We worked on Voice AI for Others for 18 Months Before We Built It for Ourselves
How 18 months of working with voice AI for clients finally pointed inward, and what happened when it did.
Sometimes the hardest problem to see is the one sitting directly in front of you.
Over the past 18 months, we built voice AI systems for two very different businesses. A major pharmaceutical company needed a smarter way to manage their sales pipeline: getting the right information captured at the right moment without slowing their team down. An agricultural business in Northern Ireland needed on-farm data collection tools that worked with how their people actually operated. Out in fields, hands occupied, no time to stop and type.
Different industries. Different problems. Same fundamental challenge: how do you make capturing information feel natural rather than like a chore?
We learned a lot across those 18 months. About voice technology, about what makes a voice interaction feel human rather than robotic, about the gap between a system that technically works and one that people actually want to use.
And then we kept taking phone calls and sending emails to manage our own enquiries. Because somehow, despite everything we'd learned, we hadn't pointed any of it at ourselves.
The problem
Bluprint was getting busier. More enquiries coming in, more initial conversations to have, more potential clients to understand before we could do anything useful for them.
The process was straightforward enough. Someone reaches out, we have a call or exchange some emails, we try to understand what they want to build and why, and eventually we get to a point where we can put together a brief and a quote.
But straightforward isn't the same as efficient. And as the volume grew, the pressure grew with it.
Every initial conversation started from scratch. We knew someone's name and that they wanted to build something, and that was usually it. So the first meeting became an extended discovery session, trying to understand the business, the problem, the users, the constraints, the budget: all the things we needed to know before we could say anything useful back.
Those meetings were long. And because we were going in without context, they were harder to convert. You can't give someone a confident recommendation when you're still figuring out what they actually need halfway through the conversation.
The irony wasn't lost on us. We were spending significant time and energy on exactly the kind of inefficient, friction-heavy process that our clients come to us to fix.
Our approach
Once we knew what we wanted to achieve, we built it the way we build everything: as a prototype first, tested against the real problem before committing to anything permanent.
The concept was straightforward: a conversational AI that could have the initial discovery conversation for us. Not a form. Not a chatbot that asks ten questions in sequence and feels like a questionnaire with a voice. A real conversation. Adaptive, intelligent, capable of following a thread when something interesting comes up and knowing when to move on when it doesn't.
We called it Blu.
The technical build came together quickly once the brief was clear. The harder work, the work that always takes longer than you expect with voice AI, was making it feel human.
This is the problem that doesn't get talked about enough in voice development. A system can be technically correct and still feel deeply wrong to use. The questions can be sensible, the responses can be accurate, the logic can be sound. And the whole thing can still feel like talking to a machine that's pretending not to be one.
Getting past that requires iteration. A lot of it. Listening back to conversations and noticing the moments where something felt off. Adjusting the phrasing. Changing the rhythm. Rethinking how a question gets asked when the previous answer was vague. Testing it with real people and watching where they hesitate or disconnect.
There's no formula for it. It's trial and error (mostly error at first) until gradually the conversation starts to feel like something you'd actually want to have.
The outcome
Blu now handles the initial discovery conversation for every new Bluprint enquiry.
A potential client visits the site, clicks to get started, and has a conversation with Blu that covers everything we used to spend the first meeting uncovering: the idea, the problem, who uses it, what success looks like, budget and timeline. By the time that conversation ends, a structured brief has been generated and reviewed by the client before it reaches us.
We go into every follow-up call already knowing what we're dealing with.
And that's where something unexpected happened.
We assumed the main benefit would be time saved on intake. And it is. Going into a call with full context is significantly faster than starting from scratch. But the bigger change has been in the quality of those conversations.
When you already understand someone's idea before you speak to them, the conversation goes somewhere different. You're not gathering basic information. You're building on it. You can challenge assumptions, suggest approaches they hadn't considered, spot the constraint they mentioned in passing that's actually the most important thing. You can think alongside them rather than just listening to them.
The briefs are more complete. The ideas are more developed by the time they reach the build stage. And the clients feel heard before the work even starts, because they've already had a conversation that took their idea seriously.
Blu isn't just a more efficient intake process. It's a better first impression than a contact form, and often a better first conversation than a cold introductory call.
The cobbler's children finally have shoes.