The Imagination Gap: Why FOMO Gets You a Pilot and Nothing More with Paul O’Sullivan, Salesforce

John Emmerson: I’m joined today by Paul O’Sullivan, Senior Vice President of Solution Engineering and CTO of UK and Ireland at Salesforce. You’ve written previously that something might be missing from the AI revolution and that something is imagination. What do you mean by that?

Paul O’Sullivan: When I think about imagination, I try and think about it in the context of your existing business and how your business operates. There is a real opportunity in front of every business leader out there. The challenge is: do we operate our businesses the same way we’ve always operated them? Can we reimagine? Can we rethink? Can we apply a level of creativity to understand where we can apply AI to reach completely new customer segments, to drive significant hyper-personalisation that builds incredible loyalty? The constraint within an organisation is just the habit of going: this is the way we’ve always done it, therefore this is how we’re going to do it. When that happens, we just apply AI to point-to-point solutions and miss this opportunity to reimagine what our businesses could look like in the future. That’s really the essence of what I was trying to get to.

John Emmerson: You’ve also said that FOMO is driving more AI investment than actual business purpose. How much do you see that specifically in the mid-market?

Paul O’Sullivan: FOMO will get you to a pilot, but having real intention around your transformation and where you’re going to apply agentic AI is what’s going to drive real business results. What I saw very early on was the hype cycle. Everyone had a jaw-dropping moment: this technology is great, imagine what we could do with it. And then we limited ourselves with our imagination. Lots of organisations and senior executives immediately went: well, they’re doing it, I’ve got to do it. Hence the FOMO. I think organisations started to dip their toe in, but because there wasn’t a meaningful intention around: what is my business need, where are the problems I’ve got, how can I actually deliver greater results — that’s where customers are now winning. When you’ve applied strategic thinking to what you’re trying to unlock within your business, it will see you beyond the pilot phase.

Paul O’Sullivan: I was quite worried at one phase over the last eighteen months, because I think the FOMO action meant that a lot of organisations did it for the sake of doing something — to report to their investors or board members to say: yes, we’re doing AI. But it lacked the link between doing something chatbot or co-pilot-like and actually driving business results. That link was missing. We’re now bridging that gap and showing the opportunity for organisations to actually start with agentic AI, having these interactions with technology and agents that can then go and fulfil tasks on their behalf. Some customers are really pushing the boundary on this and being very creative, but I still think we’ve got a lot of work to do to get everybody to that position.

John Emmerson: What do you think the rate of change is? How fast is this moving?

Paul O’Sullivan: We’ve got to remind ourselves we’re eighteen months in — if even that. We launched Agentforce in October or November 2024 and we were the first agentic solution. So when we put it in the context of time, the adoption rate is higher than I probably would have seen anywhere else. When ChatGPT launched in November 2022, within six months it was the fastest adopted technology we’d ever seen. So the adoption curve is there. But I still think we’re very much proof of concept and pilot orientated. The percentage of organisations that are saying: now let’s make it really tangible and drive some business results — that’s still not where we need it to be. Whether that’s improving productivity and efficiency, which is typically the first thing people go after, or using it for growth — marketing campaigns, outreach, lead nurturing. The adoption is still fast for where we are in the broader context of technology. But we’ve got to get really intentional about where we’re going to unlock business value, because that’s where it becomes really sticky.

John Emmerson: Are there other blockers you’re seeing firms face when it comes to adopting AI?

Paul O’Sullivan: Yes. The first is data foundations. Within a mid-market organisation that’s rapidly growing, you’ve typically got a technology ecosystem where you’ve got components over here, different components over there, and a lot of that has been managed to serve the business needs but the systems don’t talk to each other — there are lots of data silos. That prohibits you from having a single view of your customer, and it also prevents you from giving the context to an AI around what your business is trying to achieve. Getting your data foundations in good shape is really crucial to anyone’s success.

Paul O’Sullivan: The next piece is the skills of the future. We are still people-centric in everything we do. We’ve got to look at this and say: humans are the people who are actually going to interact with the technology and unlock the value from it. But do they understand it? Do they understand what data they can give it, what they shouldn’t give it? Are they using a publicly available LLM where they’re sharing information out in the wild, or do they have a ring-fenced local environment? So you think about security, guardrails, controls, compliance — we need to educate people on this. And we’ve got to overcome the inertia around: will AI take my job? Personally, I believe it won’t. I think it will elevate you as an individual. It will give you a superpower to do things differently, to hopefully get more cognitive capacity, to do the strategic thinking that your organisation really wants you to do but you’re prohibited from because you’re bogged down in day-to-day operations. All organisations have a duty of care and responsibility to educate and support employees in learning how to adopt AI safely and securely.

Paul O’Sullivan: And then the third big piece: the world has been ignited and excited by a large language model. But I truly believe that these LLMs are commodities — amazing pieces of technology, but in isolation and on their own, not connected to a bigger system, they’ll just give you a request and response. You need to be able to bring the data foundation so you can give context to what you’re trying to achieve. You need to safely check for things like hallucinations, toxicity and bias. And you still need to be able to activate AI wherever the work is happening. 89 per cent of where AI is operational in our own organisation is where the work is actually getting done — in the applications employees are using. Standalone chat interfaces on their own are probably not enough. Don’t plan to DIY your AI, because you’re going to lose time to value. Look at your technology partners that can help you move at pace.

John Emmerson: People are already really time-poor. There’s no bandwidth. How do organisations help drive that change, free up that space for people to learn the skills they need?

Paul O’Sullivan: This will boil down to really good leadership and great prioritisation. Pick one or two use cases that will give you incremental margin gain. That margin might be: I need that little bit more capacity. And then when you’ve got a little bit, you can reinvest it into how you unlock the next thing, and then the next thing. Very similar to the data challenge — customers say to me: we’ve been trying to solve the data problem for the last ten years, we’ve had three big transformations around data and still haven’t cracked it. The principle is the same. Your volume of data is huge. But what do you really need that’s going to drive value in that one particular use case? Probably a small subset — maybe a couple of objects, one schema rather than much larger stuff. If you can focus on the things that are going to add value to your organisation now, and then over time take the next piece and then the next piece, that’s the approach. And on the time-poor piece: start with the basics around safety and security. Just understand that you’re not going to give a publicly available LLM corporate information. That’s 101. Then build that onto how do you write a really good prompt. I would encourage people to have a twelve-month plan, but go: what could I do every single month in a bite-sized capacity that could level up all of my organisation to be AI-native?

John Emmerson: You’ve previously called it conversation design rather than prompt engineering. Tell me more.

Paul O’Sullivan: Conversational design is also a future skill. Customer experience is really crucial to adoption rates, success, how people are going to interact with your technology. In the world of AI, conversational design is equally crucial. You need AI to first represent your organisation — how does it represent you the same way that a person in your service centre who picks up the phone and talks to a customer would? The second piece is understanding sentiment. Could we have AI understand, based on word pattern matching, if somebody seems distressed, angry, upset, frustrated — and respond in an empathetic suggested manner? What we’ve actually done in our AI centre at Salesforce is resurrect the Alan Turing test. We have two conversational panels — one has a human behind it and one has an AI agent behind it. We ask our customers to come along and try it. They have five minutes, with the same conversation running in parallel with both sides. After five minutes, they have to guess which is the agent and which is the human. The reason for this: if you can’t distinguish between the human and the AI agent, you’re more likely to just feel: this is fine, I can have a conversation with this. And if you have a good experience, you’ll come back. Conversational design for AI is absolutely crucial for successful deployment in the future.

John Emmerson: What have the results of that Turing test been?

Paul O’Sullivan: Really mixed. With about twenty customers, the room was divided — you could see people genuinely disagreed about which was which. It’s such a good test because it has stood the test of time, literally. I would encourage people: when you deploy your AI agents, get not just your employees to test it, but get your customers involved. Do small pilots, but get it out there, because I think the results would really surprise you. Move beyond the proof of concept and get some real user feedback and usability testing as well.

John Emmerson: You cited an MIT study where 95 per cent of generative AI pilots collapse after the demo stage. What do organisations need to get right in the first 90 days?

Paul O’Sullivan: The first 90 days are absolutely crucial. The use case you pick has to be realistic, achievable and linked to your business and how you run it. There has been a lot of tinkering and a lot of disconnected things — agent sprawling, things being popped up across organisations all over the place. In that first 90 days, you really have to be intentional. Pick one use case and take it all the way. Get the data sorted because you only need three or four objects to augment your prompts and get meaningful responses. You only need one workflow to be triggered or one API call to be made. Being able to narrow the scope into one use case, then measure the value from that, and then actually move on to the next — that is a fundamental component. There is risk that organisations try and create monolithic transformations and end up with a proliferation of different tools in different areas. You’ve really got to cut through the noise and look for the signal on what is going to unlock business value. Start small on the single use cases, but have a real look at your overall strategic direction — because otherwise agent sprawling is going to add to technical debt really quickly and organisations are going to see a lot of sunken cost. Can you keep up with the R&D effort of the technology partner you’re working with? The pace of change is significant. You need the freedom to swap in and out an LLM as it improves. Have a conversation with your technology partners. Invest in understanding their strategic direction because they might actually help you move quicker.

John Emmerson: Final question. If you were advising a mid-market CEO, what’s the one piece of advice you would give them right now when it comes to AI adoption?

Paul O’Sullivan: For them to get their hands on the technology. There is nothing quite like a leader who turns around and says: look at what I did — whether it’s over the weekend or on a Wednesday afternoon. Look at how I’m using the technology. Set the tone and example for your organisation because others will follow. They’re looking to the leaders of these organisations to say: this is it, this is how it’s working, this is how I can get benefit from this technology. Even if it’s something as simple as summarising your last meeting to draft a follow-up email. Just use it, show it to your employees, and what you’ll find is that will drive the momentum you need within your organisation for full adoption. The time is now.

Why does FOMO lead to pilot collapse in AI projects?

Paul O’Sullivan’s argument is that fear of missing out creates urgency to start something visible rather than something useful. Organisations stand up a pilot to demonstrate activity — to report to the board that they are doing AI — without first establishing what business problem it is supposed to solve. Without that intentional link between the pilot and a real business result, the pilot stalls after the demo stage. An MIT study he cites puts the collapse rate at 95 per cent. The organisations that push through are the ones that started with a clearly defined business need, not a technology.

What are the three main blockers to AI adoption in mid-market businesses?

Paul O’Sullivan identifies three blockers in sequence. First, data foundations: most mid-market organisations have accumulated technology ecosystems over years where systems do not talk to each other, creating data silos that prevent AI from having the context it needs to do useful work. Second, people skills and security literacy: employees need to understand what data they can and cannot feed into AI tools, particularly regarding publicly available versus enterprise-controlled LLMs. Third, LLM connectivity: a standalone large language model without retrieval augmentation, grounding, safety checks and integration into the systems where work actually happens will only give you a request and response — not business value. Most organisations attempt to solve the third problem without fixing the first two.

What is conversation design and why does it matter more than prompt engineering?

Paul O’Sullivan uses the term conversation design to describe the discipline of deploying AI in customer-facing roles in a way that represents the organisation’s brand, recognises customer sentiment, and creates an experience customers are comfortable with. It goes beyond prompt engineering because it requires thinking about tone of voice, escalation logic, empathy in the response, and the overall quality of the interaction — not just whether the AI can answer a query correctly. At Salesforce’s AI centre, the team uses a modern version of the Turing test — running a human and an AI agent in parallel conversations and asking customers to identify which is which — to assess whether deployments meet a standard customers are comfortable with. If customers cannot tell the difference, they feel fine about the interaction and are likely to return.

What should organisations focus on in the first 90 days of an AI project?

Paul O’Sullivan’s advice is to pick one use case and take it all the way through to a measured result before touching anything else. Scope it to the minimum viable data — probably three or four objects, not a whole data estate. Trigger one workflow or one API call. Measure the value properly. Then move to the next use case. The risk he identifies is organisations trying to create monolithic transformations that lead to agent sprawl, technical debt and sunk cost. Being intentional and narrow in the first 90 days is what builds the momentum and internal credibility to expand. He also recommends starting the people skills piece with bite-sized monthly learning rather than a single big training rollout.

Why should mid-market CEOs get their hands on AI tools themselves?

Paul O’Sullivan’s single piece of advice to mid-market CEOs is to use the technology personally and then show their team what they did with it. He argues that people in organisations are looking to their leaders for a signal about whether AI is real and worth engaging with. When that signal comes through demonstrated personal use — a specific thing the leader did, even something as simple as summarising a meeting and drafting a follow-up email — it sends a message that no all-hands meeting or policy document can replicate. The barrier is not usually capability: most tools are accessible enough for a non-technical leader to use within an hour. The barrier is permission. Leaders who demonstrate use give their teams licence to experiment.

How does Salesforce use a Turing test to assess AI deployment readiness?

At Salesforce’s AI centre, two conversational panels run in parallel: one staffed by a human, one by an AI agent. Customers interact with both for five minutes with the same conversation. They then have to identify which is which. The result, consistently, is that the room divides — customers genuinely disagree. The purpose of the test is not to fool anyone. It is to set a deployment standard: if customers cannot distinguish between a human and an AI agent in a live conversation, they are comfortable with the interaction. If they are comfortable, they come back. Paul O’Sullivan uses this as a practical quality bar for conversational AI deployments and encourages mid-market organisations to run their own version — testing with real customers rather than just internally — before scaling any customer-facing AI.

Why is AI not enough on its own without integration into business systems?

Paul O’Sullivan argues that standalone large language models are commodities. They are powerful in isolation but without retrieval augmented generation to ground them in an organisation’s specific data, without safety checks for hallucination and bias, and without integration into the applications where employees actually do their work, they will only ever produce a request-and-response output rather than genuine business value. His data point from within Salesforce: 89 per cent of where AI is operational is where work is actually getting done — in existing applications. Standalone chat interfaces are not where value compounds. He advises organisations not to try to build this integration themselves but to work with technology partners who have already invested in the trust layer and the connectivity infrastructure.