A meeting ends. The notes are in one place, the relationship history is in another, the proposal is somewhere else, and the promise I just made is still in my head. The final follow-up might take five minutes to write. Rebuilding the context can take much longer.
That is the kind of work I wanted AI to handle when I began developing Sidekick in December 2025. I incorporated the business and began paid client work in March 2026. The plan was simple: use AI to build and operate a real one-person company, then use what held up in practice to help clients improve their own work.
The company still had to earn revenue while I built it. Every internal improvement competed with proposals, client work, follow-up, agreements, and the ordinary work of running a business. That constraint changed the goal. I was not building the most advanced system possible. I was building the smallest system that created more value than it cost to run.
The founder became the system
Before AI could remove work from me, I had to see how much of the business depended on me connecting everything by hand. Important work moved only when I remembered it, found the source, understood the current state, and decided what happened next.
AI is useful here because it can prepare the decision. After a meeting, it can bring together the notes, relationship history, active commitments, and work already in progress. It can propose a follow-up, next steps, missing information, and record updates. I check the sources, correct the draft, decide what I will commit to, and approve any message before it leaves.
The expensive part is often not making the decision. It is getting everything ready so the decision can be made well.
What AI handles inside Sidekick today
Sidekick keeps its current priorities, decisions, working methods, approval rules, and records in one shared operating core that both ChatGPT and Claude can use. I can switch between the two without rebuilding the company from chat history. Their capabilities are not identical, but the work follows the same source and rules.
Here is the practical split. These states describe what is in use, not what exists as a document or a demo.
| Area | AI prepares | I keep | Current state |
|---|---|---|---|
| Meetings | Context, notes, follow-up, and record updates | Commitments, relationship judgment, and send approval | Operating |
| Sales | Research, proposal drafts, and next steps | Pricing, promises, negotiation, and final approval | In use with approval |
| Marketing | Research, drafts, comparisons, and checks | Positioning, taste, claims, and publication | In use, human-led |
| Legal and finance | Research, issue lists, document review, and coordination | Authority, money, signatures, and professional review | Assisted only |
| Internal operations | Priorities, records, checks, and decision packages | Direction, exceptions, and final decisions | Operating and improving |
The current Inside Sidekick operating map goes one level deeper. It shows which parts are operating, which are still being tested, and where a person must step in.
The shared context also needs care. Old decisions, duplicate instructions, and stale records can quietly make the work worse. I explain that maintenance problem in Why Your AI Gets Worse the More You Use It.
Where the human boundary sits
The tools still fail in ordinary ways. A connection may allow reading but not editing. A scheduled task may time out. Access can change. A request may not trigger the right process. For important work, the system stops clearly when it cannot complete a step reliably.
I also keep approval for consequential actions: external messages, publication, pricing, legal terms, money, account changes, permissions, and client access. AI can prepare a legal review, but it cannot be my lawyer or sign an agreement. It can prepare a public article, but it cannot decide that I should publish it.
Some work stays human because judgment is the value. Relationships, company direction, negotiation, taste, unusual exceptions, and decisions with real reputational or financial weight still depend on the responsible person.
I do not automate a task simply because I can. Some work happens too rarely. Some is faster to handle directly. Some creates more maintenance than it removes. The test is whether the change earns its cost, risk, and complexity. When NOT to Use AI explains how I make that choice.
I learned this by building the more complicated version first
I have tested direct API setups and agent systems that can run longer, connect more deeply, and complete more work without intervention. I would get another layer working, then spend time maintaining connections, changing configuration, watching runs, and finding out why something that worked yesterday had stopped.
The extra autonomy was real. So was the extra system I now had to secure, monitor, debug, and pay for.
Desktop Claude and ChatGPT usually give the individuals and small businesses I serve most of the valuable first uses I see: research, preparation, documents, follow-up, quality checks, and work across familiar tools. They are easy to open, relatively inexpensive, and simple to maintain.
A custom system can do more. That matters when the extra capability solves a proven operating need. At the beginning, it can also make a useful setup harder to afford, understand, and keep running. If you are comparing the common business options, I cover the practical differences in Microsoft Copilot vs Claude vs ChatGPT.
The goal is not maximum automation. It is the best balance of useful work, human control, cost, and simplicity.
The principle transfers. The architecture does not.
Sidekick is still incomplete. Public marketing remains human-controlled. Legal, finance, permissions, and client access have stricter boundaries. I still choose between improving an internal system and doing revenue-producing work directly. That pressure forces every new layer to prove it will remove more work than it creates.
Clients do not need a copy of Sidekick's internal operating system. They need the same discipline applied to their own work. That lesson started with five early owner setups. You can read what those builds changed about my method in What Five Early AI Setups Taught Me About Where to Start.
Solo applies the approach around one owner, executive, or independent professional. Fractional applies selected parts across a business, with clearer ownership, stronger boundaries, and closer ties to company tools.
In both cases, the method is the same. Start with work that should work better. Decide what must remain human. Build the smallest useful change. Put it into everyday use. Expand only after real work justifies the next layer.
That is what an AI-operated small business looks like in practice. The person stays responsible. The work arrives better prepared, so less of the company depends on one founder's memory and available hours. For a plain-English view of the method, see How Sidekick Works.
