There is a quietly provocative paper making the rounds on arXiv, and its title says it all: “Optimization Is Not All You Need.” It is a deliberate jab at the “X is all you need” naming convention that has dominated AI research for years, but the argument runs much deeper than a title punchline.
The authors trace something they call “optimisation culture” — the belief that if you can measure something and improve the number, you have done the job. This mindset did not start with AI. It has roots in the audit culture that spread through public services, healthcare, and education over the last few decades. AI just gave it a shiny new engine.
When You Measure the Wrong Thing
Here is the problem in plain terms. An AI model can be optimised to produce text that looks fluent, scores well on benchmarks, and passes automated checks. But the same optimisation that makes it fluent also makes it confident when it is wrong. The model cannot tell the difference between an unusual but correct answer and a confident mistake, because from its perspective, both are just improbable sequences of words.
Think about what that means for a business relying on AI to write customer emails, generate quotes, or summarise contracts. The tool has been trained to sound right, not to be right. Those are different things, and the optimisation process does not distinguish between them.
What This Means for Your Business
If you are a letting agent using AI to draft property listings, or a contractor using it to write project quotes, the output will look professional. That is the easy part. The hard part is knowing whether the details are correct, whether the numbers add up, and whether the promises being made are ones you can keep.
The paper’s central insight is that optimisation can measure how unlikely a piece of text is, but it cannot tell you whether that unlikelihood is a creative insight or an error. A human being reading the output needs to make that call. The AI cannot do it for you.
This is not an argument against using AI. It is an argument against trusting the polish. The more fluent and confident the output looks, the easier it is to skip the verification step. That is exactly when the mistakes that matter slip through.
The Bigger Picture for Small Business Owners
The paper is also a warning about how the industry sells itself. AI companies compete on benchmark scores, and those scores are real. But a benchmark is a narrow test designed to produce a number. Optimising for that number does not guarantee the tool works well in the messy, varied conditions of an actual business.
Here is a practical rule of thumb: if an AI vendor is leading with benchmark numbers rather than case studies from businesses like yours, treat the numbers as what they are — evidence the model is good at tests, not evidence it is good at your job.
The Takeaway
Optimisation is a powerful tool, but it is not a substitute for judgment. When you use AI in your business, build a verification step into your workflow. Read the output. Check the numbers. Apply your own knowledge. The AI can get you 80 percent of the way there, and that is genuinely useful. But the last 20 percent, the part where you decide whether the answer is right rather than just well-phrased, is still yours.