Anthropic and OpenAI are here to stay
The model labs look like the durable part of this. Anthropic and OpenAI hold the capital, the distribution and the research, and they are compounding all three. If your plan is to wait until that settles before committing to anything, the wait will cost more than the mistake you are trying to avoid.
The open question is everything being built on top of them. That is where the spending is going, and it is where the least has been proven.
Anthropic and OpenAI are going to win this AI race. It's the same today as it was in the early 2000s, when companies like Microsoft, Apple, Dell and HP persevered through that tech bubble. It was the service companies assembled around them that didn't make it, and their clients paid for it twice. You can spend a great deal of capital with the wrong partner and the wrong underlying philosophy around this great technology.
We have seen this shape before
The last platform shift of this size did not end with the platform failing. Microsoft, Apple, Dell and HP all came through the dot-com crash intact. Some were battered and some compounded for the next two decades, but the companies closest to the technology persevered. The technology itself was never the problem.
What did not survive was the service industry that assembled around them. The large web consultancies of that era, firms like MarchFirst, Scient, Viant and iXL, raised enormous sums, grew headcount faster than capability, and were gone within a few years of the peak. Their clients were left holding half-finished systems and the bill for rebuilding them.
The failure was not the technology. It was that a category of firm formed to sell it faster than the ability to deliver it existed. Enterprise software is not a discipline you pick up during a client engagement.
A foundation assembled from vendors is a foundation you have to test
Building with AI from the ground up usually means assembling a foundation out of parts. One vendor for the database, another for authentication, another for hosting, another for the API layer. Each piece is good. The assembly is the part nobody owns.
Three costs follow from that, and they arrive in order:
Hard to test
Four vendors means four failure modes, and no single place they surface. The integration is the system, and nobody wrote tests for it.
Expensive to test
Proving a foundation holds under real load and real permissions costs engineering months. It is the first line cut from a budget.
Costly to undo
By the time a foundation shows its limits, the applications on it are the business. Replacing it is a rebuild, funded twice.
This is the part we would want a board to understand. The expensive outcome is not a project that fails visibly in month three. It is one that succeeds for eighteen months and then cannot scale, at which point the sunk cost is not the software. It is every process the company has built around it.
Ten years of head start
Co-Wright was not built for this moment. It was built because mid-market companies needed a permissioned, multi-tenant portal sitting over systems that do not talk to each other, and that need is ten years old.
It happens to be the layer AI now needs to sit on. Roles and access groups, single sign-on, multi-tenancy, an API with row and column-level control, MCP servers an agent can plug into. That is the slow part to build and the easy part to get wrong.
Ten years of production use is not a marketing claim. It is ten years of finding out where the problems are: the permission edge case that only appears at the fourth tenant, the reporting query that degrades at a certain data volume, the integration that fails silently rather than loudly. None of that can be compressed. It is the reason a foundation is a poor thing to buy from a firm that started building one eighteen months ago.
It is also what makes the next ten years possible. A foundation that has been extended repeatedly is one you can keep extending. The MCP servers and agent hooks we are adding now are not a rewrite; they are the current layer on top of something already built to take them. That is the difference between a platform that absorbs the next shift and one that has to be replaced by it.
We've been building this foundation for ten years, and it happens to fit where the market is going. We have a head start on the layer AI sits on top of.
The gap between knowing and claiming is widening
The distance between people who genuinely understand this technology and people who can talk about it convincingly has never been wider. It is widening because the talking got dramatically easier and the understanding did not.
There is a second dynamic underneath that one, and it concerns us more. A growing number of engineers are skipping the fundamentals, the years of learning why systems are built the way they are, what a database does under contention, how permissions actually propagate. AI makes it possible to produce working code without ever acquiring that. The code runs. It looks right. It fails in ways its author cannot diagnose, because they never learned the layer beneath it.
That makes engineers who do understand the fundamentals scarcer every year, not more common. It makes them harder to identify in an interview, harder to hire, and harder to keep, and the salaries reflect all three.
For most mid-market companies, assembling and retaining that team outright is not a realistic plan. That is the argument for an embedded partner rather than a hire. Not because hiring is wrong, but because the market for the people you would need to hire is moving against you.
Where that leaves you
You could build a role-based, multi-tenant, permissioned portal yourself. People do, and some of them are right to.
This one is already built. It is running in production for clients today, SOC 2 Type II audited, with the APIs and MCP servers to plug your own agents into. The team that built it is the team that operates it.
Starting from zero is a decision like any other. It is worth being certain it is the one you meant to make.