Our group runs a regional networking community focused on end-user AI and application modernization. A few times a year we put together an expo: a half-day event where vendors, practitioners, and technology leaders come together to talk about what is actually working and what is not. This year I spent most of the day behind the booth rather than walking the floor, and it was a different experience than I expected.
Who comes to these things
I was prepared for a lot of vendor enthusiasm and a lot of skeptical end-users. That was roughly accurate. But the proportions were different than I expected. A meaningful number of the people who stopped to talk were practitioners: actual engineers and architects from financial services, healthcare, and professional services firms: who had been given mandates to "do something with AI" and were trying to figure out what that meant.
These were not people who needed to be sold on the idea that AI was worth exploring. They were past that. What they were looking for was specifics. What does a production AI deployment actually look like in a regulated environment? What are the failure modes that nobody writes about? How do you get a model into your infrastructure without creating a compliance nightmare?
Those are good questions. They are also the questions that the expo floor is mostly not designed to answer.
What people were actually asking
The conversations that came back repeatedly, across different companies and different job titles:
"How do we actually evaluate this?" The honest answer is that most organisations do not have a rigorous evaluation framework for AI tools. They are buying on demos and vibes and then discovering the gap between demo performance and production performance after the contract is signed. The people who were furthest along had built internal evaluation sets: representative samples of the actual tasks they needed the model to perform: and were running every candidate against them before making purchasing decisions. Not many had done this.
"What about the data?" Everybody is worried about data. Where does the inference happen? What does the vendor do with your prompts and completions? Can you get a data processing agreement that your legal team will actually sign? In financial services this is not a hypothetical concern. The number of firms that had stalled AI initiatives specifically because they could not get comfortable answers to these questions was notable.
"How do you modernize without breaking everything?" Application modernization is the other half of the expo's focus and this is where I saw the most nuanced conversations. Nobody serious is talking about rewriting everything from scratch. The conversations that were going somewhere were about incremental modernization: strangler fig patterns, gradual containerisation, pulling pieces out of monoliths without touching the core until there is enough test coverage and operational experience to do so safely. The companies that were actually making progress were the ones that had made peace with the idea that modernization is a five-year programme, not a project.
The gap between what gets announced and what gets deployed
One pattern that stood out across the day: there is a significant gap between what organisations announce they are doing with AI and what is actually running in production.
This is not surprising. Technology announcements are made by communications teams. Production deployments are made by engineering teams. The incentives are different. But the gap was larger than I expected, and the people who were being honest about it, usually the practitioners rather than the leaders, were often visibly frustrated.
The most common state I encountered was: AI initiative announced, proof of concept running, production deployment stalled somewhere between "we need to solve the data governance question" and "we need to figure out how to monitor this in a way our ops team can actually respond to." These are real problems and they are solvable, but they require the kind of sustained engineering attention that is hard to maintain when the next announcement is already being drafted.
What the good conversations looked like
The conversations I found most useful, both for the people I was talking with and for me, shared a few characteristics.
They started with the problem rather than the technology. The people who came in saying "we are trying to solve X and we are evaluating whether AI is the right tool" were having much better conversations than the people who came in saying "we need to implement AI and we are figuring out where." The first group had a way to evaluate anything I said. The second group did not.
They were honest about where they were. The organisations that were willing to say "we are earlier than we sound in our press releases" got more useful input. Pretending to be further along than you are is a natural instinct in competitive environments but it makes it very hard to get help with the actual problems you have.
They had thought about the operational question. Not just how to deploy the model but how to monitor it, how to know when it is behaving badly, how to roll back, who owns it on the operations side. The organisations that had a coherent answer to these questions were almost always the ones that had production deployments rather than stuck proofs of concept.
What I took away
Running the booth is a particular kind of information gathering. You see a cross-section of where an industry actually is rather than where it says it is. The distance between the two is usually instructive.
My read from this expo: enterprise AI adoption in regulated industries is real but slower and more complicated than the vendor narrative suggests. The technical problems are largely solvable. The governance and operational problems are harder and less glamorous and getting less attention. The organisations that will be ahead in two years are the ones that are doing the unglamorous work now.
Application modernization is having a quieter moment than AI but in some ways is more concrete. The teams that are making real progress are the ones that have accepted the pace that is actually achievable and are executing steadily against it rather than trying to move at a pace the organisation cannot sustain.
The conversations were worth having. If you are working through any of these problems and want to compare notes, the contact form is there.
