TL;DR: Today was the Applied AI New York Symposium, a half day of talks about AI at financial services companies. When anyone can build something with AI, the value moves to trust, ownership, and proving the thing actually worked. Here is what I learned and what I’m taking back to my own work.
My Interest in AI Community Events
The Applied AI New York Symposium was hosted by Indicium AI at The Malin in Chelsea this afternoon, with Anthropic delivering the keynote. The pitch was financial services moving beyond pilots to real deployment at scale, while dealing with regulation and risk along the way.
A few weeks ago, I attended an AI event hosted by Ramp and MeritFirst, which was super interesting, particularly the demo of an internal tool for AEs to listen to live calls and answer questions instantly as opposed to having to tell prospects to wait for the follow up. That project was the winner of their company AI Hackathon, and it was cool to learn from what other companies are doing, and recognizing the similarities in both excitement for and roadblocks in scaling AI tools. I received an invite to that event from a recruiter on LinkedIn. For the Applied AI Symposium, I am pretty sure I saw an ad on either Facebook or LinkedIn and I dropped my email, which converted into an invite.
One of the benefits of living in NYC is the proximity to all these fun events where you can talk to people in real life about AI. While I don’t feel particularly lonely working remotely, I still do appreciate opportunities to meet up in person. Thankful that my husband also works from home so I can attend events like this without worrying about childcare!
The common theme across two fireside chats and two speaker sessions was: when building gets cheap, the hard part is no longer the building. It’s trust, ownership, and measurement. The first part is new or something I haven’t heard before, but the second part is important and I will need to dive in further on future blog posts as I learn more and put these ideas into practice. I am super excited to apply these timely lessons to my current post-sales AI Agent program buildout at Juniper Square.
The keynote: the unit of work got bigger
The day started with the Anthropic keynote. The unit of work AI can take on is growing bigger, evolving from basic chat conversations to a whole process. The question we used to ask was “is the model good enough?” But we have come a long way since then. Now, context and architecture is what matters. We should be asking more questions like these:
- Whose credentials does it use?
- Which systems can it write to?
- Who gets paged when it’s wrong?
The keynote’s example of how we should think about implementing AI agents was a credit memo, end to end. An agent pulls the financials, spreads them, reads the loan covenants, checks the credit policy, and drafts the memo. Every step reaches into a different system, and a person approves it and puts their name on it. I really liked this example because it’s quite relevant to the work I am doing in trying to automate specific CSM workflows as opposed to plugging parts of a renewal process or parts of the EBR. My own pitch to my manager also included that nothing gets sent to a customer before a CSM reviews it.
Focus on the Goal, and have them well defined
The first panel was “Activity vs. Impact: How Regulated Financial Firms Tell Real AI Value from Noise,” with speakers from London Stock Exchange Group (LSEG), Anthropic, and Indicium AI. The LSEG speaker described the value of LSEG very similarly to how we see the value prop of Juniper Square in the age of AI. Anyone can vibe code something similar to what a vendor sells, so the actual value of SaaS providers has to be the trusted partner for clients. That means trust, the quality of your evals, and governance. Who has permission to see what, and is there an audit trail of what happened.
They also talked about creating velocity on AI use cases, and the answer was almost boring in a good way. It’s basic project management. Get the stakeholders aligned, write the problem statement, and decide up front how you’ll measure it. Making sure you understand and set the goal from the beginning, because it’s easy (and fun!) to spend time building more and more unless you know where you’re actually trying to go and how what you’re building fits in with the bigger picture.
In a world where you can build anything, going back to being goal oriented matters more. How will you measure success?
Getting from pilot to production
The speakers kept coming back to what it takes to move from a pilot to something real. Risk and governance came first. Pilots and sandboxes are how you test what the tools can do, but you also need a way to evaluate what the agent produces.
Evals came up as more than a safety net. Every time a new frontier model is released, a good eval tells you how much incremental value you actually got from it. Without one, you’re just guessing, or worse, not thinking about it at all.
The other warning was about where people point AI. If you just plug it into one bottleneck, another one pops up right behind it. That’s typically how people start when they experiment with AI, myself included. I’d say that is very much a 2025-June 2026 thing to do. We start with trying to fix specific problems, but with the capabilities of agentic AI, the better move is to rethink the whole workflow. Don’t limit yourself to how things have always been done.
A few other points from the panel stuck with me:
- AI is very good at understanding how legacy systems behave. On modernization projects, a sophisticated problem statement pays off at the end, when you have to turn the old system off and explain why you must pull the plug and have proof that you are ready to.
- For future initiatives, look at your data strategy, where AI can speed up engineering priorities you already have, and your cloud infrastructure.
Are you trying to boost individual productivity, or revamp a whole process? In financial services, human judgment was the bottleneck. If AI can scale judgment, a process you run weekly could run daily, or instantly. That changes which metrics matter, and individual productivity and enterprise productivity each need their own.
The panel also made the case for a grassroots innovation loop instead of a top-down one, so that ideas touch different parts of the org and the process can actually be restructured.
The Rise of the Citizen Developer and what to be wary of
The second fireside chat session, “Rebuilt, Not Retrofitted: How AI Is Transforming the Way Insurers Work,” had speakers from Lockton and Indicium AI.
I was super interested in this session, even though I don’t directly work with insurers. As a systems thinker, I believe everything is connected, and there are a lot of parallels across different industries. I love having the chance to learn and connect the dots on my own. There was also really nothing said that I found limited to the insurance industry.
The speakers talked about the identity crisis for software developers, and highlighted two roles that are actually more valuable now because of AI:
- Platform engineering builds the scaffolding that lets coding agents build things safely and get them to production.
- The product engineer owns the outcome of an application and instruments it so they can answer questions about ROI and show it.
They also had a hopeful view of the “citizen developer.” Ideas with value will come from people with deep domain expertise. Vibe coding lets those people turn the idea into a prototype, share it with product people, and get platform teams to help harden it.
I thought the term “citizen developer” was interesting, but I suppose that is what AI allows all of us to become now. I disagree that I myself need a product engineer to help me think through the requirements of my projects, but I agree that for those without much of a background in building software, that would be a good idea.
Three cautions came with it:
- The lethal trifecta. This is a security model for agents, a term coined by Simon Willison. The risk shows up when one agent has all three of these: access to your private data, exposure to untrusted content such as a web page or an email, and the ability to communicate externally. Put together, someone can hide instructions in that untrusted content and trick the agent into sending your data out. Take away any one of the three and the attack gets much harder.
- Observability matters even more with AI. How do you capture intent, and how do you audit it?
- When you iterate on an agent, measure it. This is the evals idea from the first panel, applied to every change: did it make the agent faster, cheaper, or better?
Another important piece of advice/requirement for enterprise AI conversations: think about how context flows through your org.
The firm was always a system
The closing talk was from Steve Byren, Indicium AI’s chief AI officer. He emphasized that AI usage isn’t value. Counting usage tells you the tool is being used, not that anything improved. A lot of finance is built around limited expert hours, so the real prize is not in completing tasks faster. It’s stopping the rationing. The table below is taken from today’s slides:
| Today, because expert hours are limited | If whole processes can be handed off |
|---|---|
| Review each credit once a year | Review every name when something changes |
| Audit a sample of claims | Check every claim |
| Save the deep prep for top clients | The same prep for every client |
| Pass on work too small to staff | Say yes to it |
The closing talk’s second point was let the business build, and make the safe path the easy one. That connects back to the citizen developer idea from the insurer session.
What makes it safe for agents to act at scale
The speaker pointed out that companies have always been a graph of work to be done: people, processes, tasks, customers, and systems. Now agents join that graph.
He named three things that make it possible for agents to act safely at scale:
- Agents run on events, not prompts. Something happens, like financials being filed or a covenant test coming due, the agent acts, and a person gets the decisions that need judgment. (How does news travel?)
- Every actor needs an identity. Think of the employee directory extended to agents and skills, each with an owner. For example, a human has their name, role, team, manager, and authority. An agent has a name, description, owner, authority, and version. A skill has a name, author, approver, agent users, version. (Who can act?)
- Govern what every actor can see. That means an enterprise context hub with fine-grained access, and every action traced and evaluated. (Who sees what?)
“The road must run safe first and fast afterward.”
That is the quote that matters most to me from the whole talk. Anyone can buy the technology. The advantage goes to firms that can let agents act safely at scale, which means learning to trust AI at scale.
The pace is why this matters now. The Anthropic keynote had already made the point that the length of a task AI can handle is doubling every 131 days. A closing slide extended it: if that keeps up, that is roughly 2,000 times longer tasks by 2030.
From my notes, the speaker laid out the roadmap like this:
| When | What it looks like |
|---|---|
| Today | Agents on bounded tasks |
| Next two years | Processes run on events |
| 2030 | The firm runs on the graph |
The speaker believes that by 2030, capacity will be counted in decisions, not heads. Instead of asking “how many people do we need?”, the question becomes “which decisions need a person?” You spend expert hours there and give everything else full attention. The slide’s example was a credit team of ten that reviews 2,000 names a year. Today, its capacity is ten people. In an AI-native firm, every name is reviewed continuously, and capacity is the exceptions it can judge well each day.
Coincidentally, this is very close to the pitch I made to my manager about my AI agents: “Every client gets our CSM’s full attention, every day.”
Basic Questions for Enterprise AI Architecture
The Anthropic keynote ended with three questions to take back to your own organization. I think they work for almost any team, not only banks:
- Which process would you hand off end to end? Pick one that matters, and draw the chain the way they did with the credit memo.
- Are you wiring each agent by hand, or building the shared pieces once? Will your 50th agent be cheaper than your fifth?
- Who owns the outcome for that process, by name? If no name comes to mind, that is the place to start.
Wrapping up
I went to a financial services event and came back with ideas I can use in customer success. The technology will keep getting more capable. What decides whether it works is everything around it: who owns the outcome, what each agent is allowed to see, and how you know it did the job.
“The road must run safe first and fast afterward” is the line that matters most to me, and it’s a good rule for any team building with AI. As a risk averse person by nature, I am also very aligned with this idea. I’m currently talking to my head of security about my plans for the post-sales org AI agents, so I really appreciate Steve Byren’s three points on how to unlock safety at scale. I am having Claude build out my events, actors, and governance chart for my team now.
If you’re working through the same questions, I’d love to hear what you’re seeing.
Frequently asked questions
It is a half-day event hosted by Indicium AI at The Malin in Chelsea, with Anthropic delivering the keynote. The 2026 edition focused on financial services companies moving beyond AI pilots to real deployment at scale, while dealing with regulation and risk.
It is a security model coined by Simon Willison. The risk shows up when one agent has access to your private data, exposure to untrusted content such as a web page or an email, and the ability to communicate externally. Take away any one of the three and an attack gets much harder.
The speakers pointed to risk and governance first, a way to evaluate what the agent produces, and a clearly defined goal with a decided way to measure it. They also advised rethinking the whole workflow instead of plugging AI into one bottleneck.
The closing speaker named three things: agents run on events, not prompts; every actor, human or agent, needs an identity; and you govern what every actor can see through an enterprise context hub with fine-grained access, where every action is traced and evaluated.
The closing speaker predicts that by 2030 capacity will be counted in decisions instead of employees. The question shifts from how many people you need to which decisions need a person, so expert hours go to the exceptions and everything else gets full attention.