AutomationAI & Technology

Before you automate: what a one-person business has to write down first

By Bhavik Sarkhedi, Serial entrepreneur, founder of Solo Not Small

When you run a business of one, the promise of automation is almost irresistible: a tool will take the invoices, the follow-ups and the first drafts off your plate. I felt that pull after I sold Write Right in 2023. I chose not to rebuild a team. The work still had to move, and I was the only person left to move it. 

The mistake I had to unlearn was starting with the tool. If I cannot write down the steps, the exceptions, the inputs and what “done” looks like, I do not have a process that is ready for a machine. I have a habit I have not examined yet, and automating that habit does not give me time back. It repeats the same uneven decisions, only faster, and often in front of a client. 

Why speeding up a messy process makes the mess worse 

A one-person business is full of work that only lives in the founder’s head. You know, roughly, how you chase a late invoice or turn a call into a proposal, and you also know that you do not always do it the same way. A long-standing client gets a softer reminder, and a messy brief gets a longer proposal. A tight week gets a shortcut you would not defend on a calm Monday. 

Those variations feel like judgment when you do them by hand, and some of them are. Many of them are just inconsistencies you have never made explicit. The moment you automate the work, the inconsistency becomes a rule. The soft reminder goes to the wrong person, or the shortcut becomes the default, and it happens every time the trigger fires. 

Official guidance on replacing old systems makes the same point in a different setting. The UK government’s service manual tells teams to use a changeover to take a fresh look at their processes, and not simply re-create the same system with new technology. New software will not repair a process you have not yet described honestly. 

What I write down before I automate anything 

For any task I am tempted to hand to a system, I write a short page that answers a fixed set of questions. If I cannot finish the page, I do not automate the task yet. I keep doing it myself and I keep notes until the page is honest. 

I start with the trigger: what happens that should start this work, such as a form being submitted, an invoice going seven days overdue, or a call ending with notes. If I cannot name the trigger cleanly, the system will either fire too often or not at all. Then I list the inputs I already have, and where they live, looking for facts that already exist in a sheet, an inbox or a folder. If a step requires a fact I would have to invent, it does not belong in an automation. 

I write the steps in the order I actually do them. I also write the exceptions: what changes if the client is new, if the job is rushed, if two rules collide, or if I would normally pause and think. The exceptions are the part most people leave out, and they are the part that causes the most damage when a system has to guess. 

I finish by defining what “finished” looks like, and how I would spot a wrong output in under a minute. If a mistake would be hard to see, I am not ready to let a system produce it without me. The last line on the page is what I will still approve myself. In my own work, nothing goes out with my name on it until I have read it. 

What the research shows when the work is already defined 

In a 2023 study later published in Science, Shakked Noy and Whitney Zhang assigned occupation-specific writing tasks to 453 college-educated professionals, including marketers, consultants, managers, grant writers, data analysts and human resource professionals. Half were given access to ChatGPT. The group with access finished about 40% faster, and independent evaluators scored the quality of their output about 18% higher. 

The researchers were careful about what that finding does not cover. The tasks did not require precise factual accuracy, or context about a particular company’s goals or a particular customer’s preferences, and the team did not fact-check the outputs. Participants were also given explicit instructions, which real work often lacks. 

An MIT News summary of the study is the clearest public account of both the gains and those limits. The study is strong evidence that AI helps with drafted writing when the brief is clear. It is not evidence that you can ship the draft without reading it, or that a vague job will suddenly become a good one because a tool is involved. 

A study of customer support work points in the same direction. Erik Brynjolfsson, Danielle Li and Lindsey Raymond looked at a generative AI assistant rolled out to 5,179 customer support agents at a Fortune 500 software firm, in the working paper Generative AI at Work from the National Bureau of Economic Research. Access to the tool increased issues resolved per hour by 14% on average, and by 34% for novice and lower-skilled agents, with little effect on the most experienced ones. The researchers offer suggestive evidence that the system helped by spreading the best practices of stronger agents. 

The design detail that matters for a business of one is easy to miss. The assistant monitored customer chats and suggested replies in real time. The agents stayed responsible for the conversation, and they were free to ignore a suggestion. That is assistance built on captured practice, not an unsupervised system speaking for the company. 

Those gains showed up where the work was defined, or where good practice had already been encoded, and where a human remained accountable. In a one-person business I am both the person whose practices need writing down and the person who has to stay accountable. If I skip the writing, there is no best practice for the tool to spread, only my improvisation at speed. 

What happens when AI is asked to decide without that foundation 

The risk is not only wasted time. In more open-ended work, a tool can make performance worse when the user cannot tell good advice from generic advice. 

A field experiment with 640 small business owners in Kenya makes that plain. Nicholas Otis, Rowan Clarke, Solène Delecourt, David Holtz and Rembrand Koning randomly gave half the owners access to a GPT-4 assistant over WhatsApp, prompted to act as a business adviser, and then tracked revenues and profits. On average, they could not show that access to the assistant improved performance. 

Owners who were already doing better at the start may have benefited by over 15%. Owners who were already doing worse did nearly 10% worse with the assistant. The difference did not come from the questions they asked, or from receiving different advice. It came from which pieces of advice they selected and put into practice. 

Weaker performers were more likely to act on generic suggestions such as cutting prices or spending more on advertising. Stronger performers used the same kind of assistant to find changes that fit their particular business. The authors’ own summary, How AI Helps the Best and Hurts the Rest, is direct about the implication: in broad, fuzzy problems, generative AI amplifies the role of human judgment. 

I take the mechanism seriously. If I have not written down my pricing rule, my follow-up rule or the kind of client I will not take, I am in a weak position to judge a tool’s suggestions. Writing those rules first is how I practice the judgment the research says I cannot outsource. 

A practical test for whether a task is ready to automate 

Before I switch a task from my hands to a system, I ask myself four questions and I write the answers down. 

Can I describe the steps and the exceptions clearly enough that another careful person could follow them without calling me? If the answer is no, I do not have a process yet. I have a personal method I have not finished explaining, even to myself. 

Are the inputs already sitting somewhere a system can find, or would I be asking a tool to invent a fact? A reminder that uses the invoice date and the client name is ready in a way that a personalised email, built from guesses about a stranger, is not. 

Would I notice a wrong output quickly? A bad date, a misspelt name or a proposal that quotes the wrong scope should be obvious on a short check. If I would only catch the mistake after a client does, the task is not ready to run without me. 

Have I decided what I will still approve by hand? For me that line is anything that carries my name, a fee, a promise or a refusal. The system can draft. I still send. 

If any of those answers is no, I keep the work on my desk and I keep notes until the missing rule is clear. Then I automate the parts that are clear, and I leave the judgment with me. The useful order is a process on paper, then a small automation, then a human check. That is slower on the first afternoon, and it is how a business of one avoids teaching a machine the habits it was trying to escape. 

About the author 

Serial entrepreneur, founder of Solo Not Small, an AI initiative that helps solopreneurs and small business owners scale with AI, Bhavik Sarkhedi is also the co-founder and CEO of Ohh My Brand, a personal branding firm. He has three bootstrapped exits: Dad of Ad (2021), Write Right (2023) and Blushush, the London brand and Webflow studio he co-founded with Sahil Gandhi (October 2026). He is based in Ahmedabad, Gujarat, India. Learn more at solonotsmall.com. 

Related Articles

Back to top button