The promise of AI agents is everywhere—but so is the confusion. As Gartner predicts that 40% of enterprise applications will embed AI agents by 2026, vendors are flooding the market with tools that blur the line between genuine autonomous agents, glorified chatbots, and rebranded RPA. For enterprise leaders trying to separate signal from noise, the stakes are real: move too slowly and risk competitive irrelevance, move too fast without proper governance and risk accumulating costly technical debt.
In this in-depth Q&A, Andreea Plesea, Co-Founder and COO of Druid AI, cuts through the marketing fog with clear frameworks for what a true AI agent actually is, why “lift and shift” deployments consistently fail, and what a minimum viable governance structure needs to look like before go-live. Drawing on real-world deployment patterns across healthcare, financial services, and enterprise operations, Plesea offers a pragmatic, workflow-first roadmap for technology and business leaders ready to move from AI experimentation to measurable operational impact.
++
VMblog: Gartner predicts that by 2026, 40% of enterprise applications will embed AI agents. From your vantage point at Druid, what’s driving this accelerated timeline, and are enterprises actually prepared for this shift—or are we seeing another case of analyst optimism outpacing operational reality?
Andreea Plesea: The prediction is directionally right, but the real story is not the technology alone: it is the business pressure behind it. On the one hand, companies are moving faster because they are under growing pressure to reduce service costs, improve response times, protect retention, and do more with the same teams, or the knowledge volume is high and the spread is wide, making it increasingly difficult to manage and assess. Still, there is also the “fear of missing out,” either from competition or from the global trend of AI adoption among the general population.
AI agents are gaining traction first in high-volume workflows where the value is easy to see: service, support, intake, scheduling, and other processes where speed, consistency, and cost-to-serve matter every day.
I would be more cautious in readiness. Many enterprises are ready to add AI that assists. Far fewer are ready to deploy autonomous AI. Adding an “AI feature” is relatively easy. Running it safely in production is much harder. That requires clear rules on what the agent can and cannot do, secure access to the right systems and data, human escalation when needed, and visibility into every decision and action. Also, evaluation plays a critical role, whether it’s a technical assessment of the performance of AI agents and models or a business assessment to see how AI agents can become better skilled at optimizing ROI.
So, this is not just analyst optimism. The demand is real, and adoption will continue. But the companies that will benefit are not the ones that add the most AI the fastest. They will be the ones that start with one high-value workflow, prove measurable business impact, and expand with control.
VMblog: The term “AI agent” has become almost meaningless in vendor marketing. Can you draw a clear line for our readers: what technical and functional criteria distinguish a true autonomous agent from what is essentially glorified RPA or a chatbot with better NLP?
Plesea: It is true that the term “AI Agent” can be misleading and is sometimes misused. It is important to have a high-level classification of the AI-agent types:
- Functional AI Agents are Task-Oriented
Functional or “agentic” AI agents focus on taking action to achieve a specific, pre-defined goal or to complete a workflow. They connect external tools, APIs, RPA, and databases to perform actions, such as updating customer records, booking appointments, or generating software code. Unlike chatbots, they do not just converse- they act. These agents usually automate business processes and IT operations, and manage complex, multi-step workflows.
- Conversational AI Agents
Conversational AI agents are designed to engage users in natural, human-like dialogue, handling queries through chat or voice. They use natural language understanding (NLU) and large language models (LLMs) to interpret intent, manage conversational context, and deliver personalized interactions. While basic chatbots are rule-based, modern conversational agents are LLM-powered, capable of memory retention and context awareness.
- Reasoning AI Agents (Cognitive/Agentic)
Reasoning agents focus on complex, non-linear problem-solving and planning. These agents break down high-level, ambiguous, or complex goals into smaller sub-tasks, develop plans, and iteratively refine their actions based on results. They use “symbolic reasoning” to map environments and understand relationships between objects or steps. They move beyond pattern matching, using “chain-of-thought” or similar mechanisms to simulate logical, human-like reasoning.
The simplest distinction is this: a true AI agent is responsible for completing a task, not just responding to a request. A chatbot answers. Traditional automation follows a fixed script. A true agent can understand what needs to be done, take the right steps across systems, move the work toward a clear business outcome, suggest the next best action, and perform reasoning.
In practical terms, it can maintain context, choose the next action, use connected systems or tools, complete a multi-step process, confirm that the job is complete, and escalate when it reaches a case that requires human judgment. That is the difference that matters to a business leader: not whether the interface sounds more intelligent, but whether the system can reduce effort, remove delays, and deliver a measurable result under clear controls.
The practical test is straightforward: if it only answers, it is not an agent. If it only follows a fixed script, it is not an agent. If it can act within defined boundaries and show what it did, then you are dealing with a true agent.
VMblog: You’ve suggested this transition requires redesigning enterprise workflows rather than simply adding AI assistants to existing processes. Can you walk us through a concrete example of what this redesign looks like in practice, and why the traditional “lift and shift” approach fails with agentic AI?
Plesea: Redesigning does not necessarily mean rip-and-replace. Processes should be adapted to onboard AI agents.
A good example is patient access in healthcare because it clearly shows the difference. In a traditional “lift and shift” approach, a hospital adds a smarter assistant at the front end: it answers questions and collects information, but then the request still moves into the same scheduling backlog, the same fragmented systems, and the same manual handling. The customer experience may sound better, but the operational bottleneck remains. In some cases, it simply moves further down the process.
A true redesign adapts the workflow itself. Instead of just collecting information, the AI can verify the patient, understand the need, check availability, apply the appropriate scheduling or eligibility rules, complete the booking or rescheduling directly in the core systems, and escalate only cases that genuinely require a person. That is where the value comes from: fewer handoffs, fewer delays, lower workload, and faster resolution.
If you only place AI on top of an old process, you improve the conversation but not the outcome. If you redesign the workflow so AI helps move the work through to completion, you start to see real business impact.
Druid AI’s own healthcare model reflects exactly this pattern: AI agents that handle booking, rescheduling, and inquiries across channels while feeding data directly into core systems, rather than creating another disconnected layer.
VMblog: Many organizations are rushing to deploy agents without establishing proper governance frameworks. What are the biggest risks you’re seeing in the field, and what should a minimum viable governance structure look like for enterprises in the early stages of agent deployment?
Plesea: A minimum viable governance structure is not complicated, but it has to be explicit. Enterprises need a named business owner for the use case, plus a small cross-functional control group spanning operations, IT, security, and legal/compliance. They need to define the agent’s scope, action boundaries, data permissions, and escalation rules before go-live. When choosing an agentic technology, an organization must consider full audit and logging of the agent’s actions, human override, automated testing, and continuous monitoring against both risk thresholds and business outcomes. That is the baseline. Those same control principles – governance, secure access, auditability, and visibility – are also the ones Druid AI is built around, because in enterprise deployment, speed without control does not scale value; it scales exposure.
VMblog: Traditional software ROI metrics don’t always translate well to agentic systems. How should enterprises be measuring success with AI agents, and at what point in the deployment cycle should they expect to see meaningful, measurable returns?
Plesea: With AI agents, the wrong way to measure success is by looking at activity alone: how often it was used, how many conversations it handled, how many features were turned on, or how many transactions might be considered successful because they were not routed to a human. The right question is much simpler: how was the business improving in terms of efficiency, growth, and CSAT?
The most common KPIs are the ones executives are already tracking to measure their business performance: faster resolution, lower cost-to-serve, better completion rates, less manual effort, and stronger retention or conversion where service quality affects revenue, newly generated revenue, better customer satisfaction scoring and NPS rates, plus very particular KPIs that are measuring the specificity of a certain process. In some cases, risk metrics also matter just as much: fewer exceptions, better escalation, and more consistent policy handling.
In the first 30 days, the most important signals are operational performance improvements and immediate results: whether the agent is completing tasks correctly, routing properly, reducing time to resolution, and achieving a straight-through processing rate of 20-40%. In the next 60 to 90 days, if the use case is well chosen and tightly scoped, companies should start seeing significant measurable business impact.
The key is to prove value in one workflow first. The companies that struggle are usually the ones that expect broad, enterprise-wide ROI before they have demonstrated a clear result in a single high-value area.
The worst decision a company can make is to overanalyze and delay a decision to implement AI.
VMblog: For an enterprise starting its agentic AI journey today, what should its roadmap look like over the next 6-12 months? What are the critical foundational steps that often get skipped in the rush to deploy?
Plesea: First 30 to 60 days: choose one bounded, high-volume workflow or a process that brings the highest ROI if automated entirely or partially using AI and requires the least effort on the organization’s side in terms of AI adoption. In this way, it will ensure visible economics and will build trust for AI adoption across other processes. In Druid AI we work with our customers to advise, identify and define such scenarios and, depending on the vertical, some of them might be: offering patients information around services, access, procedures, facilities, etc., appointment management, account servicing, claims intake, or order lifecycle, as those are tied directly to revenue, retention, and cost-to-serve, so value is easier to prove.
Next 60 to 90 days: redesign that workflow for agent execution, not just AI assistance: define exactly what the agent is allowed to decide, what systems it can touch, where it must escalate, and how success will be measured.
From 90 days onward: expand the AI agent’s team with new joiners who cover internal processes; expand omnichannel capabilities; add skills to the already implemented agents based on user interactions; monitor and improve; and expand on niche processes that would create a competitive advantage for the organization in the space.
VMblog: You mentioned that certain areas—customer service, operations, internal workflows—are likely to show gains first. Can you rank these by speed-to-value and explain why some use cases deliver faster returns than others? Where should a CFO or COO focus on initial investments?
Plesea: If I rank them by speed-to-value, it is customer-facing service first, and internal workflows second. Customer-facing service delivers faster because the economics are visible immediately: high interaction volume, clear pain around cost-to-serve, resolution time, abandonment, and retention, and a narrower path from automation to measurable business impact.
Internal workflows usually come next: HR, IT, shared services (accounts payable, accounts receivable, collections, procurement), and internal support. They are often more structured and easier to control, making deployment straightforward. But the return is usually seen as efficiency, cycle-time reduction, and released capacity, not as immediate revenue or retention impact. In other words, the ROI is real, but it is usually less visible to the board than frontline service gains.
Therefore, first money should go into one bounded, high-volume workflow where the leakage is obvious, and the metric already matters to the business – for example, account servicing, patient access, claims intake or status, order lifecycle, or scheduling. Those are the best first bets because they combine three things: visible economic pain, manageable scope, and fast proof. The smarter move is to prove one workflow end-to-end, with clear ownership, measurable outcomes, and tight controls, then expand from there.
VMblog: Looking ahead to when Gartner’s prediction materializes, what will differentiate the enterprises that successfully transformed their operations with agents from those that simply accumulated technical debt with poorly integrated AI tools? What patterns are you seeing now that predict future success or failure?
Plesea: The enterprises that will come out ahead are the ones that treat agents as part of an operating model or as part of their staff, as the worst mistake is considering AI Agents as yet another software solution. In this way, even for business owners who are not highly technical, the immediate business outcome could be more tangible. Companies that succeeded are the ones that picked a high-volume workflow with clear economics, redesigned that workflow for decision-and-action execution rather than layering AI on top of old friction, and put one governed platform around integrations, visibility, and control instead of accumulating disconnected tools.
That matters because the market is moving towards systems that decide, act, and shape real user journeys. Those agentic systems have a far lower tolerance for ambiguity, weak controls, or fragmented ownership.
The success pattern is driven by a focus and a targeted approach when choosing candidate processes for automation; identifying an accountable business owner; secure integration into existing systems rather than rip-and-replace; full observability into what the agent decided and did; and business metrics tied to throughput, cost-to-serve, and resolution quality from day one.
Those are also the same design principles that Druid AI emphasizes at the platform level and in the Business AI Agents already available in the marketplace, through orchestration across systems, secure integrations, and end-to-end traceability.
##
