Varick's FDE lead: most enterprise agent failures come from missing context, not from using an older model

Daniel Kornum, who leads the forward deployed engineering team at Varick, a company that helps large enterprises adopt AI agents, argued in a post on X titled "You Don't Need Better Models" that most operational agent work does not need the newest frontier model. Digital Today reported the post on October 20.

He used a hypothetical company where 40 staff check 30,000 supplier invoices a month. If people and an agent process the same invoices for a month, results differ on about 6% of them. In his breakdown, most of those mismatches come from missing context (such as billing terms agreed by email years earlier), exceptions that were never written down, and bad system data. Only about 100 cases, or 0.3% of invoices, were model errors, and a newer model would fix only about half of those. He also put numbers on the cost: at 40,000 tokens per invoice, a $10-per-million-token model costs $12,000 a month against $2,400 for a $2 model, and every model swap means re-tuning prompts and re-testing the agent.

His suggested alternative is to choose a model per task, build an internal evaluation set of 500 to 1,000 real invoices including the exceptions, and put a routing layer between agents and models so the deployment isn't tied to one vendor.

For FDEs, the point is that the job is mostly the unglamorous part: mapping how a customer's workflow really runs, digging out the context that lives in people's heads, and building evals. That work matters more to whether an agent succeeds than which model sits underneath it.