ARTICLE
It's not drift. It's spaghetti.
July 24, 2026
The wrong number
I was recently talking to a friend who runs a service business. She told me a story about her new AI assistant quoting a client the wrong price. It wasn't a wild number, but wrong enough to matter. Her first thought was, "The AI is hallucinating. Or broken." It wasn't. It performed the task exactly as instructed, but it went off track because it couldn't find the right version of the truth. There were three pricing documents floating around, and because it had no way of knowing which one was current, it just picked one, as nothing in its instructions told it otherwise. AI didn't invent the mess. It revealed it.
The compensation nobody noticed
For as long as most businesses have existed, small inconsistencies have been survivable, because a capable human was quietly patching them. Someone always knew which spreadsheet was right. Someone remembered that the policy had changed. Someone had been there long enough to know the real answer wasn't the one on the shared drive.
So, the question follows: why didn't a human just review it before it reached the client? Plenty of businesses already build a review step into their AI workflows. It's a solid safeguard, and it catches a lot. But it wouldn't have caught this. A reviewer facing the same three pricing documents has to guess exactly the way the AI did. Nobody had settled which one was actually current. Review catches an obviously wrong answer. It doesn't catch a wrong answer that looks like it's right.
My friend didn't catch it in the moment either. She was mid school-pickup when the quote went out, and by the time she circled back to check, it had already reached the client. The person best placed to know what's actually current is often the same person with the least room left to give it real attention: raising kids, caring for parents, doing both at once, right in the prime years the business relies on them most.
None of this is a business character flaw. It's what happens when growth and speed eclipse the effort of keeping things consistent, and nobody's noticed until now.
Four problems, same mask
Once you start looking, "AI got it wrong" frequently turns out to be one of four causes.
Stale information. A policy that wasn't updated. Pricing that hasn't moved with the market. Information decays, and until recently a human quietly kept pace with it. A reviewer glancing over an AI's output won't necessarily catch that the number is six months stale either, not unless staying current is somebody's actual job.
Differing versions of the truth. Marketing says one thing, sales says another, and nobody's sure what's current, because they exist in siloed systems that don't talk to each other, or break somewhere in the handoff between them. This is often where the real damage starts: not one wrong document, but several right-sounding ones that disagree.
Instructions that have quietly piled up. Every team adds a new rule, a new exception, a new "just do it this way for now." Nobody goes back and removes the old ones. Eventually the instructions contradict each other, and whoever's executing them, human or AI, is forced to guess which one wins.
The system behaves differently, for no single visible reason. This is the least obvious face of the problem, and often the most consequential. Nothing was edited. Nothing was announced. But something upstream changed, and the output downstream changed with it. This is where the cases that make headlines tend to live: a chatbot inventing a refund policy that never existed, another inadvertently giving advice that was arguably illegal. In both cases, the AI exposed a gap nobody knew was there.
None of these are new. What's changed is that AI agents now act on them immediately, at full confidence, before anyone gets the chance to catch them.
Bigger kitchen, same mess
A bigger kitchen just means more counters to lose things on. It doesn't mean fewer spills. The same pattern shows up in a five-person business as easily as a five-thousand-person one. Smaller businesses just have less redundancy to absorb it when it surfaces, and fewer people to catch it before a client does.
The readiness gap plays out the same way review does: having a policy doesn't count for much if nobody acts on it. Gartner's most recent Australian research found only 17% of employees are what it calls "AI Champions" reaping value, while 56% sit at the other end, using it rarely and without much confidence, and half say they've had no real guidance on how to use it well.1
Keeping a human in the loop still matters here, more than ever, in fact. But "in the loop" only works if that human actually has what they need to catch the problem: the current version, the whole picture, enough time to look properly. Microsoft's own research across 20,000 workers backs this up: the environment around a worker accounts for more than twice the real impact of AI as that worker's own mindset or behaviour toward it.2 Culture, manager support and whether things are actually documented are what matters more than how good any individual is at prompting. It's not a talent problem, and it's not a human-oversight problem. It's a plumbing problem.
Clearing the counter
You don't fix a spaghetti-mess kitchen by wiping the counters once and calling it done. You fix it by throwing out what's gone off, agreeing where things actually live, and giving one drawer, not five, the job of holding what's current. It's not one dramatic clean-out. It's deciding clearly what's true and necessary, then keeping it that way. Four moves make that possible, matched to the four faces already named.
Define it before you automate it. Must-do for stale information and instructions that have piled up. Before handing a process to AI, ask the boring question first — which version is actually current, and who says so? If nobody can answer that quickly, that's the thing to fix.
Make someone responsible for an accurate single source of truth. Essential to close the gap between differing versions of the truth. Not an extra duty tacked onto someone already stretched thin. A named owner — whether it's pricing, policy, or whatever core your business actually runs on.
Treat your documentation like infrastructure, not admin. It's what stops a system from quietly behaving differently for no visible reason. Businesses getting real value from AI are the ones capturing what they learn, clearly and cleanly documenting it, then building it into new ways of working. Not leaving it scattered in someone's head or an old folder. That's not a technology project. It's treating "keeping things straight" as mandatory, not optional.
Look for the four faces before you blame the tool. Stale, split, piled-up, or shifted. The category tells you exactly where to start looking and saves you from fixing the wrong thing. The goal isn't to slow down. It's to get the house in order, because now there's a genuinely good reason to.
My friend fixed her pricing as soon as she discovered the problem. It took her ten minutes once she knew what she was looking for. The hard part was never the fix.
About the Author
Stephanie Silvester is the founder of The UnConsultants, The Agentic CMO Lab and Uncomplicated AI for Business. After 30 years leading marketing inside global organisations, she now helps small business owners make sense of AI without the jargon - as someone who's made the calls and knows what actually matters when you're running the whole show yourself.
Sources
The information on this website is general in nature and does not constitute professional advice. All rights reserved.
Other names and brands may be claimed as the property of others.
© 2026 The UnConsultants Pty Ltd
ABN 87 689 593 333