# The Missing Pieces of Workflow Automation Shirsha Chaudhuri, Thomson Reuters Labs | AI Engineer Summit 2025 | 14:37 Source: https://www.youtube.com/watch?v=xq9sz5MksVI Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/the-missing-pieces-of-workflow-automation Published: 2025-04-23 Tags: deployment, enterprise, human-in-the-loop, workflows ## TL;DR - Enterprise AI has moved from shared generative AI tools and RAG toward agent-based workflow automation. - Connectors, reliable access to distributed context, and support for older systems are still missing from many agentic workflows. - Successful workflow redesign needs domain experts, standards, human control, governance, and a way to update deployed agents. ## Summary Shirsha Chaudhuri describes how Thomson Reuters and other enterprises moved from experimenting with generative AI to building RAG systems, evaluating ROI, and testing agents. The next step is to redesign whole business processes rather than automate isolated tasks. She uses customer support and content publishing workflows to show where agents could work alongside people. The hard part is connecting agents to existing systems, including mainframes and fragmented sources of business context. Business stakeholders also need evidence that agents are reliable enough to replace manual work. Chaudhuri says AI practitioners cannot redesign processes alone. They need subject matter experts who understand the work. She also calls for standards around how agents are built, packaged, and deployed. Human control, collaborative user experiences, governance, deterministic steps, and agent life cycles all need to be designed into the workflow. ## Key ideas ### Enterprise AI moved from experimentation toward whole-workflow redesign [02:00](https://www.youtube.com/watch?v=xq9sz5MksVI&t=120s) Chaudhuri describes a progression inside enterprises. Teams first democratized generative AI through tools such as Thomson Reuters' Open Arena, then moved into RAG and prompt engineering for knowledge tasks. In 2024, they began testing tools and frameworks as agents became more common. The next goal is to automate entire workflows and rethink business processes, because automating one task at a time may leave the original process unchanged. She presents this as a shift from adding AI capabilities to an existing design toward creating a different design around what AI can do. ### Customer support workflows contain many possible agent tasks [03:16](https://www.youtube.com/watch?v=xq9sz5MksVI&t=196s) A customer may report a billing problem, an invoice issue, or a product feature that does not work. Support staff search with RAG, check internal tickets, and consult other teams. If they cannot answer the question, IT operations investigates, followed by engineering support when needed. Engineers may use observability tools and scripts, then run builds, regression tests, and integration tests before updating the ticket and meeting the service-level agreement. Chaudhuri says many of these steps could be automated with agents, but asks whether automating each step preserves a workflow that should instead be redesigned. ### Content publishing also needs a different workflow design [05:24](https://www.youtube.com/watch?v=xq9sz5MksVI&t=324s) Her content example starts with an author or content specialist identifying an alert or trigger. Research and approvals follow, then writers create content and editors review it. Important material may pass through several rounds of review before reaching publishing teams, which handle formatting and styling. Agents could perform many of these activities while humans remain in the approval flow. Chaudhuri says this still leaves a design question: organizations should decide how the process ought to work with agents rather than simply insert agents into every existing step. ### Connectors must bridge agents to established enterprise systems [07:06](https://www.youtube.com/watch?v=xq9sz5MksVI&t=426s) Chaudhuri identifies connectors as the first missing piece. Agentic solutions need a layer that links them to current IT systems. This is difficult because enterprise technology includes systems that have been in place for decades. She cites her experience at Thomson Reuters and points to the continuing use of mainframes across large companies. Her point is that an agentic workflow has to connect with the stable systems that already run important work, rather than assume an organization can replace those systems before using AI. ### Reliability and ROI determine whether an agent can replace manual work [09:26](https://www.youtube.com/watch?v=xq9sz5MksVI&t=566s) The second missing piece is a convincing answer to business stakeholders who ask about return on investment and reliability. From a business-impact perspective, the decision can become a binary one: will the organization still need to pay for manual hours, or can those hours be removed after paying for an AI agent? That makes reliability a major obstacle. A promising demonstration is not enough if the agent cannot perform consistently enough for stakeholders to trust the workflow and change staffing or operating decisions. ### Workflow redesign requires domain experts and shared standards [10:21](https://www.youtube.com/watch?v=xq9sz5MksVI&t=621s) AI practitioners can only take process redesign so far on their own. Chaudhuri says they need subject matter experts and specialists from the relevant domain to imagine the new process together. She also calls for standardization. Organizations need agreed ways to build, package, and deploy agents. Without those standards, it is difficult to introduce agents into an established technology ecosystem or explain how they should fit with existing data, systems, and operating practices. ### Agents need a connected view of enterprise context [11:06](https://www.youtube.com/watch?v=xq9sz5MksVI&t=666s) An agent needs access to context that is often spread across business systems, logs, chat messages, IT tickets, and siloed applications. Teams must identify what each system contains and work out how to correlate a single transaction across those sources. Chaudhuri calls this a stumbling block for AI adoption. The problem is not only retrieving a document. It is assembling the right operational history from systems that were built separately and may use different identifiers or structures. ### People, governance, control, and agent life cycles belong in the design [12:01](https://www.youtube.com/watch?v=xq9sz5MksVI&t=721s) Chaudhuri argues for a collaborative user experience in which agents act as assistants and humans support the agents' work. Governance rules around ethics and responsibility must be translated into the agent architecture. Teams also need to decide which steps remain deterministic, which steps an agent can control, and when a person should intervene. Finally, agents need a life cycle that accounts for fast changes in capability and provides a way to bring newer capabilities into systems that are already deployed. ## Notable quotes - "We are looking at a future where we want to reimagine business processes because just automating a task seems redundant." (02:48) - "The first thing that we're missing is connectors." (07:06) - "You need the subject matter experts, you need the Specialists from that specific domain to sort of do this together with them." (10:21) - "Agents are going to be assistants." (12:01) - "We still want to give the human control." (13:15) ## Tools & references mentioned - Thomson Reuters - Thomson Reuters Labs - Open Arena - RAG - AI agents - mainframes ## Who should watch - You are responsible for an enterprise workflow and need to decide whether an agent should automate individual steps or change the process itself. - Your team has a promising agent prototype but is blocked by legacy systems, scattered context, or doubts about reliability and ROI. - You need practical questions for designing human approval, governance, control, and deployment standards around agents. ## Related talks - [How to Build Enterprise-Aware Agents](https://aietalks.com/talks/how-to-build-enterprise-aware-agents) (Chau Tran, Glean, 19:53) - [Building AI Agents that Actually Automate Knowledge Work](https://aietalks.com/talks/building-ai-agents-that-actually-automate-knowledge-work) (Jerry Liu, LlamaIndex, 17:57) - [How agents will unlock the $500B promise of AI](https://aietalks.com/talks/how-agents-will-unlock-the-500b-promise-of-ai) (Donald Hruska, Retool, 16:22) - [3 ingredients for building reliable enterprise agents](https://aietalks.com/talks/3-ingredients-for-building-reliable-enterprise-agents) (Harrison Chase, LangChain/LangGraph, 20:55) - [Agents vs Workflows: Why Not Both?](https://aietalks.com/talks/agents-vs-workflows-why-not-both) (Sam Bhagwat, Mastra.ai, 15:37)