Twenty questions you will hear in real interviews for agentic-era data roles, with complete answers built directly on the 4 C's framework. Read them, rewrite them in your own voice, and you will sound like the most prepared person in the room.
Context is the difference between the words your business uses and what they actually mean. A model can read your data, but it does not know that available inventory means one thing to finance and another to operations. Context is the shared brain that gives AI the meaning it needs to be trustworthy.
A semantic layer maps business concepts to underlying data. It defines what a metric means, how dimensions relate, and which sources are trusted. For AI it matters because every agent that consumes the same semantic layer gives the same answer to the same question — and you stop arguing about whose number is right.
I would pick one business domain with the most disagreement. Define the top 10 metrics and 20 terms with the owning team. Map metrics to trusted source tables. Add lineage. Express business rules in code. Test with real business questions. Expose the layer to one consumer with a real go-live date. Ship in eight weeks, then grow.
Two layers. Access control decides what data the agent can see. Action control decides what it can do. I would inventory every agent action, classify it by risk, write policy as code, route risky actions to humans through a review ladder, and log every decision in a structured audit trail.
I use a five-rung ladder. Rung 1 is full automation with later audit, for low-risk, reversible actions. Rung 5 is human ownership, for actions that change customer outcomes irreversibly. The middle rungs are notify, approve, and review. Different actions deserve different rungs, and the assignment is revisited as evidence accumulates.
Structured, queryable, retained. For each action: the user, the agent, the inputs, the context loaded, the policies checked, the model used, the output, the citations, and the final decision with reason. Compliance should be able to answer regulator questions without engineering help.
Start with model routing — a cheap classifier in front that sends easy work to a small model and only escalates when needed. Add a semantic cache for repeat queries. Shrink prompts by moving to structured-first retrieval. Set a unit cost target per use case and monitor it. These together usually cut cost by 50 to 80 percent.
I keep a model choice matrix mapping task types to model tiers — tiny for simple classification, medium for standard QA, large for complex reasoning, specialized for domain tasks. A cheap router decides at runtime. The expensive model is reserved for cases that genuinely need it.
Per unit of value delivered, not per month. Cost per resolved ticket, per generated report, per answered question. That number gets a target, an alert, and a weekly review. Tokens, tool calls, retries, and model tier are logged per request so the cost can be diagnosed quickly.
Put a thin internal interface above every vendor — generation, retrieval, action, governance, observability. Vendor SDKs sit behind these interfaces as adapters. Use open formats and standards wherever possible. Test a backup provider quarterly with a small slice of traffic. The goal is not multi-vendor at all times; it is the ability to change vendors without rewriting business logic.
Because the AI landscape changes faster than infrastructure cycles. Open formats and standards let you adopt new engines, models, or tools without painful migrations. Lock-in usually hides in the surrounding code — the prompt formats, the evals, the integrations — not in the model itself.
A catalog tells you where data lives, who owns it, and how to access it. Business memory tells you what concepts the business uses, how they relate, what the rules are, and how meaning has changed over time. Catalogs were enough when humans interpreted data. Agents need both.
Treat them as a triangle and pick two hard constraints per use case. Write those targets into the workload's contract. Measure all three in production. Re-negotiate the targets only when business needs change. Optimization should always respect the chosen tradeoff.
A nine-step reference flow every agent follows: user question, context load, permission check, model selection, tool access check, cost check, answer generation, audit trail, and human approval if risk exceeds threshold. Implementing each step as a shared platform service lets new agents inherit the trust path for free.
The execution parts of data work — pipelines, queries, dashboards — are being automated with agents in the loop. The judgment parts — defining meaning, designing trust, managing cost, choosing platforms, talking to the business — are growing. The agentic data professional moves up the stack and gets more interesting work, not less.
Context modeling, governance thinking, cost awareness, platform design, business understanding, and communication. Tools change every year. These skills compound across every tool transition.
Master the concepts in vendor-neutral language — lakehouse, semantic layer, model serving, governance. Maintain a personal mapping sheet of how each major platform implements each concept. Spend one Friday a month learning a platform you do not currently use. Your work stays portable.
Three lenses. Outcome: what does this do for the business. Risk: what could go wrong, and how is it controlled. Cost: what does it cost, and how does it scale. Engineers default to architecture; leaders default to outcome, risk, and cost. Translation is the skill.
Days 1–30 understand: read deeply, score my current work on the four C's. Days 31–60 build: ship one small real thing — a retrieval prototype, a context layer for one metric, a router. Days 61–90 design: turn the lessons into a pattern adopted at work. Then repeat the loop on the next weak C.
Treating every problem as a model problem. When the agent gives a wrong answer, the instinct is to swap the model. Usually the real issue is in one of the four C's — missing context, weak control, runaway cost, or a tool we cannot change. Diagnose by C before changing the model.