GUIDE

GenAI at work: what employers actually expect in 2026

Prompting alone is not a job. Hiring managers now look for evidence that you can build an AI system on real data, prove it works, and keep it safe and affordable.

Ashwood Education · 4 min read · Sept 2026

An abstract network of connected points of light in purple and pink

A couple of years ago, a portfolio with a chatbot in it stood out. It no longer does. Almost anyone applying for an AI-related role can call a model through its API and get a sensible answer back. What separates candidates now is everything that happens around that call.

This guide covers what hiring managers tend to probe for in generative AI roles, and how to show it.

The demo is not the hard part

Getting a large language model (LLM) to produce an impressive answer once is easy. Getting it to produce a reliable answer thousands of times a day, on the organisation’s own data, at a cost the business accepts, without leaking anything it should not, is the actual job. Interviews increasingly focus on that gap.

Expect questions along these lines:

  1. Where does the model’s information come from, and how do you keep it accurate?
  2. How do you know the answers are right?
  3. What does each request cost, and how would you reduce that?
  4. What happens when a user tries to make it misbehave?
  5. How would you notice if it got worse after a change?

Retrieval: grounding the model in real data

Most business uses of LLMs involve the organisation’s own documents, policies or records. The common pattern is retrieval-augmented generation (RAG): find the relevant pieces of information first, then give them to the model along with the question.

Employers want to know you understand the parts that go wrong. How documents are split into chunks. How search results are ranked. What happens when the answer is not in the documents at all, and whether the system says so instead of inventing one. Being able to talk through these trade-offs on a project you built yourself is worth more than any certificate.

Evaluation: proving it works

This is the skill most candidates lack. An evaluation is a repeatable test of whether your system gives good answers. At its simplest, it is a set of realistic questions with known good answers, run every time you change the prompt, the model or the retrieval.

Strong candidates can explain:

  • How they built their test set, and why it represents real use
  • What they measured, such as accuracy, faithfulness to the source documents, or declining to answer when it should
  • How they used a model to help grade answers, and where they checked its grading by hand
  • A change that looked better but scored worse, and what they did about it
A laptop on a desk displaying a set of bar and area charts

Cost and speed

Every request to a hosted model costs money and time. Employers like candidates who think about this unprompted: choosing a smaller model where it is good enough, caching repeated work, keeping prompts lean, and setting limits so a bug cannot run up a large bill overnight.

Safety and security

LLM systems bring their own security problems. Prompt injection, where text in a document or a user’s message tries to override the system’s instructions, is the best known. Others include leaking personal data, confident wrong answers on sensitive topics, and AI agents with access to tools taking actions they should not.

You do not need to be a security specialist, but you should be able to explain how your project limits these risks: what the model can see, what it is allowed to do, and what a person checks before anything important happens.

Working with the business

Generative AI projects fail more often for organisational reasons than technical ones. The system solves a problem nobody had, or nobody trusts it enough to use it. Employers value people who start from a real workflow, measure the time saved or errors avoided, and involve the people who will actually use the tool. If your portfolio project includes a short note on who it was for and what changed, make sure it is easy to find.

What to put in your portfolio

One well-documented project beats five demos. A strong example:

  • Solves a specific problem, such as answering questions over a public set of policy documents
  • Includes an evaluation set and its results, honestly reported
  • Explains the main design choices and what you would change
  • Shows the cost per request and how you kept it down
  • Notes the risks and the safeguards you added

Prompting still matters, as a craft rather than a job title

Writing clear instructions for a model is a real skill, and it runs through everything above. It is just rarely a job on its own any more. Treat it as one part of building systems, not the whole of it.

Our Generative AI course starts with machine learning foundations, then covers LLMs and prompt engineering, retrieval and fine-tuning, agents and responsible AI, and finishes with a capstone you build end to end in an industry you choose. For a longer route that combines data science with AI agents, see the Data Science + Agentic AI course.

Want to work at Ashwood?

We hire practitioners to teach and mentor each cohort.

See open roles