The Interview Spine · Chapter 8
Which market are you interviewing for?
AI-native and enterprise are two different jobs wearing one title. Studying for the wrong one is the most expensive mistake on this site.
Every figure elsewhere on this site comes from one corpus: postings pulled from public ATS boards, which skew heavily toward AI-native and product companies. A reader supplied a second, independently collected sample — 48 postings gathered by hand from LinkedIn, weighted toward enterprise and financial services. Running the identical analysis over both is the closest thing this site has to a control, and it produced one confirmation and one correction.
What was confirmed
The applied-versus-frontier split holds. Agents and customer-facing work dominate both samples — more strongly in enterprise, not less. Research markers stay in single digits in both. If you were preparing on the assumption that AI hiring is mostly research hiring, two independent sources now say otherwise.
What was corrected
Retrieval was badly understated. RAG appears in 11% of AI-native postings and 72% of enterprise ones. That is not noise, and it is not a measurement error — it is a real difference in what the two segments build. An AI-native company treats retrieval as one technique inside a product; an enterprise is very often building exactly one thing, which is a grounded assistant over its own documents.
Governance was missed entirely. Responsible AI, model risk and regulatory language appear in 87% of the enterprise sample — the highest single term in it, and one this site had not measured at all. In a regulated employer, governance is not a separate team's problem that arrives after you build. It is a stated engineering requirement in the job description.
The two jobs, side by side
| AI-native product company | Enterprise / BFSI | |
|---|---|---|
| What you are building | A product feature many customers use | One grounded assistant over internal documents |
| Named in postings | Capabilities — “agentic systems”, “evals” | Specific tools — LangGraph, AutoGen, Bedrock, Azure OpenAI, Vertex (41% and 27%) |
| Retrieval | One technique among several (11%) | The centre of the job (72%) |
| Governance | Rarely stated in the JD | An explicit requirement (87%) |
| Observability | Product metrics | AI observability, drift, monitoring (43%) |
| Cost | Cost per resolved task | Token economics and usage forecasting, often to a budget |
| Who you persuade | Product and design | Risk, compliance, audit and a model-risk committee |
| Seniority on offer | IC ladder, Staff and Senior | Skews to Lead, Principal, VP and Head-of |
So what do you actually do differently
- If you are targeting enterprise, build a RAG system and know it to the bone. Chunking strategy, hybrid search, re-ranking, citation fidelity, permission-filtered retrieval, and how you evaluate all of it. In this segment that is the job, not a component of it.
- Learn the named frameworks, not just the concepts. AI-native interviewers ask what you would build; enterprise postings list LangGraph, AutoGen, Bedrock, Azure OpenAI and Vertex by name. Having used one of each is worth more here than it is anywhere else.
- Be able to hold a governance conversation. Model risk, responsible AI, explainability, regulatory reporting. Read Model risk management and Auditing an AI system — in this segment they are not the auditor's chapters, they are yours.
- Expect the title to be senior and the scope to be advisory. Enterprise AI postings skew to Lead, Principal, VP and Head-of, and the work involves persuading a model-risk committee as much as writing code.
- If you are targeting AI-native, do the opposite. Depth on agent design, evaluation and product judgement; do not spend your preparation memorising vendor SDKs.
On the honesty of this chapter
The second sample is 48 postings, hand-collected rather than randomly drawn, and therefore reflects what one person was searching for. It is not a representative sample of enterprise hiring and its percentages should be read as directional.
It is published anyway, including the part where it corrects a figure elsewhere on this site, because a market map that only reports the evidence agreeing with it is not worth reading. If you have a sample of postings from a segment under-represented here, it would genuinely improve this.