We're Still Treating AI Like a Fancy Spell-Checker

Now, well into the third year since generative AI exploded into everyday use. Three years of access to arguably the most powerful reasoning and analysis tools in human history — available to any business owner with a browser and a subscription.

The overwhelming majority of small businesses, and many larger ones, are still using it to write emails faster.

They run shallow experiments month after month. They get modest time savings. And then they quietly arrive at this conclusion: "AI is useful… but not transformative."

"We tried it — it's okay for emails, but nothing game-changing."

That conclusion is correct — if you keep asking the same low-leverage questions.

The model does not magically become strategic because you have had it for 18 months or so. It remains exactly what it was on day one: a mirror.

 AI reflects back the quality and ambition of what you bring to it.

Bring shallow questions, get shallow answers. 

Bring a real business problem with real data and a real appetite for an uncomfortable answer — and the conversation changes entirely.

 

The Domain-Specific Model Argument — And Why It Misses the Point

A significant portion of early 2026's AI commentary has argued that the real problem is generalist models, meaning ChatGPT, Claude, Gemini. The thinking is they are too broad for serious professional work.

This is expected from our siloed brains. We have been trained to think in linear fashion. 

A solution was offered: domain-specific models. BioMistral for clinical contexts. Med-PaLM derivatives for radiology. Fine-tuned Llama variants for legal or financial work.

Those tools deliver genuine improvements in their lanes: better accuracy, fewer hallucinations, tighter compliance with domain-specific standards. Clinical notes models trained on medical literature perform meaningfully better on clinical documentation than a generalist model.

The uncomfortable truth is that even when a healthcare system or business fully adopts a domain-specific model, most still do not see outsized gains. Because the model is not the bottleneck. The questions are.

Readiness gaps matter — data quality, integration infrastructure, governance frameworks, workforce skills.

Beneath all of those operational blockers is a cognitive one: people are not asking the model to do the hard, high-leverage work it was built for. They are using a Ferrari to deliver pizza and the Ferrari is not the problem.

 

What the Shift Actually Looks Like

The difference between "AI saved me 30 minutes on emails today" and "AI helped me recover 25% lost revenue this quarter" comes down to one thing: the ambition and precision of the question.

Here is what that shift looks like in practice — in a small business context where the stakes are immediate and concrete.

 

Accounts Receivable

Productivity hack: "Write a polite email to a late-paying client."

Real question: "Here are my last 90 days of accounts receivable aging [paste data].

Identify the top 5 chronic late payers, explain the behavioral patterns you see, and draft a 3-step escalation sequence — with exact wording and timing — that maximizes collection probability while preserving the relationship."

 

Customer Feedback

Productivity hack: "Summarize these customer reviews."

Real question: "Here are 200 recent customer reviews [paste or describe]. Cluster them into 4 to 6 distinct pain points, quantify the revenue risk for each using rough estimates if needed, and recommend the single highest-ROI fix I can implement in the next 30 days with less than $500 budget."

 

Neither of these requires a custom model.

Neither requires an enterprise AI deployment.

They require data you already have, a willingness to paste it into a conversation, and enough ambition to ask a question whose answer might be uncomfortable.

 

Why Most People Never Make the Leap

This is not a technology problem. It is a cognitive and cultural one — and it has four distinct components.

They do not know what a real question looks like in their own business. The productivity hack questions are obvious. The strategic questions require self-awareness about where the actual leverage points in your business are — and most people have not done that audit.

They lack clean, structured data to feed the model. Garbage in, garbage out is not a cliche — it is the operational reality of AI at the business level. If your accounts receivable data lives in someone's head or an unstructured spreadsheet, you cannot ask meaningful questions about it.

The answers feel threatening. "Cut this product line." "Raise prices 18%." "Your top two clients represent 70% of your revenue, and both are showing churn signals." These are not comfortable answers. When the model starts delivering them, many users retreat to the safe shallows of email drafting.

Free public models are tuned for helpfulness, not ruthlessness. The default behavior of consumer AI tools is to be agreeable and non-confrontational. Getting strategic clarity sometimes requires explicit permission — telling the model you want the hard answer, not the polished one.

The moment someone starts asking questions that threaten the status quo, force trade-offs, or demand data-driven foresight, the model stops being a hack and starts being a force multiplier. The problem is that most people stop just before that moment.

 

The Insanity Diagnosis

There is a well-worn observation about doing the same thing repeatedly and expecting different results. It applies here with particular precision.

Running the same shallow prompts on the same repetitive tasks — month after month, across three years of increasingly capable AI tools — and concluding that AI is not transformative is not a technology assessment. It is a question assessment. The technology changed. The questions did not.

The gains that leaders talk about — recovering lost revenue, identifying your highest-risk clients before they churn, reallocating marketing spend to channels that actually convert, catching the operational inefficiency that has been quietly costing you for years — those gains are not locked inside a better model. They are locked inside better questions. Questions you have been avoiding because the answers might require you to change something.

 

Where to Start

If you want to move from productivity hack to force multiplier, the framework is straightforward. It has three steps.

1. Step one: Name your highest-stakes uncertainty. Not your most annoying task. Your most expensive unknown. The thing you have been avoiding quantifying because you are not sure you want to see the number. That is where the leverage is.
   
   
2. Step two: Gather the data that exists. It does not need to be perfect. Rough estimates, structured summaries, historical sales data in a spreadsheet — whatever you have. The model can work with imperfect data and will tell you where the gaps are.
   
   
3. Step three: Ask for the uncomfortable answer explicitly. Tell the model you want strategic clarity, not reassurance. That you want trade-offs named, not smoothed over. That you want the highest-leverage recommendation, not the safest one.
   
   

That is not a complex framework. It requires no additional technology, no enterprise software, no custom model. It requires a decision to stop using a reasoning engine for spell-check.

The gains aren't hidden in the model. They're hidden in the questions we haven't learned to ask yet. What question have you been avoiding?

 

Try This This Week

Name one high-stakes uncertainty in your business — the number you have been avoiding calculating, the client relationship you have not analyzed, the product line you have not honestly evaluated. Gather whatever data you have. Ask the question with the word 'honestly' in it. See what happens.

If you want help crafting the prompt that actually moves the needle for your specific business context, bring the question to the conversation. That is where this work begins.

 

By Mollie Barnett, Strategic Systems Architect • AI Powered Strategies for Modern Business Growth