A client in Vienna recently paused a search for a senior data engineer. The role had been scoped around building retrieval systems for an internal AI assistant. Two weeks in, the hiring manager came back with a rewritten specification. The new requirement: someone who could design systems that constrain what the AI does, someone who could define boundaries around autonomous action.
Companies built their first wave of AI tooling on the assumption that more context would mean better outputs. Feed the model your documentation, your customer records, your product catalogue, and it becomes useful. That logic held for question-answering. It breaks down the moment you ask the system to take action. A manufacturer in Zurich asked us last month for a data architect who could articulate what their procurement assistant should never do, what approvals it needs before committing spend, and how to distinguish a meaningful action from a plausible-sounding one.
The candidates getting offers have two capabilities that rarely sat together before. The first is the technical depth to work with retrieval pipelines, embedding models, and inference infrastructure. The second is the judgement to define boundaries around autonomous behaviour. One data architect we placed earlier this year won her offer by walking through how she would prevent an internal agent from auto-approving vendor invoices based on pattern-matching alone. Her interviewers spent forty minutes on boundary design and fifteen on the technical stack.
Anthropic’s recent pricing adjustments are forcing similar conversations. Frontier model costs mean companies must decide which tasks actually need that level of capability. A financial services company in Frankfurt asked us last week for an engineer who can orchestrate multiple models, routing queries based on complexity and risk, selecting smaller and cheaper options where the task allows. The orchestration role and the boundary-setting role both require the same underlying skill: knowing when a system should stop, hand off, or ask for approval.
Prompted by reporting from Diginomica.