A client of mine last year, enterprise level, handed me a 60-page PDF of “user research.” They had spent a fortune on it. Three months later I asked the product lead about an edge case in their checkout flow, and his answer was “I think the users would want X.” He had not opened the PDF. Nobody had. It was digital landfill. That gap is what virtual personas are for.
The research was not bad. It was unreachable. Mid-sprint, nobody digs through a shared drive looking for a slide deck from six months ago. If the information is not within arm’s reach, it may as well not exist. I tried to fix this once by printing persona posters and taping them to the office walls. People used them as coasters.
Why virtual personas beat a static PDF
In a previous deep dive into functional personas I argued for focusing on tasks instead of demographics. Even functional personas go stale when they sit still. Moving to virtual personas means building a RAG (Retrieval-Augmented Generation) system over your user knowledge. You stop reading about “User A” and start querying a model that has ingested every survey response and support ticket you have collected.
It also makes your UX research impact hard to ignore, because stakeholders can “talk” to the data. A developer can ask how power users would react to losing a legacy filter and get an answer drawn from 500 support tickets. That carries more weight in a meeting than a gut feeling.
Setting up the repository
None of this needs a PhD in AI. You can start with Claude Projects or ChatGPT “Gems.” The trick is the data structure. Feed the model raw, messy inputs (interviews, Discord chats, Notion databases) rather than polished summaries. It is better at finding the pattern in that noise than you are.
// Example: Structuring a persona prompt for bbioon virtual assistants
{
"bbioon_persona_identity": "Tech-Savvy Tina",
"data_sources": ["q3_surveys.csv", "support_interviews_march.txt"],
"instruction": "Evaluate the following feature request based ONLY on the provided research. If the data is missing, suggest a specific user interview question to fill the gap."
}Your virtual personas are only as good as what you feed them. Give the model generic marketing fluff and it hands back generic marketing advice. The raw transcripts are the whole point. I used “executive summaries” once and the model echoed the same corporate speak straight back at me. Never again.
What you actually get out of it
- You ask instead of hunt. No more digging through the drive for the right PDF.
- You can compare perspectives. Query several virtual personas at once and see where they disagree.
- It stays current. Add new research to the repository, as Smashing Magazine suggests, and the personas move with it.
I have spent 14 years watching good research die in an inbox. If you are tired of design decisions going to whoever talks loudest in the room, build something that answers questions instead. I have taken teams from chaotic research folders to an AI mentor they actually use. If you want your next call backed by real user evidence, let’s talk. I have seen the mess and I know how to clean it up.
Research earns its cost when the team can question it in the middle of a sprint, not when it gets filed as one more report nobody opens.
