AI Engineer Summit 2025
Missing pieces of workflow automation
Loading slides…
Read the talk
Missing pieces of workflow automation
Agents can automate support and publishing tasks, but redesigning the whole process depends on integration, reliability, shared context, and human control.
From a talk by Shirsha Chaudhuri
From model access to workflow redesign
What is missing before AI can automate a real enterprise workflow? For Shirsha Chaudhuri at Thomson Reuters, this question grows out of bringing AI into teams and business processes—and hearing similar challenges from practitioners across the wider community.
The enterprise journey began in 2023 with making generative AI broadly available. Thomson Reuters built Open Arena, an internal playground where employees could try different large language models and discover uses in their own work. Prompt engineering and retrieval-augmented generation (RAG) followed: retrieving relevant information gave models material with which to answer questions and support knowledge tasks. As those applications spread, the enterprise question became whether they produced a return on investment.
In 2024, tools and frameworks brought agents into focus. The ambition expanded from automating an individual task to automating an entire workflow. That expansion raises a further design question: should a business retain its existing process and automate each step, or reconsider the process now that different capabilities are available? Workflow redesign changes the process itself, not just who performs its tasks.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Follow a customer issue through the organization
A familiar customer-support workflow makes the distinction concrete. A customer reports a billing problem, an invoice issue, or a product feature that does not work as expected. Resolution then moves through a sequence of teams and systems:
- Frontline support searches for an answer. The support representative may use RAG, consult internal tickets, or ask someone in the support hierarchy.
- IT operations investigates unresolved cases. If support cannot answer the question, it creates a ticket for the internal IT operations team, which provides level-two support and launches an investigation.
- Engineering handles deeper escalation. Cases that remain unresolved reach engineering for level-three or level-four support. Observability tools help identify the underlying problem and a possible fix.
- Build and test work follows. Scripts create a new build, and regression and integration tests check the change.
- The case reaches closure. The bug is fixed or the billing question answered. Tickets are updated and the service-level agreement, or SLA, is met. The completed workflow diagram also shows alarms stopped and the customer notified.
Many individual tasks in this chain are candidates for agents. But automating them all would still leave the same escalation structure in place. The larger question is whether those handoffs remain the right design when agents can perform work across the process. Identifying an automatable task is only the beginning of reconsidering the workflow.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Publishing raises the same design question
Content production begins with a different trigger but has a similarly structured progression. An author or content specialist identifies an alert that warrants new content. An approval may authorize research into what should be written. Research and drafting follow, after which editors, associate editors, and reviewers assess the result.
Critical content may pass through several rounds of review before it is finalized. The finished material then reaches the publishing team, which performs its own formatting and styling work before publication. The workflow therefore includes both the creation of the content and the downstream work needed to release it.
Agents could perform many of these tasks while humans remain in the approval flow. Yet that still leaves the question of whether the original sequence should remain intact. New capabilities create an opportunity to redesign the process instead of inserting AI into every existing task. The difficulty is that the capabilities needed to make that redesign work are not all in place.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Connect agents to the systems already doing the work
The first missing piece is connectors: the integration layer between an agentic solution and an enterprise’s current IT systems. A capable agent still needs a way to reach the systems involved in the business process. This is not necessarily an organization’s first encounter with AI. Thomson Reuters has worked with AI and natural language processing for decades, and practitioners elsewhere also support substantial existing technology estates.
Chaudhuri cites two industry estimates to illustrate the continuing importance of mainframes:
| Scope | Figure cited |
|---|---|
| Fortune 500 companies using mainframes | 71% |
| Worldwide production IT workloads running on mainframes | 68% |
Credit-card transactions provide another example of important business activity that continues to depend on this infrastructure.
Editor’s note: Mainframes: The Next 60 Years repeats both estimates without measurement dates or methodology. The 68% claim already appeared in IBM’s July 2017 announcement, so it should not be read as a fresh 2025 measurement.
The integration problem spans the distance between these stable technology stacks and emerging agent capabilities. Reimagining a workflow requires connecting those worlds. An agent’s ability to perform a task does not, by itself, supply the connection to the enterprise system where that task must take effect.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Establish reliability and bring domain specialists into redesign
Once a proposed agent reaches business stakeholders, the next questions concern reliability and return on investment: will it perform dependably enough to justify paying for it? In the labor-replacement case Chaudhuri encounters, stakeholders often frame the decision as binary. Can they stop paying for manual hours when they start paying for an agent, or will the manual work still need to be funded? Reliability determines whether the expected labor savings can be realized. This makes it an adoption obstacle as well as a technical concern.
A third requirement is the domain vision needed to imagine a different process. AI practitioners can understand the technology deeply and still reach the limits of their knowledge of the work itself. Subject-matter experts and specialists in the relevant domain need to redesign the process with them. The vision for a new workflow has to combine knowledge of what AI can do with knowledge of the business process being changed.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Standardize agent delivery and assemble usable context
The fourth missing piece is standardization across how agents are built, packaged, and deployed. Established technology environments need shared expectations for those activities before replacing existing components with agents becomes straightforward. At the stage described here, the ecosystem is still nascent. The requirement is for a dependable way to deliver agents into that environment; no particular standard or deployment implementation is proposed.
The fifth requirement concerns data and context. An agent needs information that is distributed across IT and business systems. Some of it may be in logs, some in chat messages, and some in IT tickets. These systems may also sit in different parts of the organization.
Bringing that context together involves more than collecting records. Teams first have to identify which systems contain the relevant information, then work out how to correlate a single transaction across them. That correlation is a distinct obstacle: separate systems can each hold part of the evidence needed to understand the same event.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Define how humans and agents work together
The sixth missing piece is a collaborative user experience. If agents act as assistants, the workflow must define the human’s role alongside them. People need ways to support agent work, and agents need ways to support people. The interface is part of that relationship: it must make their respective roles workable within the process.
What this collaboration should look like remains an open design question. Chaudhuri asks for fresh ideas about systems that support the relationship in both directions. The requirement is to design the work humans and agents do together, including the human contribution that remains as automation expands.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Carry governance, control, and evolution into deployed agents
The next requirement is AI governance. Security testing has to address the different parts of agentic workflows. More broadly, established commitments to ethics and responsibility need to translate into the different levels of an agent architecture. The unresolved question is how to make those existing governance expectations apply within the system that performs the work.
Control adds another design decision. Some workflow steps should remain deterministic, while others can be controlled by the agent. Humans still need control over the process, including the opportunity to act at the right time. Balancing these forms of control requires deciding where the process follows fixed behavior, where the agent has autonomy, and when a person can intervene.
Finally, the agent lifecycle continues after deployment. As capabilities evolve, how can the latest ones be brought into agents that are already deployed? Delivering the first version does not settle how that system will absorb continuing changes in the technology.
Enterprise adoption is still at an early stage. The work ahead is to connect emerging agent capabilities with the established systems and processes where businesses need them. Chaudhuri closes with an invitation for questions, ideas, and feedback from people who can help bridge that gap.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Thomson Reuters and AWS explain the internal playground’s model access, retrieval pipeline, employee interface and deployment architecture.
Further reading
- Building effective agentsArticle
Practical patterns for predefined workflows, autonomous agents, tool feedback and human intervention.
Updates since the talk
A November 2025 account of Open Arena’s enterprise integration, self-service workflow creation, evaluation and governance capabilities.
Full transcript
Loading transcript…