AI agent projects are just starting and many are barely proof of concepts. What’s clear is your boardroom is about to ask for the agentic AI plan if they haven’t already.

With that backdrop, we dropped in on Boomi World 2025 to get a feel for what enterprises are doing in agentic AI and emerging best practices. Here’s a look at some early lessons to ponder.

It’s early. If Boomi World 2024 was about plotting a course toward agentic AI platforms and enabling enterprises, Boomi World 2025 was about putting the tools in the hands of what Boomi CEO Steve Lucas called “digital alchemists.” Boomi put its Agentstudio in the hands of its customers to build things with a large well of free built-in messages and consumption tiers. It’s obvious that 2026 will be about agents that graduated to production. No matter what the vendor—ServiceNow, Google Cloud, Microsoft, Amazon Web Services, Salesforce and dozens of others—agentic AI projects are just being hatched.

“I don't know if we're a best practices state at this point,” said Boomi Innovation's Michael Bachman in an analyst session. “I think we're going to be at better practices as we iterate on that, but I don't see how we can do it without something like a control tower.”

Boomi World 2025: Boomi acquires Thru, makes case for AI orchestration, automation platform | Boomi World 2025: Agentstudio, AWS pact, 33,000 AI agents deployed | Agentic AI protocols: MCP and A2A today, many more tomorrow

A screenshot of a computerAI-generated content may be incorrect.

Experiment, but don’t rush in until you’ve sorted use cases and outcomes. Ken Maglio, Principal Architect at World Wide Technology, works in a division of large IT services firm that is spending heavily on centralized AI. As a result, Maglio’s 200-person unit had to wait a bit before investing in its own AI use cases. "We have our own needs but we are not big enough," explained Maglio.

Maglio needed the company's go ahead to invest in AI for its unit and pilots started in the third quarter of 2024. That time frame worked out since Maglio had time to avoid lock-in mistakes by other enterprises. As a result, Maglio looked to build its agentic AI orchestration on Boomi with data and LLMs via Resolve.ai. "The initial questions were how can I use AI to improve experience, drive internal costs down and figure out what agentic looks like a year from now," said Maglio. “We're on our own journey. Basically, they gave us the blessing to go forth and forge our own path and that gave us the time to ask what agentic will look like.”

AI agent projects are more about continuous improvement and options than a destination. Luke Hagstrand, Head of Enterprise AI, Boomi, walked through how Boomi was working through implementing AI agents. The biggest takeaway is that it’s early in agentic deployments and that they are really a continuum from those genAI projects. Boomi began with its ChatB rollout a year ago, focused on sales and marketing for low-risk returns and now is expanding use of AI agents throughout the company.

Lucas said: “I talk about the evolution of the self-driving enterprise. It's not going to happen overnight. This will happen process by process, piece by piece. That's what we want to give organizations. Find the low hanging fruit and deliver trust with that.”

You already have legacy AI technology to worry about. I caught up with two Boomi customers that were focused on data and system integration and already frustrated with early generative AI investments at their companies. These projects worked well enough, but were started more than a year ago. These enterprises, a consumer company and a pharmaceutical vendor, entered deals with OpenAI, which was the only game in town when the projects started. Now they can’t easily go multi-model. Yes, the enterprises have trained OpenAI models on company data and sandboxed them internally. In a nutshell, OpenAI models are the front end interface that taps into the model that has the internal data. “We did it to say we had AI and be on the bandwagon,” said of the customers.

Since you already have legacy AI technology (that’s basically any technology that’s older than 3 months), you’ll have to avoid being locked in. We covered what to look for in an agentic AI vendor last week and horizontal approaches were critical. There are architecture considerations that also matter. For instance, Hagstrand said Boomi took an API-centric approach to generative AI and now agents because it didn’t want to be tethered to any one model or vendor. Boomi uses models from OpenAI, Anthropic and Google Cloud, but can work in more models depending on the use case via Amazon Bedrock. Hagstrand said:

“If we think about this API first, we can try a lot of things and not get locked into large seat based contracts with frontier LLM companies just to try their tools and not even having a baseline for what AI adoption looks like.”

Control the user interface for additional learning and use case knowhow. Boomi’s approach to its internal agents revolved around its ChatB interface. While the models underneath, data and APIs are abstracted away, the data from the interface provides useful information on usage and use cases. “One of the benefits controlling the user interface and having API connectivity is we can track and learn,” said Hagstrand.

There’s a broader point here: Agentic AI means that enterprises can control their own interfaces and more systems are likely to become headless. Boomi CEO Steve Lucas said: “I think SAP will always exist. Workday will always exist; Oracle will always exist. Here's the real question. The real question is, how much of that exists in the future? I believe is their UI will go away entirely.”

Use cases evolve. Those generative AI use cases will come in handy because they create a foundation for agentic efforts. Start with agents that are focused on use cases that can drive measurable returns and then build up to multi-agent systems covering complex workflows. You’re not going to bite off a multi-agent system out of the gate no matter what the vendor tells you. Boomi said 78% of employees now rely on AI agents for daily tasks, productivity has improved 47% and the company has spent 75% less time on customer support requests. Here’s how Boomi’s internal AI agent use cases have evolved.

  1. First focus on sales and marketing use cases.
  2. Then expand to more deterministic use cases in customer success and support.
  3. Create multi-agent systems that handle complex workflows.
  4. And then build specialized agents for specific business functions.

Data quality matters. The companies launching genAI and now agentic AI projects thought that they had their data lake houses in order. Once these enterprises scaled, they realize further data work is required to feed the models what they need to deliver accurate answers. Your data hygiene is even more important with multi-agent systems. Some companies are creating data quality agents.

Where data lives also matters. Maglio said where data lives remains a big issue and enterprises need to figure that out first. "The data has to live somewhere," said Maglio, who noted that his division decided to migrate data to Resolve.ai from multiple data repositories. "The data issue is why we slow rolled this," said Maglio. It’s not uncommon for enterprises to have more than a handful of data lakes.

Remember to dust off those old playbooks. A few folks at Boomi World 2025 made the case that agentic AI rhymes with microservices architecture. That take is on target—especially when you consider standards are just being formed and a lot has to be worked out. Microservices took a while to gain traction since tools needed to be built. In the end, AI agents like various microservices have to share data, communicate and ultimately create a modular system that can operate as one (yet be easier to maintain). It’s possible that early agentic AI rhymes with early microservices architecture.

Don’t be scared of consumption models (yet). Hagstrand said Boomi’s internal use of agents is built around a consumption model. The argument here is that the company doesn’t want to pay for seats when not all of them will leverage agents. Agentic AI is not mature enough to make a bet on usage and eating costs on seats.

Business Research Themes

Data to Decisions, Future of Work

Roles

Chief Information Officer

Tags

Boomi, agentic AI, AI agents