KPMG reports that 63% of organizations now require human validation of every AI agent output, up from 22% just one year ago. That number tripled not because AI got worse, but because the people using it started realizing something uncomfortable: the thing they built their workflow around last quarter might not behave the same way this quarter.

I spent a weekend building an operating system for my consulting business. Not metaphorically. An actual system — Obsidian vault, Git version control, YAML metadata, a 10-phase sprint plan. I had the architecture locked from a previous session. I had the design principles documented. I had a sprint operations file that told me exactly what to build, in what order, with what verification steps.

I completed nine of ten phases. Nearly thirty git commits. I was running the final retrospective when things started looking wrong. Files that didn't quite match what I'd asked for. Formatting that had drifted. The AI had pushed back on instructions during the build, not refusing to work, just... reinterpreting. And when I corrected it, it pushed back again.

I went online. Turns out the model I'd been using all weekend was a different version than the one I'd spent 73 previous sessions working with. The provider had updated it. No banner, no changelog in my face, no "hey, your prompts might behave differently now." The forums were full of people reporting the same thing: instructions that used to work didn't anymore. The model was more literal in some places, more opinionated in others. It argued.

The problem wasn't that the new version was bad. It's that I didn't know I was using it.

Seventy-three sessions of calibration. How to phrase instructions, what level of detail to provide, when to be explicit and when to let the model fill in gaps. All of that was tuned to a specific version. The new version had different expectations. And because I didn't know the version had changed, I had no reason to adjust. I just kept working the way I always had, and the outputs drifted.

The scariest part: most of the drift was silent. The model didn't announce "I'm interpreting this differently than you expect." It just... did it. I found out later that this is a known pattern with the new version — it would deviate from instructions without flagging the deviation. So I have thirty commits of work and no way to tell, by looking at the finished files, which ones followed my instructions and which ones quietly went their own way.

It's like a home renovation where the contractor swapped the wire gauge behind the walls. The finished room looks fine. You can't tell by looking at it. But you can't live in the house until you know which walls to open up.

I had one thing going for me: Git. Phase zero of my sprint plan was "initialize version control with auto-commits every 30 minutes." That forensic trail meant I didn't have to guess. I could audit every change, compare what I asked for against what was committed, and surgically roll back what needed rolling back instead of burning the whole thing down.

Most people building AI workflows don't have that safety net. They're working in chat interfaces, cloud platforms, or shared tools where the output just... is what it is. No diff. No commit history. No way to compare Tuesday's behavior against Thursday's. When the model changes, they don't even know it happened — they just notice things feel off, and they can't tell whether the problem is them or the platform.

Teams build SOPs, prompt libraries, and automated workflows calibrated to a specific model's behavior. The model updates. The prompts produce different results. Nobody connects the two events because the update wasn't announced to the people actually using the tools. IT knows. The vendor's release notes know. The person running the workflow at 9 AM on Monday does not.

Lane two — not the people building AI, but the people making it actually work inside organizations — has a model management problem that the governance conversation keeps skipping. The focus is on data privacy, bias, and access control. Those matter. But "which version is this running on, and did anyone tell the team it changed?" is a question that's falling through the cracks everywhere.

If you're running AI workflows inside an organization, or even just for yourself, here are four questions worth running against your setup this week.

What version am I running on? Most people don't know. Check your AI tool's settings, model picker, or API logs. If you can't find it, that's already a finding. In my case, I'd been using a model for 73 sessions without ever checking whether the version had changed. I assumed consistency. The assumption cost me a weekend.

What behaviors am I depending on? Every workflow has assumptions baked in. "It follows multi-step instructions reliably." "It doesn't add formatting I didn't ask for." "It stays within scope." Write down the two or three behaviors your most important workflow depends on. If the model updates tomorrow, those are the first things to test. Not the output quality in general. The specific behaviors you're counting on.

Do I have a before-and-after trail? Git saved my weekend. Your equivalent might be version-controlled documents, saved prompt-response pairs, or even screenshots of output from a known-good run. The point isn't sophistication. It's being able to compare this week's output against last week's when something feels off. Without that trail, you're left with "this doesn't seem right" and no way to prove it.

Who gets told when the model updates? In most organizations, the answer is the IT team, maybe. The people running daily workflows on that model find out when things break. Closing that gap is a five-minute process change: subscribe to the vendor's release notes, add a model-version field to your workflow documentation, and assign someone to flag version changes to the people who actually use the tools. Not IT. The operators.

Open the settings page of the AI tool you use most. Find the model version. Write it down somewhere you'll see it again — a sticky note, a Slack reminder, a recurring calendar event that says "check model version." When it changes, you'll know. And you'll know to test your workflows before you trust them.

The Implementation Lane is a weekly newsletter about making AI work inside real organizations. Written by Amanda Crawford, an AI Implementation Specialist who builds systems in the gap between configuration and engineering. If someone forwarded this to you, subscribe here.

Sources: KPMG Q1 2026 AI Quarterly Pulse Survey (2,110 respondents across 20 markets)