consulting
Saïd Business School AI strategy
AI strategy for the careers centre at Oxford's Saïd Business School: discovery, a build-or-buy call across about 30 vendors, and three AI prototypes.
the problem
The careers centre at Saïd Business School, University of Oxford, was exploring where AI could help. Two gaps stood out. Careers information is scattered across the website, the course platform, PDFs and emails. And students often reach their first coaching meeting without a clear picture of what they want.
why it hurts
Students can't find answers, so the team answers the same questions again and again, and general AI tools guess because they don't know the school's content. Coaches spend first meetings on the basics instead of the hard part. Vendors sell fixes, but most polish a CV or push jobs.
how i solved it
As AI strategy advisor to the careers centre, I ran discovery and assessed about 30 vendors for build versus buy. My recommendation: buy commodity tools such as CV checkers and interview practice on short contracts, and build only the thin layer that needs the school's own content, data and coaches. I built that layer as working prototypes, one for each gap.
the prototypes, one by one
Student chatbot
Careers assistant
the gap
Careers information is scattered across the website, the course platform, PDFs and emails, so students struggle to find it and the team answers the same questions again and again.
what it does
- Answers in plain language from the school's public careers pages.
- Shows the page each answer comes from.
- When the pages don't cover a question, it says so and points the student to the team.
- No login, and nothing the student types is stored.
31 of 33
test questions with the right page in its top three
design call
Show where every answer comes from.
General AI tools don't know the school's content, so they guess, and the guesses sound confident. A cited page lets the student check.
Earlier demo, invented data
Careers Desk
the gap
Each advisor covers a few industries and a group of students, and much of what they know sits in their head, their inbox and a spreadsheet. A student who has gone quiet is easy to miss.
what it does
- One list each morning across students, employers and alumni, ranked by urgency, each with its reason.
- One page per student: goals, what was said last time, and what to do next.
- Drafts with no send button: it never contacts anyone.
Built in September, before the two above. Every record in it is invented, and a bar on every page says so.
why it's useful
Each idea can be judged on a working tool rather than a slide. The careers assistant also comes with a first test result, below.
outcomes
- RAG chatbot on the school's knowledge base that cites its sources: right page in its top three for 31 of 33 test questions.
technical specs
- Going live would need university sign-in, the school's data protection approval, and personal data stripped out before any internal material is loaded. The stripping tool is already built.
- Careers assistant: keyword (BM25) search over 488 passages from 60 public careers pages. Code, not the model, links each numbered citation to its page and strips any web address the model writes. Tested on 33 questions it should answer and 10 it should decline.
- Career navigator: one question per turn across 15 areas, and code decides when the conversation may close. Every quote in the profile is checked word for word against what the student said; in the coach brief, the facts are copied by code and the model writes only the prose. The job radar is built to read public sources only (81 job boards, 23 feeds, 8 news queries); the worked example runs on a 60-item sample.
- Careers Desk: Next.js 16, React 19 and Recharts on a seeded, fully invented dataset (120 students, 60 employers, 200 alumni), with no keys, database or network.


