Jaime Orozco

Problem. A language model can explain a word in context. It sounds just as sure when it is wrong.

My part. I wrote the requirements and user flows. I also decided that fixed rules, covered by tests, settle questions with a single correct answer, such as which words form a German verb, and the model only writes the explanation. For the paid live-translation model, I set limits so that idle sessions close on their own.

Checked by. An automated test gives the app’s server a deliberately wrong answer from the model and passes only if the rule corrects it.

Case study:Which decisions I leave to an AI model

Question. Would London software and IT micro-firms adopt AI agents that keep their data on their own devices?

My part. I wrote the plan on my own. I designed a survey aimed at decision-makers in London micro-businesses and ran it on Prolific, a UK online research panel. I analysed the 102 responses, then followed up with interviews.

Outcome. Model accuracy (34%) and data privacy (26%) were the main barriers, and 75% rated local control of their data as necessary or critical. Most respondents had spent between £0 and £50 on AI tools in the previous month, so the plan led with results people could see and put cost savings second. The sample is an online panel, and the findings are indicative.

Case study:What small firms want before they trust an AI agent

Problem. Coding agents reintroduced mistakes I had already corrected, because nothing carried my corrections from one session to the next.

My part. I decide what each prototype must do and judge every build. When I reject an output, I record it in my own words with a one-sentence rule that can be tested, and the agents read those rules before they work. Some concern interface design. Others concern process, such as never calling a screen finished while one of its buttons does nothing.

Checked by. Where a script can verify a rule, it does, and a check is accepted only after it has been seen to fail on a deliberately planted error.

Case study:Writing down what I reject from AI agents

In coursework on privacy, personalisation and consumer trust, I studied consent requirements and GDPR principles such as data minimisation. Modules included Innovation and Digital Strategies; Analytics for Business; and Creativity and Innovation Management.

Upper Second Class Honours. Modules included Digital Business and Operations Management; Business Analytics; Innovative Consultancy Solutions; and Strategic Perspectives.

Spanish (native); English (C2); German (B1.1).