Conor Twomey
25 Aug 2026Hugging Face $13B | Jensen Taps Perplexity | Blank Street’s $75M | Straight Outta Promptin
Conor Twomey (AI One): Data Layers for the Agentic Era
Returning guest Conor Twomey, co-founder and CEO of AI One, recounts how he and co-founder Fergus Keenan left careers building data systems for Fortune 500s after realizing 95% of decision-engine effort went into understanding meaning behind and between data assets. He explains AI One's software for improving accuracy, repeatability, and token effectiveness of chatbots and agentic systems, and how Claude Code shifted market perception that AI can do real work, creating tailwinds for the company's enterprise conversations. Segment continues past the end of Part 1.
Conor Twomey: Scaling Enterprise Agentic AI & Accountability
Conor explains why enterprise agentic projects stall at scale: surfacing information is easy, but guardrails, evals, judges and sandboxes are the hard part, so his company maps meaning between data, creates guardrails by software, and promotes workflows to deterministic code, often cutting token usage by over 95%. He notes only 60-70% of any business is digitally represented, so system memory must close the tribal-knowledge gap, and teams keep hitting 'false summits' where data turns out to be only the first of eight challenges. On liability, he rejects personified 'digital workers' like the big-four accounting firm's 'Harry Holdings' agent — accountability must stay with people — and argues the technology should be invisible, amplifying humans rather than replacing them, with customer support the worst place to deploy agents.
Conor Twomey: Operational Leverage, Copilot KPI Madness & Farewell
Conor argues nobody actually wants AI — they want operational leverage — and the only measure that matters is the financial profile: growing top line or cutting costs with the same resources rather than counting pilots. Ronan shares the Atlassian CEO's internal token-budget scheme for spotting the most AI-keen employees, and Conor counters with a top-five global bank that made four daily Copilot calls a management KPI and another bank that generated 100x more code in a quarter while rebuilding one existing component 76 times. He describes slipping into existing stacks to decide whether a task qualifies for agents, when to lean on humans, and when to reuse learnings as plain code — 'good old-fashioned code is going to be cool again' — before the hosts thank him and read an Airwallex ad.