The problem with career tools
Most career tools assume the user is ready to choose.
They ask a few questions, assign a type, and return a list.
That can work for someone already at the decision stage. It is weak for someone still exploring.
Career change usually starts messier than that.
People can often describe what they do not want before they can describe what they do want. They know they are tired of a role, a manager, a commute, an industry, a pace, or a lack of growth. They may have skills they undervalue and constraints they have not said out loud.
An AI career discovery product has to work with that ambiguity.
The job is not to guess a perfect career.
The job is to help the person make the next useful step with more confidence.
What we built
At SEEK Learning, I led design and discovery for Career Compass, an AI-powered career discovery tool.
The product combined two layers.
First, structured preference collection.
A quiz captured what people enjoyed about current work, what they wanted to change, what environment suited them, what kind of impact they wanted, how they preferred to learn, and how much risk they were willing to take.
Second, LLM-powered synthesis and vector search.
The structured answers were turned into a prompt. The LLM generated a profile of the person's likely career direction. That profile was matched through vector search against SEEK role-page content, including what roles involve, which skills they develop, and where they can lead.
The result was not title matching.
It was semantic career fit.
Start with the questions
The intelligence was not only in the model.
It was in the questions.
Before building the product, I ran a Wizard of Oz test. Candidates answered draft quiz questions. I manually used those answers to generate recommendations and took the outputs into interviews.
That test changed the product.
The richest recommendations came from questions about values, environment, and what people wanted to change. Skill questions mattered, but they were not enough.
Candidates also trusted the output more when they could see how their answers connected to the recommendation.
That is the first rule of AI career discovery:
Do not start with the LLM.
Start with the information the system needs to learn.
Ground the model in real market content
Career guidance becomes thin when it is disconnected from the market.
A general model can produce plausible advice. It can say a person might suit product management, learning design, operations, or customer success.
That is not enough.
The recommendation needs to connect to real roles, real pathways, real skills, and real constraints.
Career Compass used SEEK role-page content as the searchable knowledge base. That mattered because the recommendations could point to what a role actually involves, not only what a job title sounds like.
For career discovery, grounding is not a technical detail.
It is the difference between advice and product value.
Explain why
Career recommendations are high-trust moments.
If the product says "you should explore learning and development", the user needs to know why.
The explanation should connect the recommendation back to the person's answers:
- what they enjoy
- what they want to avoid
- the kind of environment they prefer
- how much change they can tolerate
- which transferable strengths appear relevant
- which gaps they may need to close
This does two things.
It builds trust when the recommendation is right.
It creates correction signal when the recommendation is wrong.
When people can see the reasoning, they can push back on the part that missed them.
Limit the output
More recommendations do not create more confidence.
In the Wizard of Oz test, five options felt overwhelming. Three felt curated.
That mattered.
Career discovery is already cognitively and emotionally loaded. A long list can make the product feel like another search results page.
AI products often make this mistake because generation is cheap.
The model can produce ten paths, twenty role ideas, and a detailed explanation for each.
That does not mean it should.
Good AI career discovery should reduce the surface area of the decision.
Design for negative preference
People are often clearest about what they want to avoid.
No more retail.
Not mining.
No relocation.
No customer support.
No more weekend work.
Those signals are valuable.
They are also easy to lose if the product treats them as conversational colour rather than structured data.
An LLM can acknowledge a negative preference in the moment. The product needs a schema that remembers it.
If negative preference is not captured explicitly, the system will keep recommending things the user already ruled out.
That is how career products lose trust.
Make career discovery adaptive
A career discovery product should not be a one-time quiz.
Preferences change as people explore.
Someone may start by asking for remote work, then realise the real issue is autonomy. They may say they want a career change, then discover they only need a different company size. They may browse roles in a new industry and reveal a preference the original questionnaire never asked about.
This is where LLMs are useful.
They can help choose the next best question based on what the system already knows, what remains uncertain, and what the person is currently exploring.
The product should ask fewer, better questions over time.
Not more questions upfront.
Measure useful outcomes
Career Compass was measured across satisfaction, engagement, adoption, retention, and task success.
In the first version, 26.2% of people who completed the quiz clicked directly through to a role page. Including indirect pathways, 38.3% continued to role pages. Positive sentiment reached 74%.
Those numbers mattered because they showed that people were not only playing with an AI tool.
They were exploring specific career options after using it.
For AI career discovery, useful measurement should include:
- quiz or question completion
- recommendation click-through
- role-page exploration
- correction behaviour
- return visits
- positive and negative sentiment
- whether people act on the next step
Engagement alone is not enough.
The product should help people progress.
What I would design earlier now
The biggest thing I would move earlier is the feedback loop.
A career discovery product needs to know why a recommendation missed.
Was the role too senior?
Too risky?
Too similar to the work the person is trying to leave?
Too far from their life constraints?
Did the explanation misunderstand their motivation?
Click data cannot answer those questions by itself.
A lightweight feedback pattern would have improved the first version faster.
The second thing I would move earlier is content quality review.
Vector search is only as good as the content being searched. Rich role pages produce better matches than thin ones. If the source material is uneven, the AI will inherit that unevenness.
That is a design problem as much as a retrieval problem.
The product lesson
AI career discovery works when the system does four things well.
It asks the right questions.
It grounds the answer in real market content.
It explains the reasoning in language the person can judge.
It learns from correction.
The LLM is useful because it can synthesise messy intent and make semantic connections across role content.
But the product value comes from the design around the model: information architecture, trust, feedback, measurement, and restraint.
That is the work.
About the author
Richard Simms is Principal Product Designer at SEEK, where he leads design for the Career Discovery Agent and Career Feed. He also builds Sentiuma, a personal AI knowledge infrastructure layer.