On July 2, Microsoft launched Frontier, a new unit it's backing with $2.5 billion and roughly 6,000 engineers, consultants, and specialists whose entire job is helping enterprises implement AI, according to CNBC. Two days earlier, AWS had committed around a billion dollars to embed its own engineers inside AI customers. Two of the largest cloud vendors on earth stood up implementation arms in the same 48 hours, and that timing is the real story.For years, the industry sold the model as the finish line. The pitch was that a capable enough system would more or less deploy itself, and that a company's main job was to get access to the best one. These billion-dollar bets on people are the same companies quietly conceding the pitch was wrong. If the model were the hard part, you wouldn't need 6,000 people to make it work.
Why enterprise AI actually fails
The evidence had been piling up for a while. MIT's Project NANDA found that 95% of enterprise generative-AI efforts delivered no measurable return, and the reason was rarely the model itself. The failures clustered in the unglamorous work that comes after you pick one: connecting it to the systems a business already runs on, redesigning the workflow around it, handling security and compliance, and getting people to change how they actually work.That's precisely the part Microsoft's Frontier and AWS's Forward Deployed Engineering are built to sell, because it's the part where enterprise AI breaks. Microsoft was fairly explicit about it, positioning Frontier as an answer to the pilots-that-go-nowhere problem. When the company that makes the model spends billions on people to deploy it, the message to a CIO is hard to miss: the model has commoditized, and the value moved to everything around it.
The takeaway for a CIO
The useful thing to take from this is validation. If your AI program has stalled somewhere between an impressive pilot and a system anyone relies on, that rarely means you chose the wrong model. It's the normal place enterprise AI gets stuck, and the vendors just spent a combined $3.5 billion confirming it. The bottleneck was always implementation, and budgeting an AI program as if the model were the hard part is how companies end up with capable tools and no outcomes.
The catch in buying it from the vendor
There's a wrinkle worth naming before you sign up for a vendor's deployment arm. A team that works for Microsoft or AWS has one structural incentive underneath everything it does, and that's to grow its employer's platform consumption. The engineers are genuinely excellent, and they are also, in the end, paid to embed you deeper in Azure or AWS. That isn't a scandal, it's just how incentives work, and it means the architecture a vendor's team recommends will tend to be the one that runs on more of that vendor's stack.For some companies that tradeoff is fine. For a $100M+ enterprise trying to stay portable and avoid lock-in, it's worth knowing that the implementation partner and the platform salesman are, in this case, the same organization.
How True Horizon does it
This is the work we've been doing all along, and it's the layer these announcements just validated. We build AI around how your company operates rather than around a vendor's roadmap: connecting it to the systems you already use, handling the security and compliance, redesigning the workflow, and running it after launch. The difference is that we don't have a platform to sell you, so our only incentive is getting your AI to work and keep working, even when that means the best model for a job isn't the one a hyperscaler would steer you toward.
What to do with this
Treat the news as permission to reprioritize. Budget and staff the implementation as the hard part of your AI program, because the people who sell the models just told you it is. Keep a layer between your workflows and any single vendor's stack so switching stays possible. And when you bring in help to deploy, ask the one question these announcements make unavoidable: is this team optimizing for my outcome, or for their platform's growth?The model was never going to be the thing that set your AI program apart. What you build around it, and who you trust to build it, is.If you want implementation help with no platform to sell, take our AI assessment and we'll map where your AI is actually getting stuck.
