Free AI Isn't What It Used to Be: ISACC Institute's Founder, SerVanes Wilson Sr., on the Real Cost to Solopreneurs

For a solopreneur, human review is essential, not optional, and SerVanes Wilson Sr. of the ISACC Institute knows that firsthand.

One piece of advice he no longer gives solopreneurs and small businesses: rely on several chatbots' free tiers instead of paying for a tool.

"I used to tell solopreneurs and small businesses to use multiple free AI tools versus paying for them. At the time, it was good advice. The free tier output and token limits increased productivity for them and allowed for previously ignored tasks to get completed. Then the free tools seemingly got worse, and the advice followed them," he wrote in response to questions from Cranberry Journal.

"In my own use, better prompting didn't restore the quality I had come to expect," he added. "I also heard similar complaints from other business owners who had been rotating among free tools. That's not a controlled study, but it was enough to make me stop recommending the strategy as a reliable way to get work done. Paying for access didn't remove every interruption, either: I still hit usage limits during routine tasks."

Wilson is the founder of the ISACC Institute, a consulting and design firm, and describes himself as a generative AI specialist and spokesperson with Outlier AI, an OpenAI beta tester and a creative strategist with more than 20 years of experience. The ISACC Institute markets strategy, branding and business consulting to corporations, small businesses and solopreneurs. His current argument is narrower than his old one. Before a small firm automates anything, he says, it should decide where a person has to check the output.

A market of one-person firms

That question matters more for small companies than for large ones, because many have no one else to do the checking. The SBA Office of Advocacy's 2025 Small Business Profile counts 36.2 million small businesses in the United States. By our arithmetic from its tables, 29.8 million of them, or 82.4 per cent, had no employees in 2022, the year of the underlying Census data.

The share is higher among the owners Wilson often has in mind. The same profile counts 4.63 million Black-owned businesses, and 4.44 million of them, or 95.8 per cent, had no employees. For White-owned firms the share was 82.1 per cent.

When the owner of a firm without staff hands a customer email, a price quote or a contract summary to a chatbot, the owner is also the quality-control department.

Two cases that set the terms

Wilson leans on two well-documented failures.

The first is Zillow. In November 2021, the company said it would wind down Zillow Offers, its business of buying and reselling homes, after concluding that further scaling was too risky and too volatile. In its shareholder letter it admitted that it had been unable to forecast home prices accurately, missing in both directions. Bloomberg News reported at the time that Zillow planned write-downs of as much as $569 million and would cut about a quarter of its staff. The final inventory write-down for 2021, reported in its annual filing, was $407.9 million.

Wilson notes that this was a pricing model, not a chatbot. "The lesson for entrepreneurs isn't that AI has no place in business," he wrote. "It's that a system's output should never be mistaken for certainty."

The second case was smaller and closer to a one-person shop's risk. In February 2024, British Columbia's Civil Resolution Tribunal ordered Air Canada to pay a passenger CA$812.02 after the airline's website chatbot told him he could get a partial refund under its bereavement policy after buying his ticket. Air Canada argued the chatbot was responsible for its own actions. Tribunal member Christopher Rivers disagreed. "It makes no difference whether the information comes from a static page or a chatbot," he wrote, as reported by CBS News.

The award was modest. The principle was not. A business owns what its software tells customers.

Where the check goes

Wilson's consulting pitch is an audit of the whole workflow before any tool is switched on. His firm, he wrote, helps clients find "where it can help, where their expertise needs to verify its work, and where human review is essential before a decision or message reaches a customer."

The table below is ours, built from the documented cases above. It is not Wilson's framework.

Wilson's own rule follows the stakes. Automate low-risk steps freely, he says, and put stronger review where customers, money or important decisions are involved.

When the tool changes under you

The third row explains why Wilson changed his advice on free tools. In an April 23 post, Anthropic traced complaints about its Claude Code assistant to three separate changes. It had lowered the default reasoning effort to reduce latency, shipped a caching bug that made the assistant drop its earlier reasoning, and added a system prompt line to cut verbosity. Because each change hit different users on a different schedule, the company wrote, the effect looked like broad and inconsistent decline. All three were resolved by April 20.

OpenAI's tiers differ in quieter ways. When it made GPT-5.5 Instant ChatGPT's default model on May 5, the company said the update would give tighter, more concise answers and fewer hallucinations. A June note on the same page said responses on the free tier would draw on a reduced set of a user's past chats.

Wilson's reading is that limits and context, not price, became the constraint. "The expensive, single-provider, deep-context workflow started beating the free, multi-model, shallow-context workflow for anything past a few hundred tokens of sustained reasoning," he wrote. He frames it as a trade-off and an engineering choice, not as anyone setting out to make products worse.

Research gives the same caution a longer horizon. A 2024 study in Nature by Ilia Shumailov and colleagues found that models trained indiscriminately on machine-generated text develop irreversible defects and lose the rarer parts of what they learned. A paper presented at ICLR 2025, by Jiaxin Wen and co-authors, found that training models on human approval made them more persuasive to evaluators without making them more correct. Output that sounds right still needs checking by someone who knows the subject.

The partnership pitch

Wilson's thesis is that the advantage goes to people who pair AI with human judgment, creativity and empathy. For a one-person business, he says, that starts with a clear business problem and a workflow built around human checkpoints, not automation for its own sake.

Cranberry Journal has reported on how small firms are adopting generative tools for paperwork rather than decisions, and on why companies struggle to tell whether their AI works. Wilson's revision adds a practical point for the smallest of them. The cheapest setup is not always the one that costs least once the owner's own review time is counted.

Mr. Wilson's quotations are from his written answers to questions from this publication, and his titles, credentials and work history are as he describes them. The ISACC Institute's services are from its website, read on 3 November 2026. Business counts are from the SBA Office of Advocacy's 2025 Small Business Profile for the United States, which draws on 2022 Census data, and the shares are this publication's arithmetic from its tables. Zillow figures are from its 2 November 2021 shareholder letter, its Form 10-K for 2021 and Bloomberg News, as published by National Mortgage News. The Air Canada ruling is as reported by CBS News and Legal 500. The Claude Code account is from Anthropic's 23 April 2026 post. OpenAI's statements are from its 5 May 2026 GPT-5.5 Instant announcement. Research findings are from Shumailov et al., Nature, 24 July 2024, and Wen et al., ICLR 2025. The table is this publication's own.

Topics businesssmall businessartificial intelligence