Problem
Many corporations deploy LLM agents, but projects get closed and no profit appears. The speaker runs the robots team in customer service at T-Bank and has deployed agents for two years. He names two causes. The old way of funding AI (one team, one model, one task) pays off only in a few huge processes, and those are already automated. The cost of an agent cannot be estimated in advance: nobody knows which LLM will work or what quality it will reach. Low-code platforms leave the hard parts open: process descriptions, tools, inference and evaluation.
Example from the article: a manager says "I want an agent in my department", and nobody can name the cost with a 20% error.
Solution in brief
Pick processes whose protocol can be written down (intelligence, not judgment). One team builds a reusable agent platform, and a second team of forward deployed engineers rolls it out process by process. Compute the smallest number of processes N at which discounted profit covers discounted deployment cost.
Questions this case answers
- How to choose which operational processes are worth automating with LLM agents when the cost of each agent cannot be estimated in advance
- How to decide between buying a platform, configuring a builder tool or building in-house for a portfolio of many small automation use cases
- How to choose which corporate processes are worth automating with LLM agents
- How to estimate whether an agent program pays back when the cost of one agent is unknown in advance
Tools and methods
- Hybrid deployment model — a platform team plus a deployment team that works inside the customer company
- Intelligence vs judgment split — a test for whether a process can be handed to an agent
- Discounted capex and profit over N processes — a rule for choosing N
- Agent platform — context collection, tool guidelines, LLM serving, evals
- Forward deployed engineers — roll out the platform per process
Google or a one-time email code. No password.
