By 2030, according to Gartner, 90% of enterprise AI agent workflow steps will run on open-weight models, meaning models companies can download, inspect and run themselves. The forecast comes from a document by analysts Mark Petty and Raymond Paquet, titled "Open-Weight Models Will Run 90% of Enterprise Agentic Workflow Steps", published last week.
The argument starts from a simple figure: nine in ten steps in an agentic workflow are routine. These include classification, tool calling, passing data, validating and reconciling it. A typical operational workflow contains hundreds of such executions, and the analysts say open-weight models can handle them at parity with closed models for a quarter of the price of the cheapest closed model. For the rare steps requiring complex reasoning, frontier models would remain in use.
The most favourable numbers compare against a frontier model run on every step: there, open models cost as little as a twentieth. The analysts admit, however, that current open models tend to be more verbose and retry more often, so the real saving per task is closer to five or six times.
What Gartner considers most important is not cost. The document notes that an enterprise's workflow history weighs heavily in the valuation of a SaaS company, and that every routine step sent to a frontier API moves part of that asset outside the company's perimeter. Open weights, Petty and Paquet argue, preserve control; the saving is a bonus.
The analysts therefore expect enterprises to adopt mainly domain-specific open models, which work well where a few very stable steps run millions of times a month. The document also says frontier models will keep serving as the engine of research and development, treated as capital expense, while open models become an operational expense once frontier models have done the heavy lifting.
On where value lies, the reasoning is that most models are becoming commodities, while the value sits in harnesses: the tools that run and manage agents and route each task to the most suitable model without users noticing. A good harness saves money, imposes strong guardrails and governs agent fleets; a weaker one leaves users more work.
The forecast concerns enterprise operational workflows, not the whole universe of AI agents. For frontier model providers, which are spending billions on infrastructure, the implicit picture is that revenue would concentrate on the hardest tenth of steps. Whether enterprises will actually shift their workloads remains to be seen.
The topic fits a visible trend: Aiven says it cut its internal token spending by 80% by moving work to open-weight models, while IBM's CEO says open-weight models are now good and cost a tenth. American companies are also moving away from exclusive reliance on frontier models because API costs have become unsustainable.
The open question is how much of this shift will depend on the real quality of open models on repetitive tasks, and how much on price pressure.
No comments yet. Be the first!