The honest risks of using AI wrong.

AI is not coming for the world. It is already in your business in small ways. The real risks are not science fiction. They are boring, specific, and already happening to people who deployed too fast.

Risk one: the model is confidently wrong

Hallucination is the polite word. The plain word is fabrication. The model generates text that reads correct, cites sources that read correct, and is not correct.

This is not a rare edge case. Stanford researchers tested specialty legal AI tools. Lexis+ AI returned incorrect information more than 17% of the time. Westlaw AI-Assisted Research hallucinated more than 34% of the time. When researchers asked general purpose models specialty legal questions, hallucination rates ran between 69% and 88%.¹

In medicine the picture is similar. A 2025 study measured large model hallucination rates on clinical case summaries at 64% with no mitigation. Structured prompting brought that down to 43%. The best performing model still hallucinated 23% of the time.²

This matters in your business if you use AI to draft anything that ends in a client deliverable, a tax filing, a contract review, or a medical or legal opinion. The model will sound certain. It will be wrong in ways a junior team member would never be wrong, because a junior team member knows when they do not know.

The fix is not avoiding AI. The fix is treating its output as a draft and having a human verify the parts that matter before anything leaves your hands.

Risk two: bad data in, bad system out

The most underrated risk is not the model. It is the data the model is fed.

In early 2025 three of the largest US health insurers were sued for using AI to deny medical claims. One filing cited an algorithm that reviewed and rejected more than 300,000 claims in two months, spending an average of 1.2 seconds per claim.³ Patients were discharged too early. Some later died.

The model was doing what it was told. The system around it was the failure. No human review. No appeal flagged. No threshold for when a denial should bump to a person.

A small business will never be sued for 300,000 wrongful denials. But the pattern repeats at smaller scale every day. A CRM that auto-tags leads as cold based on bad activity data. An accounting tool that miscategorizes invoices and quietly distorts margins. A copywriter that pulls from your old positioning instead of your current one.

Garbage in is not just a quality problem. It is a credibility problem. Once the system is making decisions on bad data, the cost of cleaning up after it is higher than the cost of getting the inputs right at the start.

Risk three: real money on the table

Forrester estimated that AI hallucinations cost businesses globally $67.4 billion in 2024. That works out to roughly $2,130 in losses every second.⁴ Most of it is invisible: bad recommendations followed, returns processed in error, errors compounding in downstream systems.

A public hallucination case database has documented 1,455 court filings where AI generated fake citations or fake facts. 486 are recorded with full details. 324 of those are in US courts. Sanctions exceeding $10,000 have been imposed in at least five recent cases.⁵

The lawyers who got sanctioned were not idiots. They were practitioners who used AI to draft a brief and did not check the citations before filing.

Risk four: job displacement is real but uneven

The honest version of the displacement question is not a single number. It depends on what you are measuring.

The World Economic Forum's Future of Jobs Report projects 92 million roles displaced globally by 2030 and 170 million new roles created. Net positive on paper.⁶

McKinsey estimates that today's technology could automate roughly 57% of current US work hours, though "could" is doing a lot of work in that sentence. Across the entire working population, McKinsey says 30% to 50% of current work activities could be automated.⁷

The OECD is more conservative. Their task-level analysis concluded that only 9% of jobs in OECD countries face a high risk of automation.⁸ The gap between McKinsey's 57% and OECD's 9% is not noise. It is the difference between automating tasks within a job and automating the job itself. Most jobs are bundles of many tasks, and the easy tasks get automated faster than the whole role disappears.

Inside a job, the affected categories are clear. Administrative and office support tasks: about 46% automatable. Manufacturing: 45%. Customer service: 41%. Data processing: 38%. Basic financial services: 37%.⁶

Telemarketing and data entry roles are at the top of the at-risk list. Sales managers, skilled tradespeople, and roles that require physical presence are at the bottom.

For a small business with a small team, this is less abstract. If you have an admin assistant doing scheduling, intake, and basic data entry, parts of that job will get faster with AI. The right move is usually not to fire the assistant. It is to widen their scope using the time AI gives them back.

Risk five: over reliance on the tool

The least talked about risk is the one I see in the diagnostic most often. Owners hand over judgment to the model because it sounds confident.

The owner asks the AI what their pricing should be. The AI gives a number. The owner uses the number. The owner stops asking the question themselves.

Six months later the owner cannot explain their pricing to a client who pushes back. Because the owner never built the reasoning. The model built it and the owner copied it.

This is the most expensive kind of AI risk because it is silent. The number was fine. The capability loss inside the owner is the cost.

How to use AI without these risks landing on you

Three rules.

Verify every output that matters. If the cost of being wrong is high, a person checks the work. No exceptions.

Audit your data before you automate anything. If the inputs are dirty, the system will be worse than the manual process it replaced.

Use AI to expand your thinking, not replace it. Ask the model to challenge your reasoning, list counter arguments, or surface what you missed. Then decide yourself.

The owners getting hurt by AI right now are not the ones using it. They are the ones using it without these three guardrails.