Conversational support AI, built before ChatGPT.
Co-founder and CTO. Raised $50K to build a conversational AI support platform out of Entrepreneur First's EF5 cohort.
01 · The problem
What Assist AI actually needed.
Customer support in 2015 was a queue problem dressed up as a people problem. Teams paid agents to answer the same handful of questions thousands of times (where's my order, how do I reset my password) while genuinely hard tickets waited behind them. Repetitive volume scaled linearly with the business, so support cost grew in lockstep with growth. Assist AI bet that most of that volume was answerable by software, if the software could understand the question rather than match keywords. We placed that bet in late 2015: no GPT, no off-the-shelf LLM APIs, no transformer ecosystem. The hard part wasn't the idea. It was building something that understood intent reliably enough to put in front of real customers with the tooling of the time.
02 · Context and insight
The reframe that set the direction.
Assist AI started inside Entrepreneur First's EF5 cohort in London, a talent-first accelerator where you form a company and a co-founding team in parallel under real time pressure. The thesis we landed on sat exactly where the market went years later: conversational AI applied to a concrete, expensive, repetitive business workflow. Customer service was the right wedge because it has clean economics, a constant supply of training data (every historical ticket is a labeled example), and a buyer who feels the pain monthly. The moat wasn't access to a model. It was the product judgment around when to answer, when to escalate, and how to earn enough trust that a business would let software speak to its customers at all.
03, The approach
The decisions that mattered.
Pick customer service as the wedge, narrow on purpose
Conversational AI in 2015 could go a hundred directions, most of them demos. We chose customer support deliberately because it bounds the problem: a finite intent space per customer, an objective success signal (did the ticket resolve without a human), and a buyer with a budget line already allocated to the pain. Narrowing the domain was the seniority call. It traded a flashy general assistant for something we could actually make work and actually sell.
Build for intent, and make the handoff a first-class feature
The core was understanding what a customer meant, not what they typed: an NLP and intent-classification pipeline built on the tools of the era, with every resolved ticket becoming a training example. A confident wrong answer kills support automation, so we architected the product around graceful escalation. The assistant handled what it understood and handed off cleanly, with context, the moment confidence dropped. That decision made it deployable in front of real customers instead of stuck in a closed beta.
Raise on a working foundation, not a deck
We raised $50K at a $625K post-money valuation coming out of EF5, a pre-seed round whose job was to buy runway to prove the wedge, not scale it. The pitch was the one the industry later proved correct at scale: repetitive support volume is a software problem, and we'd built the technical and design foundations to attack it.
04 · How it's built
Close to the stack, not above it.
As CTO I owned the stack end to end. The center of gravity was a natural-language understanding pipeline that classified customer intent from free-text messages and routed each to an automated answer or a human, built without the transformer and LLM tooling that makes this routine today. I built the platform's foundations: the data model that turned historical tickets into training signal, the surface where customers interacted with it, and the agent-facing interface for escalations. In a two-hat role, what I coded and what I led were the same person on different days.
Assist AI proved a thesis the whole industry would validate years later: conversational AI could absorb the repetitive core of customer service. We built the foundations of a conversational assistant, an early version of what tools from OpenAI and Anthropic do now, and raised against it out of one of Europe's most selective accelerators. The company didn't reach scale. That's the honest outcome: right thesis, early. But building NLU before it was solved, owning both engineering and product as CTO, and raising on a working foundation seeded the AI-for-consumer-products work that came after.
What I’d carry forward
Being early is not the same as being right, and it's expensive. The thesis held (conversational AI did eat repetitive support), but in 2015 the technology forced you to build the hardest part by hand, so more of the budget went to capability that's now a commodity API call. The durable lesson: in AI products the defensible work isn't the model, it's the product judgment around trust. When to answer, when to escalate, how to make confidence visible. Run it again and I'd validate distribution and willingness-to-pay harder before deepening the tech.