How Druid is approaching agentic AI with a governance and composability mindset

There is a growing consensus that the future of generative AI lies in finding better ways of discovering how agentic AI could help discover new ways of working. The big challenge is that the industry is still in the early days of knowing what these are and how agentic workflows can support them.

Existing vendors are championing their own approaches that leverage their strengths in enterprise platforms, low-code, and robotic process automation (RPA), or domain expertise. Also, the recent hype around OpenClaw is bringing excitement to the thesis that personal knowledge management workflows might scale to enhance many enterprise processes. Druid AI is championing the thesis that this agentic future could build on its track record in conversational AI, compliance-ready integrations into enterprise platforms, and extensibility of RPA platforms.

Against this backdrop steps Joe Kim, CEO of Druid AI, who took the helm last September after a successful career at Sumo Logic, Citrix, SolarWinds, Hewlett Packard Enterprise, and GE. So, what got him excited about Druid AI:

Obviously, AI is all over the place. People say that it will be as big as the Internet. For agentic AI, it is starting to see a tailwind where people can now start implementing capabilities to gain real business value. And actually, this is how I ended up joining Druid AI. I did a market analysis of which products and platforms were ready to deliver to enterprises. 

There was only one company that was ready for that, as Druid AI had all the fundamental capabilities that you needed to be successful in this space. And I think that's where Druid AI was for a long time. Ever since its inception, almost eight years ago, the company has been working on things that would bring value to end customers and the enterprise. And it just so happened that AI capabilities, like generative AI and, more specifically, agentic AI, help deliver those business values to customers, thereby positioning us better to be more innovative than any other player on the market.

Kim also believes it’s important to differentiate the excitement surrounding personal agentic AI exemplified by OpenClaw from enterprise agentic AI constraints and requirements:

Using AI for personal tasks is one thing. You can tolerate lower accuracy because you’re the final decision-maker. But in the enterprise, the bar is completely different. Accuracy needs to be far higher, and the expectations around security, compliance, and trust go up dramatically when you're handling other people’s data. That’s exactly why Druid AI is built the way it is. Our platform includes the controls, certifications, and safeguards that let government, healthcare, higher education, retail, and other regulated industries use AI confidently. You won’t see many competitors with the same level of readiness.

Different agentic approaches

With enterprise AI agents reaching peak hype cycle, it's worth unpacking what that means for enterprise workflows. All enterprise platform and LLM vendors have their own spin on the idea, along with traditional RPA, process intelligence, and API integration vendors.

The big shift with enterprise AI agents in particular is about flexibility. Traditional automation tools focused on predefined workflows. RPA, for example, traditionally required teams to script every step of the process in advance. As a result, the bots could not easily adapt when information changes, the user’s intent shifts, or the workflow diverges from what was originally mapped. They perform exactly what you programmed, nothing more. Kim observes:

AI agents have a completely different approach. They can interpret intent, adapt to variability, and make quick decisions. A typical IVR system at a doctor’s office is very structured: ‘Press 1 for billing, press 2 for appointments.’ Options are strictly limited, and users quickly become frustrated. On the other hand, agentic AI enables full customization. The system verifies the caller's identity, understands their needs in natural language, and completes the transaction without forcing the user through the company’s internal process map. It’s faster, more accurate, and increases the number of completed transactions considerably.

The seminal difference is that AI agents are getting much better at elevating the entire interaction rather than just automating tasks. This can remove friction, adapt in real time, and scale the number of transactions a business can close. Kim says:

This is how people get delighted, and enterprises finally get past the ‘glass ceiling’ created by rigid automation, delivering both better user experiences and stronger operational outcomes.

Shifting the practicality of agents

Druid has a long history of creating tools for conversational agents by thoughtfully combining a slots-and-intents conversational model with extensive platform integrations. There have been many experiments over the last couple of years to assess how LLMs can process unstructured data while minimizing problems such as hallucinations and ungrounded responses.

A popular notion is that this could be done by combining LLMs with MCP support, retrieval-augmented generation (RAG), and a handful of other tools to quickly produce systems capable of better decision-making and more sophisticated actions. Kim argues the fundamental problem is that this approach fails to consider LLMs' weaknesses as generation engines against the need to take action and make decisions on behalf of users that require a far higher level of accuracy and reliability:

To get there, you can’t rely on LLMs alone. You need structured data, enterprise systems, rules, context, and modern tooling all working together. And you need a way to orchestrate both the ‘old’ structured logic and the ‘new’ reasoning capabilities in one place. Our internal orchestration layer, Druid Conductor, is designed to do exactly that. What the industry is realizing now is that true agentic AI, the kind that can act with enterprise-grade precision, is far more complex than many early adopters expected. It’s not just stitching tools together. It’s orchestrating an entire ecosystem to deliver accuracy, governance, and real outcomes.

One important distinction to consider is how well a given agentic approach might support or create frictions for building new workflows that span the enterprise. RPA approaches risk being constrained by a process-centric view of agents, while enterprise platforms might limit extensibility to processes that operate outside their scope. This can create riction when new processes could benefit from crossing boundaries within HR, finance, operations, customer service, and legacy tools, as well as multiple domains.

Kim argues that, when done well, agentic AI can introduce flexibility across the entire business process. This means thinking about how to coordinate every AI and automation asset in an enterprise to ensure they work together coherently and support true end-to-end processes. That’s the level of integration enterprises need as they move from isolated automation to full agentic operations.

Druid’s strength has been building this orchestration tier for eight years, long before the LLM boom. This long history of working with enterprises to solve orchestration challenges across industries and tools has helped shape their approach to challenges such as versioning, CI/CD integration, and observability. Kim predicts:

You’re going to hear a lot of discussion about orchestration because it’s becoming central to how enterprise AI actually works. We’re confident in our approach because we’ve been building orchestration capabilities for eight years, and the market is only now recognizing how critical this maturity really is. Once you start deploying AI agents at scale, you immediately face requirements such as versioning, CI/CD alignment, observability, and lifecycle management. These are the areas where our orchestration layer, Druid Conductor, is significantly more advanced than many players on the market.

Mapping the value of the agentic proposition

In the long run, it's worth taking a step back from the technical details of different agentic approaches to consider how they might go beyond existing platform approaches. Kim believes important areas to consider include:

1. automated decision-making & action-taking within existing systems, and
2. elevating the user or customer experience.

One practical example of automating decision-making might start by considering a process like collections. Every industry has to manage collections, whether it’s tuition in higher education or net-30 invoices in SaaS. That process often runs inside Salesforce or another CRM. When done well, Agentic AI could reduce the number of human touchpoints, turning a five-step workflow handled by multiple people into a single human-approved step with the agent handling the rest.

The notion of elevating the customer or end-user experience refers to finding ways to personalize the experience even when the underlying process is structured. In this context, the AI agent helps personalize each interaction, understand intent, gather the right information, and close the transaction quickly and accurately. The combination of more complete, faster, and easier transactions translates directly into higher revenue or greater service outcomes.

What’s the future of agentic development

Another important consideration is the best approach for developing agents to support new workflows. For example, vendors like GitLab, Google, Replit, Lovable and many others are exploring ways to build on their strength in code generation to support agentic development. But Kim argues these are fundamentally designed to generate code based on patterns that already exist on the internet:

They’re excellent for accelerating traditional software development, but that’s not what we’re doing with agentic AI. We’re not authoring applications in the classic sense. We’re authoring new agents. And the challenge is that there isn’t a large body of ‘agent development’ out there for these tools to learn from. No one has built enough of these agents for a GitLab or Google to say, ‘Here’s how you create an agent that does X or Y.’

In contrast to these approaches, Druid has recently introduced a virtual authoring environment for agents that builds on its eight years of experience working with enterprise customers to understand how agents behave, what makes them effective, and what development patterns actually work in production. That experience led to a different set of processes for creating and refining agents than those used with standard SDLC tools designed for websites or Java applications. Kim explains:

Agentic AI also introduces completely new engineering challenges. For example, when you upgrade from one LLM version to another, the question isn’t just whether the code compiles, but rather whether the agent’s conversational behavior changes. Does it respond better? Worse? Does it deviate from expected patterns? Traditional unit and regression testing frameworks weren’t built for that. These are new problems that emerge only when you’re building agents that act, reason, and interact in dynamic environments.

New agentic trust considerations

Many aspects of security and governance mirror what enterprises have always needed, like control, visibility, and traceability, even if the technologies are more advanced. But agentic AI can also introduce new risks. Kim says:

The real risk emerges when organizations rely solely on an LLM to make decisions and take actions. LLMs are extraordinary technologies, but they learn autonomously, and their decision paths aren’t always explainable. If you hand all logic to the model, you lose visibility into why it chose a particular action, and that lack of transparency has already led to some well-publicized failures. It’s not enough to trust that the output is ‘probably fine.’

Mitigating these new risks requires enhancing the ability to log, inspect, and control every step in the decision chain. This can help account for every action the agent takes, which is essential if you’re handling sensitive healthcare, government, or financial data. Kim says:

Without that level of transparency and governance, I don’t see how enterprises can responsibly deploy agentic AI into production environments.

Kim believes that trust and safety will become increasingly vital, especially as new regulations develop. This will increasingly shape enterprise procurement processes when assessing new agentic platforms and capabilities. It will become increasingly important to ask whether they can inherit core components such as authentication and authorization from existing systems. Kim walks through one example:

When I make a request or ask for an action, can the AI platform use current security protocols? Sometimes, these security measures are managed on separate systems. Can your platform inherit those? Also, if some components aren’t available, can you add trust, security, compliance, and identity features directly within the platform? Druid AI offers these features, but I recommend confirming that whichever platform you select supports them.

Growing traction by growing value

Kim says he often starts discussions about the value of agentic AI with this question:

“o you want to grow revenue, or do you want to reduce costs?

The answer is usually both, which is why the demand for these capabilities is so broad. Druid’s 300-plus enterprise customers span retail, banking, financial services, insurance, leasing, healthcare and government agencies. However, they are seeing the fastest disruption in a few key sectors, and a few have even surprised him. For example, banking, financial services, and insurance are moving aggressively because the opportunities for automation and accuracy gains are so significant. Also, healthcare and higher education are experiencing massive disruption as institutions rethink how they support students and staff. Retail has been a bit of a late mover, but is now accelerating quickly.

Part of Druid’s value proposition across these industries has been its intentionally simple pricing model. They charge a platform fee, and then a per-agent fee, whether a customer deploys an existing agent or builds a new one. Kim says:

Customers appreciate the clarity and predictability, and the value they get depends on how they use the agents: some see impact on the top line, others on the bottom line, and many on both. What I hear consistently in meetings is that our pricing model should stay exactly as it is, because it’s easy to understand and aligned with the outcomes they’re achieving.

So, how does Kim reconcile this growing traction with research suggesting that many AI capabilities are not delivering ROI for end users? He reframes it as a challenge to figure out what the leaders are doing right, taking inspiration from the MIT study, which found that 95% of generative AI projects were not delivering ROI:

At our events, we even themed around the '5% club' because that's where our customers are. That said, I believe we will see a significant rise in the number of customers gaining substantial value from utilizing agentic AI capabilities. The reason for this is that many people have conducted science projects, and a few have failed in numerous ways. They now understand how complex this technology can be, and they will seek help from experts like us and hopefully others who have extensive experience, to leverage platforms that deliver value quickly. I anticipate there will be a lot more value created in the next 12 months.

My take

In many respects, we are still in the early days of figuring out what AI agents are and how they might deliver enterprise value, particularly when LLM capabilities are thrown into the mix. Druid has approached this question with extensive experience in working with pre-LLM agent workflows to solve practical enterprise problems. It will be interesting to see how well Kim can build on this foundation to navigate the agentic turbulence the agentic marketplace will likely face over the coming years.