AI is powering Wall Street’s earnings boom. Analysts expect S&P 500 profits to jump about 31% this quarter, with two thirds of that growth coming from tech and AI giants. But some on Wall Street warn that growth may be peaking — and investors want proof that all that AI spending is paying off. Joining us now is Conor Twomey, Co-Founder and CEO of AI One, a New York company that helps big businesses put AI to work. Conor, thanks for being with us again this morning.
Johnny, great to be here. Thank you so much.
AI is expected to drive most of the profit growth this quarter. Is that growth built on real demand or is it still on hype?
Demand is still insatiable at the moment. Our projection is that about 5% of organisations have successfully absorbed this technology — there is a long way to go. What we often see is the evolution of AI starting with chatbots, moving to AI that can complete code, then write code, then the rise of the agent. An agent, simply put, is daisy-chaining a bunch of prompts together to get from the beginning to the end of a specific task. We are at peak desktop agent — working on small problems where it is quite easy to input the entire problem statement into large language models. But that is a precursor to the next battleground: enterprise-wide agent AI. Using this new probabilistic technology to do meaningful work. Although companies are showing a huge willingness to make the leap from desktop to enterprise level agents, McKinsey backs up that only about 6% of companies are doing this in the right way today. That is the final step before what we call the emerging agent economy — and it is going to be as radically transformative as the internet itself.
Most of the money for AI has gone into chips, data centres, and power. When do everyday companies like banks and insurers really start seeing real profits?
One of the key things stopping people is that they are sprinting too quickly from desktop agents to enterprise agents — and it is a non-linear challenge. There is a giant gap between what it takes to build an agent that sits on my personal laptop and reviews emails versus AI that can do meaningful work for an organisation. A very interesting analysis from Gartner last week shows that people who take the time to actually figure out the meaning behind their data assets before just throwing agents at the problem can improve the accuracy of their end-to-end operations by over 80% while simultaneously reducing cost by 60%. We spoke on the show last time about the evolution from steam to electricity. People are sticking a Ferrari engine on the side of a horse and expecting it to drive 200 miles an hour. There is some pre-work and rethinking needed. The chips continue to be where value accrues — chip providers win during both the creation of new models and the usage of those models. What we are seeing is large companies wanting to control AI on their own physical architecture. They will treat models like a commodity but as soon as mission-critical operations are inserted into agent experiences, it is something they need to control.
Your company focuses on why many corporate AI projects stall. You say most companies are teaching every AI agent their business from scratch. What does that mean and how much does it cost?
When dealing with a small data problem, the amount of tokens — the currency for large language models — needed to define and work through a problem is very nominal. But as you explode the amount of data to enterprise scale — petabyte scale — it becomes a non-trivial, non-linear problem. A killer pattern in financial services is customer onboarding — KYC and AML. AI may take 200 steps on the end-to-end journey. Is the information complete? Have we seen this customer before? Are they entitled to work in that jurisdiction? Has anything changed in the regulations? Each decision point could be as many as 200 large language model interactions. If you have 100 steps and 200 models, that is 20,000 LLM interactions just for one business operation — and the cost against a frontier model might be $4. When companies invest the meaning behind and between their data, each new agent learns from the last and pushes what it has learned back into the enterprise instead of selfishly holding on to that information. That is what we call a thick versus thin agent. The benefit is material both in accuracy and cost. Most of our customers heading into 2027 — the number one conversation is how do we break the hockey stick growth of token consumption, which is directly correlated to dollars and cents.
Conor, thank you so much for joining us again this morning.
Thank you Johnny. Great to see you.
