AI Redefined Reliability for Enterprise. Make Peace With It.

We are running our companies on software that has not finished being built, and we are paying full price for it. The outage numbers show it, the vendors put it in their own documentation and we have lived it over the last few months. Here is what it costs, why this is the new norm and the four things (and a fifth one at your discretion) we put in place to work around it.

That ladder is gone. With AI we are buying the scaffolding, what used to be the very primary release, the alpha and running it as if it had reached general availability.

Unreliable like never before

Look at what reliability means now across the tools your company actually runs on. Ookla counted the days the major AI platforms took a severe hit, and that went from six in the first quarter of 2025 to fifty-one in the first quarter of 2026: Anthropic's Claude 39 of them, Google's Gemini seven, Microsoft 365 Copilot three, and OpenAI's ChatGPT two. The foundation these are built on broke too. One DNS fault in Amazon Web Services in October 2025 drew 315,342 reports in a single day. These are not just fringe startups cutting corners to survive, these are the biggest names in software and they fail often enough that uptime is now something you hope for rather than assume.

While we certainly can't single out Microsoft, an incident that made me realize we are really in uncharted territory came from Microsoft itself. In August 2025 it put an AI function directly into Excel cells: The =COPILOT() function that can do several things. Then it published the warning that goes with it. "Do not use it for financial reporting, legal documents or anything with regulatory implications. And be aware that formula results may change over time, even with the same arguments". Imagine reading that ten years ago in the documentation of an enterprise spreadsheet you were paying for. That is a calculator sold with a note asking you not to use it for arithmetic. Still, I give Microsoft real credit for that warning. Putting it where anyone can find it is more than most vendors would do and it reads to me like the AI race forcing even Microsoft to release before it is ready.

A year later, in August 2026, the function is still not generally available, and the limits around it tell you Microsoft is not fully comfortable with it either. It runs only in the Frontier and Microsoft 365 Insider programs, needs a paid Copilot license, is capped at 100 calculations every ten minutes, will not run in a workbook labelled Confidential, and the warning text has not changed.

Our own experiences

We run Claude across the agency, plenty of it on premium seats, and the monthly bill is big. If you use it heavily, you know they have a mechanism to pace you: a session limit that resets every five hours and a weekly cap sitting on top of it. One afternoon a handful of us, heads down on a hard deadline, blew clean through both. It was very unexpected; we were doing the same work that always kept us at 30% usage max. But desperate, I switched on the extra paid usage, the pay-as-you-go credits that bill at standard API rates and we burned through a lot of it in minutes too. I contacted support, which of course, is a bot. It told me there was a known service issue that caused severe and abnormal token consumption, it was being worked on and a fix was coming soon. So, I asked the obvious thing: "if this is a known problem on your end, refund the extra usage I just bought or credit my account so I can pick up where I left off once you fix it". The answer came back instantly. "That's not within our policy".

Picture saying that to a paying customer of a SaaS product they depended on five years ago. It would not have survived the quarter.

OpenAI is not immune either. One mistake took every company license we hold offline for thirteen days, and the only way back was to repurchase all of our memberships and cut our losses.

I understand why this is happening. These companies are in a race none of them can afford to lose, where being second with something finished is worse than being first with something unfinished, and that pressure runs all the way down into the release schedule and the support policy.

To be fair, this is being worked on and not quietly. The four biggest cloud companies are putting about $725 billion into capacity in 2026, up 77% on 2025, and Microsoft still expects to be capacity-constrained into 2026. Microsoft has also made Copilot multi-model, routing work between OpenAI's and Anthropic's models and falling back automatically when one of them is unavailable. That sounds like the no-brainer winning strategy. However, in our own testing, we saw that 50% of what makes one of these tools good is the instruction layer wrapped around the model, not the model itself. Getting Claude-quality output from a raw model, without the system instructions its maker puts on top, is hard and rarely achieved. So, routing helps, but do not treat it as a solution you can stop thinking about, at least not yet. That is exactly why waiting for them to slow down and get stable is not a plan.

Of everything here, I expect customer service to be the last thing that actually gets better. Capacity gets fixed because competitors are watching and there is money on the table. Support does not, because everyone selling these tools knows we have to use them whether we are happy or not.

So here is where all of this leaves us. We are tying the future of our companies to infant technology that reaches us without the testing we used to take for granted, backed by support that cannot or will not help when it breaks, on terms that bill us for their own failures. And we do it anyway, because the upside is too large to walk away from.

What we are doing about it

The stages did not disappear; they just became ours to run. You are the testing group now, at production prices, and the fine print says so in writing. Which means the stability you used to buy is something you have to build yourself, and the only place worth building it is the layer no vendor controls. Here is what we have in place:

1. Our policy asks everyone to add extra time to delivery deadlines, because at some point the tools will be down.
2. We write everything the work depends on into markdown files: decisions, memory, history, actions, all in plain text we own. If a provider goes down, or a better one comes along, we switch and lose nothing, because none of it lived inside their product.
3. Our prompts and agent definitions sit in those same files, not inside a vendor's tool. Most people move their data and forget this part. A custom GPT, a Project, a Gem or a saved prompt built inside someone else's interface belongs to them, not to you. Keep your instructions as files you own and they move with you.
4. Everyone is trained to assume the answer might be wrong, especially the sources and the citations, and to know how to check it. We also built tools that verify sources and numbers for us, because checking all of it by hand would cost more time than the AI saves.
5. And last but not least, take a deep breath, relax and make peace with this new norm. This is not a rough patch that clears up with next spring's release. It is the standard now, and it will be for a while. Every recovery program on earth opens with the same step, which is admitting the situation is exactly what it is. That is all this is. Say out loud that the tools you run your company on are unfinished, unguaranteed, and going down again this quarter. Then plan like someone who believes it.

I was a freediving addict in my younger years. And this feels very similar to my biggest learning then. You do not negotiate with water. You cannot make the current calmer or the pressure gentler, so you have to become obsessive about everything you can control instead: the training, the breathing, the safety diver, the line, the plan for the moment it goes wrong. You build all of that right up to the waterline, and then you go down anyway.

The tools are extraordinary. The ground beneath them is not solid yet, and pretending otherwise is how you get caught flat. Find the line where your control ends and build everything up to it, because that is the only ground you actually rely on.