Ramp's Shevchenko: waiting for a better AI model means finance users wait too

Alexander Shevchenko of Ramp Labs argued that an application company should not sit and wait for the next frontier model. He made the case on the AI Engineer podcast. Ramp builds finance software. His reasoning was that its users, accountants and finance staff, need tools that work now. In our reading, a vendor that waits hands the cost of waiting to those users.

What was said

Shevchenko went back to 2023, when the best model was GPT-4. Building an agent, a program that carries out a task by itself, took heavy effort. Teams needed strict guardrails. Models could not back up, so a wrong turn went unnoticed. Prompts were so fragile that, in his words, "some synonyms could change the behavior of your agent".

Ramp shipped agents anyway. Shevchenko said the company would not hold off for GPT 4.5 or GPT 5, because it had to serve its finance users at once.

Building agents is easier in 2026, he said. He still thinks a team can dig a layer deeper and give users a better experience, though it takes more effort. His example was Ramp Sheets, a spreadsheet agent. It struggled when users dropped in many sheets and asked for a financial model. Ramp worked with Prime Intellect to train a smaller model for finance questions. Frontier models still plan the work and hand questions to it. Shevchenko said: "we actually got above Opus 4.6, which was the frontier back then, exact task accuracy, while also being able to serve it at haiku level latencies." That is Ramp's own claim. The excerpts give no figures.

Why it matters

Our reading: the cost of an AI gap is easy to miss because the vendor does not pay it. The user does. In the Ramp Sheets case, the friction was an agent that struggled on a hard task. Ramp chose to close that gap with a specialist model rather than wait for a better frontier model.

Shevchenko did not claim that future frontier models would fail to fix it. Still, for buyers, this suggests a question. Ask a software vendor what it does about known weak spots today, not what the next model might fix. For builders, the Ramp case shows one answer. Keep the frontier model as the planner and train a small specialist for the slowest, weakest step.

The other side

Shevchenko did not say waiting is always wrong. He said frontier models keep arriving cheaper, faster and better. That lets Ramp offer the same accuracy at lower cost without doing anything itself.

He also conceded a risk. He described KV-cache cartridges, which pre-load compressed context into an agent ahead of time, for example for month-end close. He said the big labs might ship something like this in their push into continuous learning. Ramp would be happy if so, he said, because its staff would already know how to use it.

Two questions stay open. The talk did not measure what users lost before the specialist model shipped. It also did not say whether a company without a research team could afford this path. The cost to users is our inference, not a number Shevchenko gave.