AI automation and the march of nines

Most of us have now had the same experience.

We’ve opened ChatGPT or Claude, asked it something hard, and were astonished by the answer.

A natural thought follows: If AI can do this, surely we can use it to solve real business problems quite quickly.

And some of this is correct, we solve real business problems every day using AI. But it’s not quite as simple as your LinkedIn feed would have you believe.

One of the core challenges is how accurate the result actually was.

In a chat, 90% accurate can feel magical.

If the answer is slightly wrong, you can ask a follow-up question, rewrite the prompt, ignore the weak part and use the useful part, or you make the final judgement yourself. There is always a human in the loop.

But if an AI is classifying customer requests, updating records, generating compliance documents, routing operational tasks or making decisions inside a workflow, 90% accuracy could be a terrible outcome.

At 90%, 1 in 10 outputs is wrong.

Even at 99%, 1 in 100 is wrong.

For many processes, that is nowhere near good enough. And this isn’t necessarily down to user error. A recent 5,000-prompt study spanning factual recall, citation accuracy and code reference found frontier AI hallucination rates in 2026 sit between 3.1% and 19.1% depending on model, task family, and reasoning configuration.

Perhaps the real challenge is not whether AI can produce impressive results, it clearly can. It’s how reliable the results need to be once AI is operating inside a real business process.

There’s a concept that’s common in technology, especially in areas like uptime, manufacturing and infrastructure, called “the march of nines”.

A system might start at 90% reliability.

Then it improves to 99%.

Then 99.9%.

Then 99.99%.

Each extra nine looks small on paper. But in practice, it is huge.

Going from 90% to 99% means reducing errors from 1 in 10 to 1 in 100.

Going from 99% to 99.9% means reducing errors from 1 in 100 to 1 in 1,000.

Going from 99.9% to 99.99% means reducing errors from 1 in 1,000 to 1 in 10,000.

The numbers get smaller, but the difficulty increases.

The general rule of thumb is that to add an additional 9 of reliability is roughly as hard as everything that came before. Some say ten times harder.

And this is why it’s not as quick and easy as it seems to spin up production AI products within a business.

Deep thought needs to go into your requirements and system design.

To illustrate the point, the right way will depend purely on your requirements and what drives ROI.

For example you might need to:

1. Increase the accuracy of the AI results

One way might be to use extended thinking, this has been shown in research to double accuracy on many tasks. Equally, prompt and workflow engineering, strong guardrails, and high-quality knowledgebases, examples and training materials may all improve results.

The challenge often comes in ‘later 9s’ because, once the low hanging fruit has been done, you cannot reach the next 9 by doing more of what you were doing. You might need new architecture, new monitoring, new fallbacks, new ways of working.

And this is where budgets can really start getting punchy. The answer to getting the best ROI might be, option 2.

2. Use good product design and traditional software best practice so that the “nines” you already have are enough

A 90% reliable model can be enormously useful when wrapped in good product design.

For example, you might attach a confidence score to each output and route anything below a certain threshold into a “drafted, not actioned” queue. A human reviews, approves or amends, and moves on. A task that would have taken an hour from scratch now takes a few minutes.

The non-acceptable 10% is contained, most of the AI’s value is preserved, and the ROI is still excellent.

This might be the optimal balance for you to get the highest ROI. It doesn’t take huge investment, but still drives great staff time savings (even if not 100% eliminated). The cost of investing more to drive further accuracy might not get paid back in the business benefits the system provides.

The same logic may take many forms.

1. You might narrow the job so the model is solving something it can actually solve well.
2. You might let deterministic code handle the bits that must be right, totals, payments, who an email goes to, and reserve the AI for the bits where approximately right is genuinely fine.
3. You might have a second model check the first.

There is no single right answer, but these decisions are often the difference between an immensely valuable system that drives huge savings for a business, and a failed project.