Druid AI Targets Agent Ontology Gaps

Enterprise software vendors are building agent platforms that connect processes within their own systems, but their agents still struggle to align what terms mean across different systems. Druid AI co-founder and COO Andreea Pleșea says resolving this ontology problem—agreeing on the meaning and relationships of business concepts—is becoming a key challenge in integrating agents. Druid, founded in 2018, began by building conversational assistants to address the front end of an integration problem that robotic process automation (RPA) tackled on the back end.

Pleșea began working on software agents 15 years ago, when they were a research topic, and wrote her PhD thesis on business ontologies: formal maps of concepts used within and between businesses and their relationships. She described agents trying to buy and sell bricks: they had to match different brick types and sellers’ synonyms with buyers’ requests. Today, she likens that challenge to matching users’ intent with the jobs agents perform. After her PhD, she led development of self-service applications at Romanian business software company TotalSoft. The applications sat on top of its ERP system and helped employees navigate information spread across systems.

Pleșea says prebuilt agents in ERP and CRM products risk recreating a situation she says companies faced in 2011 or earlier: workers could not tell where information came from and lacked a single point of truth. An employee seeking information might have to ask a Salesforce agent, then another CRM’s agent. Druid aims to provide a secure, governed environment to orchestrate agents from its platform and other vendors, and an evaluation framework to check whether agents complete their assigned jobs. Pleșea says desktop agents also make it harder at enterprise scale to control, measure and audit each agent’s activity.

Romanian company UiPath, founded in 2005, built its business on RPA, in which software robots click and type through applications, often moving data between systems without usable APIs. Druid began with conversations; in 2018 it connected its chatbots to UiPath robots so a conversation could start a long-running back-office task and receive a response when it finished or encountered an exception. The companies partnered for several years, but Pleșea says they now compete. She says RPA handled back-office work, while enterprise customers lacked access to that automation. Druid CEO and President Joe Kim, who joined in September 2025 after running Sumo Logic, says agents can integrate systems around the intention behind work, such as a salesperson trying to close a deal, rather than only following a designed process. Kim says agents need behavior and trust built into their design and ongoing monitoring.

Kim says large language models (LLMs) make the ideas in Pleșea’s thesis easier to scale because they interpret meaning more readily than older conversational platforms, which matched phrases to predefined intents and slots. He says expectations should reflect the use case: lower accuracy may be acceptable for summaries read by experts able to challenge them, but business decisions and actions involving customers, employees and personal information need to approach 100% accuracy. Kim says LLMs can provide reasoning, while writes to business systems should use existing structured, deterministic systems so organizations can log and control what is committed.

Kim says a common design uses deterministic routing: a router sends a request to a workflow or to a less deterministic path when safety is considered less concerning. He identifies two problems: separating structured and unstructured routes can prevent a system from combining their strengths, and agents on a routed path can have containment issues if they are not controlled. The article’s author recalls seeing RPA robots copy and paste between terminal windows around 2016, and says people still move information between cloud apps with RPA about ten years later. It says UiPath, Automation Anywhere and Blue Prism are pushing into agent communication, while ERP and CRM vendors, integration companies and model providers are also pursuing the agentic layer. The author describes the market as a work in progress and says multiple approaches may have room to compete.

Enterprise software vendors are building agent platforms that connect processes within their own systems, but their agents still struggle to align what terms mean across different systems. Druid AI co-founder and COO Andreea Pleșea says resolving this ontology problem—agreeing on the meaning and relationships of business concepts—is becoming a key challenge in integrating agents. Druid, founded in 2018, began by building conversational assistants to address the front end of an integration problem that robotic process automation (RPA) tackled on the back end.

Pleșea began working on software agents 15 years ago, when they were a research topic, and wrote her PhD thesis on business ontologies: formal maps of concepts used within and between businesses and their relationships. She described agents trying to buy and sell bricks: they had to match different brick types and sellers’ synonyms with buyers’ requests. Today, she likens that challenge to matching users’ intent with the jobs agents perform. After her PhD, she led development of self-service applications at Romanian business software company TotalSoft. The applications sat on top of its ERP system and helped employees navigate information spread across systems.

Pleșea says prebuilt agents in ERP and CRM products risk recreating a situation she says companies faced in 2011 or earlier: workers could not tell where information came from and lacked a single point of truth. An employee seeking information might have to ask a Salesforce agent, then another CRM’s agent. Druid aims to provide a secure, governed environment to orchestrate agents from its platform and other vendors, and an evaluation framework to check whether agents complete their assigned jobs. Pleșea says desktop agents also make it harder at enterprise scale to control, measure and audit each agent’s activity.

Romanian company UiPath, founded in 2005, built its business on RPA, in which software robots click and type through applications, often moving data between systems without usable APIs. Druid began with conversations; in 2018 it connected its chatbots to UiPath robots so a conversation could start a long-running back-office task and receive a response when it finished or encountered an exception. The companies partnered for several years, but Pleșea says they now compete. She says RPA handled back-office work, while enterprise customers lacked access to that automation. Druid CEO and President Joe Kim, who joined in September 2025 after running Sumo Logic, says agents can integrate systems around the intention behind work, such as a salesperson trying to close a deal, rather than only following a designed process. Kim says agents need behavior and trust built into their design and ongoing monitoring.

Kim says large language models (LLMs) make the ideas in Pleșea’s thesis easier to scale because they interpret meaning more readily than older conversational platforms, which matched phrases to predefined intents and slots. He says expectations should reflect the use case: lower accuracy may be acceptable for summaries read by experts able to challenge them, but business decisions and actions involving customers, employees and personal information need to approach 100% accuracy. Kim says LLMs can provide reasoning, while writes to business systems should use existing structured, deterministic systems so organizations can log and control what is committed.

Kim says a common design uses deterministic routing: a router sends a request to a workflow or to a less deterministic path when safety is considered less concerning. He identifies two problems: separating structured and unstructured routes can prevent a system from combining their strengths, and agents on a routed path can have containment issues if they are not controlled. The article’s author recalls seeing RPA robots copy and paste between terminal windows around 2016, and says people still move information between cloud apps with RPA about ten years later. It says UiPath, Automation Anywhere and Blue Prism are pushing into agent communication, while ERP and CRM vendors, integration companies and model providers are also pursuing the agentic layer. The author describes the market as a work in progress and says multiple approaches may have room to compete.