You Don’t Choose a Model, You Govern a Constellation
IDC ·
A CTO explaining, for the third time this quarter, why the company runs five different models across four teams describes most enterprise AI programs heading into the new year. No one outside the AI group can name all five with confidence. Eighteen months ago, the question was simple: which frontier model to standardize on. Today […] The post You Don’t Choose a Model, You Govern a Constellation appeared first on IDC .
A CTO explaining, for the third time this quarter, why the company runs five different models across four teams describes most enterprise AI programs heading into the new year. No one outside the AI group can name all five with confidence. Eighteen months ago, the question was simple: which frontier model to standardize on. Today the answer is a constellation of small, regional, and domain-specific models, each chosen for a defensible reason and none of them mapped against the others.
What worries the CTO now is accountability: who owns the five models already in production, and what happens to the roadmap when three of them get deprecated in the same product cycle. That’s a governance and org-design question, and right now it’s being handled like a procurement one.
The instinct to default to the biggest, most capable frontier model for every task is the assumption worth retiring. IDC tracked 66 open language models released in 2024 alone , with 34 more in the second quarter of 2025, and the release pace hasn’t slowed. A meaningful share of that growth is smaller, domain-tuned, and regionally optimized models built to do one job well.
That matters directly for a model-selection decision happening this quarter. Model choice should follow the use case. A smaller or domain-specific model, sometimes a regional one, frequently wins on cost and fit for a given task , where a general-purpose frontier model pays for breadth the task never uses. The best model for the job is rarely the biggest one.
That accountability question is getting harder to answer, because the roster it’s about keeps growing. Enterprises are entering what IDC calls a multimodel, multimodal, multiagent era , one that needs evaluation, architecture, and orchestration competencies that didn’t need to exist two years ago. A dedicated GenAI evaluation category has caught up fast enough that IDC’s 2025 MarketScape already assessed 13 vendors in it , a category that barely existed the year before.
That governance work doesn’t end once a model is approved. Approval isn’t the finish line. The evaluation category emerged specifically because the portfolio keeps changing after signature, and testing has to run alongside it, not just at procurement. Your engineering team can track any one model closely. The portfolio as a whole is the harder watch, and right now it’s nobody’s job specifically.
The organizational response to that growth is a named owner. IDC is tracking a new role, the Chief AI Officer , with a formal strategic mandate most org charts didn’t carry two years ago. The role exists specifically to own decisions that used to default to whichever engineering team shipped first. Enterprises without that named owner are accumulating sprawl and data debt , plus the integration risk that comes with both, none of it budgeted for.
Expect this to formalize rather than stay ad hoc: model-selection governance tied to use case, cost, and compliance, continued funding for evaluation tooling as its own budget line, and the Chief AI Officer, or an equivalent, picking up ecosystem-level decisions no single engineering team can make alone. None of that requires waiting for the title to exist on your org chart first.
None of this sits apart from the governance questions already landing on the CFO’s and CISO’s desks. Choosing a model is no longer a one-time technical decision made at kickoff. It’s now tied to what it costs and who’s accountable for it, the same fight already playing out over AI budget ownership and agent audit trails elsewhere in the organization.
Whether or not the Chief AI Officer title exists yet in your org chart, three questions are worth asking now:
You don’t choose a model anymore. You govern a constellation. That starts with naming an owner this quarter, before the next model gets added to the roster without one.
Curious how your model portfolio compares heading into the new year? Check out IDC Quanta for an independent, vendor-neutral read on where your approach stands relative to the market.
The post You Don’t Choose a Model, You Govern a Constellation appeared first on IDC .