Caddi launches an agent that builds and governs back-office agents

Caddi, a Seattle-based software company, launched an AI agent on August 24 that discovers repetitive back-office processes, learns how employees handle them, and builds specialized agents under centralized oversight for wealth managers, law firms, and insurers. The company, which announced a $5 million seed round on March 5, 2025, aims to provide a governed operating layer where adoption depends on auditability, permissions, and predictable execution. Co-founder Alejandro Castellano emphasized the complexity of such processes, noting that 'every tool that underestimated it broke in week three.'

Caddi launches an agent that builds and governs back-office agents Alejandro Castellano is turning workflow demonstrations into audited automation for law, wealth management and insurance operations. By RuntimeWire Staff /author/runtimewire-staff · Published Primary source: PR Newswire https://www.prnewswire.com/news-releases/caddi-launches-an-ai-agent-that-builds-and-governs-a-firms-back-office-agents-302857872.html Why it matters Caddi is pushing agent software past isolated task automation into a governed operating layer, where adoption depends on auditability, permissions and predictable execution. Alejandro Castellano https://www.linkedin.com/in/alejandro-castellano/?ref=runtimewire launched a new layer of Caddi https://www.trycaddi.com/?ref=runtimewire on August 24 that finds repetitive back-office processes, learns how employees handle them and builds specialized agents under centralized oversight. The Seattle software maker is targeting wealth managers, law firms and insurers, where a process that looks like routine data movement often contains dozens of exceptions known only by the employee who has handled it for years. Caddi's launch announcement https://www.prnewswire.com/news-releases/caddi-launches-an-ai-agent-that-builds-and-governs-a-firms-back-office-agents-302857872.html?ref=runtimewire describes a system designed to capture that operational knowledge before automating the work. "It is a process with fifty branches that one person carries in their head, and every tool that underestimated it broke in week three," Castellano said in the announcement. That line captures the bet Castellano was pursuing when Caddi announced its $5 million seed round on March 5, 2025 https://www.trycaddi.com/blog/caddi-raises-5m-to-transform-professional-services?ref=runtimewire . Originally from Peru, he spent about five years managing companies and investments before earning a master's degree in engineering from Cornell and joining the AI2 Incubator as an entrepreneur-in-residence, according to GeekWire https://www.geekwire.com/2025/this-seattle-ai-startup-watches-you-work-then-automates-the-tedious-back-office-tasks/?ref=runtimewire . He co-founded Caddi with Aditya Sastry https://www.linkedin.com/in/adityasastry/?ref=runtimewire , an engineering leader who previously worked at insurance software provider AgentSync. Caddi initially described its technology as "record-to-code": an employee demonstrates a task on screen, and Caddi converts the demonstration into an API-driven workflow. The August 24 launch expands that pitch. Caddi now wants to discover the work worth automating, build agents for it and provide the control layer through which operations and IT leaders supervise those agents. Finding the work before automating it Caddi says the new agent reads activity across software already used by a customer and identifies recurring processes, ranking them by frequency and cost. That process-discovery function addresses a basic obstacle for automation projects: an operations leader may know employees are spending too much time moving information between systems without having a reliable inventory of where those hours go. Once a customer selects a process, the employee who performs it opens Loop Studio and demonstrates the workflow through a screen share. The system asks questions when it encounters an ambiguous step, and Caddi says each answer becomes a rule governing the resulting agent. An executed-contract workflow, for example, has to distinguish a final agreement from a draft, verify that the necessary parties signed it and decide how to handle an existing copy. The visible action may be filing a document. The underlying job includes classification, verification, naming rules, permissions and exception handling. Caddi uses language models for steps that require interpretation, then relies on deterministic code for routine execution. Its technology documentation https://www.trycaddi.com/technology?ref=runtimewire says Caddi can use models from OpenAI, Anthropic and Google, while every API call and action is logged. Caddi also says it connects through OAuth and SSO, runs under users' existing credentials and maintains SOC 2 Type II compliance. The architecture is central to Castellano's sales case. A model asked to improvise an entire multi-step process can introduce uncertainty at each decision. Caddi instead constrains model use to defined tasks such as reading a document or drafting text, while fixed code controls the surrounding workflow. Customers can replay runs and inspect what an agent touched and produced. That approach also moves Caddi into the governance budget. Operations employees get a way to create automation without mapping every trigger and data field themselves. IT leaders get deployment controls, run histories and one place to inspect agents operating across systems. A broader product than the one Caddi funded in 2025 Caddi raised a $5 million seed round https://www.trycaddi.com/blog/caddi-raises-5m-to-transform-professional-services?ref=runtimewire on March 5, 2025, led by Ubiquity Ventures https://www.ubiquity.vc/?ref=runtimewire , with participation from Founders' Co-op https://www.founderscoop.com/?ref=runtimewire and the AI2 Incubator, which has since rebranded as AI House https://aihouse.vc/?ref=runtimewire . At the time, Caddi was onboarding customers through a managed program and said it planned to introduce a self-service product. As of August 24, 2026, Caddi's pricing page https://www.trycaddi.com/pricing?ref=runtimewire lists Business starting at $5,000 per month and Corporate starting at $10,000 per month. The plans include process-discovery hours, multi-instance support, SCIM, centralized integration governance and action-level audit logging. Those controls matter when agents begin touching client records, billing systems and compliance processes. Caddi faces established workflow platforms including Zapier, Make, UiPath and Automation Anywhere, along with a growing crop of AI-native agent builders. Castellano's differentiation rests on the combination of demonstration-based setup, API execution and centralized governance. Each part exists elsewhere. Caddi is packaging them for professional-services operations where reliability and audit records can decide whether an automation reaches production. Production claims focus on headcount avoided Caddi says The Planning Center has used its workflows in production for over a year across seven workflows. Caddi's financial-services product page https://www.trycaddi.com/finance?ref=runtimewire describes the wealth manager using Caddi to move financial plans between systems and generate client meeting invitations from Salesforce. According to Caddi's launch announcement https://www.prnewswire.com/news-releases/caddi-launches-an-ai-agent-that-builds-and-governs-a-firms-back-office-agents-302857872.html?ref=runtimewire , Palace Law absorbed nearly three times its previous daily inbound-mail volume without adding an employee, while an unnamed Am Law 100 firm automates 150 conflict checks each day. The same announcement says another unnamed registered investment adviser is using Caddi as it works toward expanding from 1,500 advisers to 5,000 without increasing the transitions staff. Those results are customer and Caddi claims. They also reveal the economic argument behind the product: Caddi is selling capacity that does not require back-office hiring to rise in step with transaction volume, acquisitions or adviser growth. The hardest part of that proposition remains the knowledge trapped in existing operations. Castellano is betting that employees can teach an agent the exceptions through demonstration, and that Caddi can turn those lessons into workflows reliable enough for regulated customers to run unattended. The new governance layer gives Caddi a way to keep responsibility visible as those workflows multiply.