots.

A July 2026 Mercatus Center paper finds indications of rising independent work in some sectors more exposed to AI. Its findings are preliminary and descriptive; a causal relationship has not been established. Mercatus Center: AI and the Rise of Independent Work, July 17, 2026 ↗

That scenario is possible in particular occupations and companies. It does not establish that AI is already shrinking the entire labour market. In a September 28, 2026 speech, Federal Reserve Governor Lisa Cook notes limited evidence of large-scale employment changes while discussing risks for particular groups and opportunities for small businesses. Federal Reserve: Lisa Cook speech on AI, September 28, 2026 ↗

AI can strengthen an existing job, support a new business line or help test an independent service. It reduces the effort involved in some operations; it does not automatically create demand, customer trust or a cash reserve. Starting a business remains a choice with responsibilities, rather than an obligatory response to labour-market change.

Start with a specific operation: “prepare three versions of an offer using the verified service description,” rather than “we need an AI marketer.” This makes the input, output and acceptance criteria clear.

These are possible workflows, not a list of ready-made AI Office features. Each requires appropriate models, data and integrations. Writing an email, sending it to a customer and approving a discount involve three different permissions.

“An AI accountant” is too broad a description. AI can organise documents and flag discrepancies; calculations and reporting need reliable tools and, where necessary, a professional. A list of contract risks helps prepare a legal consultation rather than replacing one.

Consider a fictional small business supplying equipment and coordinating installation. Its owner has a catalogue, calculation rules, supplier documents and several concurrent projects. A customer requests a proposal for a specified set of equipment and quantities.

The assistant retrieves relevant information from authorised documents and helps prepare the proposal. The owner checks requirements and product selection. Prices come from the current catalogue; software applies discount and total-calculation rules. A disputed equipment substitution requires a separate human decision.

After checking, the owner approves a specific proposal version. The current AI Office prototype supports a deliberately limited transition: an approved proposal becomes a locally assigned task. Customer delivery, CRM updates, payments and supplier orders do not happen automatically; they need separate workflows or integrations.

The benefit is connecting documents, calculations and the next step without repeatedly moving context between chats. The assistant still cannot independently promise a deadline, reduce a price or make a financial commitment.

You do not have to become a professional marketer, lawyer and accountant simultaneously. You do need to understand your business's main decisions and recognise when a specialist is necessary.

• Customer and offer: who needs the service, which problem it solves and why someone would pay for it.
• Sales: the difference between interest and an order, how to check promises and what counts as a completed deal.
• Economics: revenue, costs, margin and cash flow; accounting profit cannot pay a bill by itself.
• Agreements: scope, payment terms, deadlines, responsibility and the limits of your authority.
• Quality: an acceptable result, what must be checked against a source and which mistakes are unacceptable.

The central skill is defining a task that can be checked. “Do good marketing” has no clear acceptance criterion. “Prepare three versions of an email for this service, use only approved features and make no unsupported delivery promises” can be evaluated.

Recognising missing information matters too. A confidently invented answer about a warranty or equipment compatibility can cost more than the time saved. In these situations, the assistant should ask for clarification.

Price lists, supplier terms and proposal versions are a small business's memory. When that memory is scattered across messages and folders, its owner must reconstruct the picture repeatedly.

AI Office is being designed as a workspace for documents, models and controlled actions. It is currently a prototype with a limited workflow. Four areas are particularly relevant to entrepreneurs.

First, a knowledge base. Text-based PDF, DOCX, TXT and Markdown are supported, with retrieval respecting permissions and source versions. An answer's basis can be checked against documents. Scan recognition is outside the current scope.

Second, proposals: an Excel/CSV catalogue, reviewed line items, software-based calculation of RUB amounts and DOCX/PDF export. Spreadsheet formulas are not used as trusted calculations; the model does not set prices itself.

Third, tasks and approval: assignees, deadlines and working task lists. A person approves the specific proposal version before it becomes a local task. The next step retains a clear responsibility.

Fourth, an analyst council. The basic prototype examines a question through several roles, including a critic—for example, finance, operations and contract terms. Every role uses the same configured model. These are different analytical perspectives, not independent experts. The owner decides; real-model quality must be evaluated separately from the demo mode.

Email, CRM, accounting and payments are not currently connected. Integrations require separate work with access rules and acceptance criteria.

Contracts, purchasing prices and customer records are sensitive regardless of team size. Sending an entire archive to an external chatbot requires a deliberate decision.

AI Office's architecture allows a local model and knowledge base. The prototype supports demo mode, Ollama and a compatible HTTP provider. Quality, speed and hardware compatibility need testing in a pilot; a local computer alone does not provide security.

Cloud services can help with authorised tasks, but the data route must be understood. A published service description and a private contract need different rules. Calling text “anonymised” does not establish that commercial secrets have been removed.

Define the process and workload before selecting hardware. An existing compatible computer may be sufficient for an initial test; buying a powerful server before evaluating the task is unnecessary.

Control includes permissions, updates and recovery. The prototype has manual encrypted offline backup; automatic scheduling and off-site storage still need development. Losing an archive can stop a small business too.

We begin by examining the process. Which documents enter it, who checks the result, where do amounts come from, and what does “finished” mean? We can then choose one workflow, prepare the data, configure an assistant and agree pilot criteria.

A trading business might test catalogue-based proposals. A consultant might start with retrieval across their own materials and drafting. A small installation business might examine the connection between documents, approval and tasks. These are possible starting points, not claims about completed deployments.

Customer enquiry handling, marketing, CRM and document workflows can then be designed separately. The label “AI salesperson” does not mean every sales channel is already connected.

For entrepreneurs working on physical sites, AI Office Field is a relevant development concept: assignments, site photos and voice messages, completion evidence and document preparation. It is a concept rather than a connected prototype feature. It could become a separate pilot where collecting field information is the actual bottleneck.

Alongside configuration, the owner needs operating rules, access boundaries, an error-correction procedure and a way to continue working when AI is unavailable.

Take an illustrative example, not an AI Office measurement. An owner prepares 40 comparable proposals per month. Before automation, each takes 60 minutes including checking. Afterwards, preparation takes 20 minutes and mandatory human review another 10.

The difference is 30 minutes per document, or 20 hours per month. This is potential time released under the stated assumptions. Proposal complexity and the number of corrections change the result.

Those hours can go to sales, customers or rest. An economic assessment must include setup, hardware, software expenses, cloud calls, support and the owner's time. A financial benefit requires higher margin, lower paid costs or fewer losses.

Measure the entire workflow, including rework. A draft that needs rewriting from scratch does not make the business faster.

Repeatable documents, clear rules and a verifiable result are a sound basis for automation. A catalogue with known prices is easier to control than “find a niche and generate sales.”

If demand is unconfirmed, test the offer in conversations with customers first. Producing more content for an unwanted service is unhelpful.

Where work requires a licence, professional qualification or physical inspection, a bot does not remove that requirement. High-consequence errors need an appropriate specialist and stricter review. Unprepared data may make organising catalogues, documents and rules more useful than launching ten agents.

AI does not eliminate contractors. You can delegate preparation to a model and complex legal, accounting or technical work to people. Minimal headcount should not become a goal in itself.

• Week one: choose one repeatable process and measure time, volume, errors and rework. Prepare comparable examples with clear correct outcomes.
• Week two: assemble current documents and rules, define permitted data and actions, and identify approval points. Configure a minimal working scenario.
• Week three: run a limited pilot. Review every result and count total working time, including corrections and owner checks.
• Week four: compare quality and costs with the original process. Decide what to continue, what to change and which task to connect next.

This is a sample plan rather than a promised implementation deadline. An unsuccessful pilot is informative too: the task may need better data, a different model or, for now, the familiar method.

A one-person business with AI is an opportunity to extend your capacity while keeping deliberate control. The owner still needs customers, competence and responsibility for outcomes. Assistants can take on some preparation, retrieval and routine work.

To test the approach, start with one process that regularly consumes your time. We can examine it together, match it to AI Office's capabilities and propose a limited pilot with clear quality and control criteria.