Infosys BPM is rapidly transitioning from traditional ‘AI-augmented operations’ with AI in the loop to ‘AI-first operations’ with humans in the loop, embracing all aspects of people, process, and technology that need to evolve for this transition. In particular:  

• People: will need to develop AI-first and transaction elimination mindsets and increasingly become process influencers
• Process: domain and context knowledge will be complemented by an emphasis on responsible AI, leading to the introduction of new metrics such as process resiliency and trust indices. The nature of process design will change to include exception support by design, since AI inherently generates exceptions, and decisions will be AI-led and data-driven
• Technology: will increasingly involve purpose-built co-pilots and autonomous agents.

In line with the increased client requirement to demonstrate the responsible use of AI, Infosys BPM ensures that each AI agent developed for a client undergoes DPO, information security, and RAI approval. The company has developed a Responsible AI (RAI) framework, which is used to ensure that every AI use case is validated, with any risks identified as part of the implementation cycle.

Infosys BPM’s RAI engine is integrated into each of the company’s AI platforms, so that every request is filtered through the RAI engine to eliminate factors such as racial bias from both input and output perspectives.

Getting AI-ready

Infosys BPM is getting its own organization AI-ready by:

• Getting the workforce AI-ready, with initial AI training rolled out to 46K personnel
• Creating new AI roles, redefining some of the old roles through an AI lens
• Redefining the AI-First vision for each Infosys BPM practice
• Developing new domain-focused solutions to complement the AI-first vision of each practice
• Creation of playbooks; for example, for identifying use cases, delivering them, and taking AI to market.

It has developed ~10 in-house AI solutions and platforms. These include the development of agentic AI platforms and solutions such as Agentic APOC, Agentic AR, and data profiling assistants for master data management.

Overall, Infosys BPM has 26 clients where agentic AI has already been implemented, with 30-40 AI agents involved, and another 25 where implementation or discussions are underway. Prominent examples are covered below.

Agentic AI in Accounts Payable for restaurant chain

Infosys BPM has deployed eight AI agents to support accounts payable for a leading MENA-based conglomorate specializing in restaurant operations with ~2,600 restaurants in the Middle East and Northern Africa, handling around 400,000 invoices per annum across five languages, including Kazakh and Russian. The service went into production in February 2025.

The documents received from the restaurant chain are complex and include a mixture of invoices, purchase orders, and goods received notes, all in various formats, typically dependent on the expense type. Each expense type also requires a potentially unique set of values to be captured for the ERP system.

Infosys BPM has developed two sets of AI agents to support processing these documents. Firstly, ambient agents process documents in the background, identifying the expense type and the required values, and categorizing each invoice as green, amber, and red. Secondly, assistant agents guide human agents in processing.

“Green” invoices are ready for posting. “Amber” invoices require a human in the loop to validate the processing that has been done. In this case, all invoices with values exceeding $5K are required to be manually checked. “Red” invoices are typically those where some processing has taken place but, for example, the purchase order number was incorrect or particular information could not be found.

Infosys BPM is finding traction for agentic AI in accounts payable, with five implementations underway since May 2025.

Agentic AI for Collections

Infosys BPM has developed AI agents for AR collections and applied them internally within Infosys BPM’s operations. The AI agents monitor multiple client payment portals, read email messages, and create follow-up emails for agents. Here, Infosys BPM has seven AI agents in production, resulting in faster cash collection and a 40%-50% reduction in subprocess effort. Infosys BPM’s wider agentic AI for Bill to Cash (collections) solution combines AI agents and RPA agents, which are autonomously integrated to achieve their goal.

Agentic AI to Drive Autonomous Sourcing

Infosys BPM has also implemented AI agents for spend data cleansing for a client, classifying and enriching 700K MRO data records. If the part number is available, browser-based agents with multilingual capability find the relevant information, curate it, and enhance the product descriptions.

This downstream data cleansing is facilitated by Infosys BPM’s sourcing team’s ongoing identification and uploading of relevant catalogs into a catalog repository within the sourcing solution.

Infosys BPM has further automated the sourcing process, with every purchase requisition going through the sourcing solution. This includes generating the product code, performing first-level negotiation, and contract negotiation. The results of this process are summarized by an AI agent to enable the selection of the most appropriate supplier. This process has now been further enhanced to include spend aggregation, enabling the achievement of additional savings.

Agentic AI Supporting Data Acquisition for Investment Firm

Infosys BPM has deployed six AI agents in support of data acquisition for an investment firm. Infosys BPM was previously capturing information from a limited number of websites, all in the English language, and now utilizes AI agents to create new customer contacts, update existing contacts, and maintain the client database. These AI agents can handle a fivefold increase in the number of websites and process multilingual information. The speed of information capture has also improved, enhancing the client’s time-to-market by 35%.

In the future, F&A will continue to be a key area for the further development of agentic AI at Infosys BPM. For example, in R2R, Infosys BPM is looking to create AI agents to interface with Blackline in areas such as journal entry uploads.

Infosys BPM also recognizes that, in much the same way AI agents combine multiple task skills, so agent orchestration platforms will increase in importance. Infosys BPM will utilize multiple AI agent orchestration platforms, including in-house platforms such as the EdgeVerve AI Next orchestration platform, leading third-party platforms, and solutions from new start-ups, depending on the client’s situation.

Moving Rapidly Ahead with A New Mindset

Infosys BPM is rapidly identifying use cases and building platforms to support client requirements for AI-first operations, developing new domain-specific AI-first solutions at pace. Much of its initial activity has been around Infosys BPM’s traditional focus areas of finance & accounting, sourcing and procurement, and order management, where the company has already developed and implemented end-to-end AI-first offerings.

Infosys BPM also recognizes that implementing AI-first operations requires a fundamentally different mindset and approach from more traditional AI-in-the-loop operations. Two key developments here are a greater emphasis on building exception support into the initial design and the integration of the company’s Responsible AI (RAI) engine into each use case to identify risks and eliminate bias in both inputs and outputs.

Summary

Infosys BPM is moving from ‘AI-augmented operations’ with AI in the loop to ‘AI-First operations’ with humans in the loop. In doing so, it:

• Is preparing its workforce by creating new AI roles and redefining existing ones
• Has developed ~10 AI platforms, including agentic AI solutions for APOC and AR, with 30-40 AI agents deployed
• Deployed its Responsible AI (RAI) framework in every agentic AI use case for risk identification and to eliminate biases in inputs and outputs
• Will leverage multiple AI orchestration platforms, and further embed exception handling and Responsible AI in all AI-first operations.

UiPath’s FUSION 2025 event marked a pivot for the company to agentic automation. Formerly known as Forward, the event has been rebranded FUSION to underscore UiPath’s vision for integration — the fusion of agents, robots, and humans, all orchestrated in a single unified platform. In effect, the previously launched Maestro steps up as the master orchestrator, handling not just classical RPA, but also AI agents and human interaction as part of agentic automation.  

How UiPath is addressing the agentic AI challenge

UiPath acknowledges that building agentic systems is inherently difficult. It requires domain knowledge, systems integration, automation engineering, and governance from day one. To address this, the company is investing in a library of prebuilt agentic solutions aimed at accelerating time-to-value and reducing complexity for customers.

These prebuilt solutions are not automation templates or components to automate a task; each solution includes:

• AI agents trained on domain-specific reasoning tasks
• Robots to execute structured, repetitive workflows
• Human-in-the-loop mechanisms for oversight, approvals, and judgment
• Apps and dashboards tailored to business roles and operations
• Embedded governance for auditability and policy compliance.

Maestro plays a central role: it governs the handoffs between agents, robots, and humans; tracks progress through cases or workflows; and maintains end-to-end traceability (including the actions of AI agents) to ensure audit‑ready compliance.

As part of the announcement, UiPath launched eight prebuilt solutions:

• Consumer Loans: automates document validation and onboarding for home equity lending, reducing cycle times and manual input
• Commercial Loans: supports loan QA with robotic document reviews and agent-based exception handling
• Financial Crime Compliance: as above, orchestrates fraud detection with agents, robots, and human review
• Healthcare Claims & Denials: automates intake, denial prediction, and appeals generation, improving speed and compliance
• Order Management: aombines agents and robots to manage supply chain exceptions and fulfillment
• Inventory Management: optimizes stock handling with demand forecasting agents and robotic transactions
• Commercial Pricing: Uses AI agents to optimize pricing strategies, with human sign-off through tailored dashboards
• Merchandising & Promotions: supports campaign rollout through agent-driven planning and robotic execution.

Maestro delivering results

UiPath has indicated it aims for 95% agent accuracy as a benchmark and reports mid‑90% accuracies so far across the published solutions. The company stated that the human-in-the-loop nature of these solutions ensures that even inaccurate model choices are given to a human for review before they are acted upon and that, as models improve, so should the accuracy. We do wonder, however, about the complacency of humans reviewing data from models that is this accurate, which may make human operators less observant of errors.  

UiPath will continue to build solutions and additionally intends to support partner solutions through a marketplace. Importantly, UiPath intends to vet the agentic solutions added by partners so that there will not be too many agentic solutions covering the same problem statement.

Another message throughout FUSION was that agentic solutions built using Maestro will improve over time as the underlying agent models improve, and that organizations should be looking to implement these solutions and expand scope and improve accuracy over time; in one such example, Capgemini spoke of an internal query resolution solution, developed using Maestro, with which it targeted the top nine queries it received and automated 90% of the workflow. Following that success, the company is expanding the use across different segments, including HR and finance.

Organizations such as Allegis spoke at the event about how this shift to agentic automation hardened its previous automation efforts, which, due to their brittle nature, only delivered ~40% efficiency; this increased to 70% through the use of Maestro.

Agentic AI platform enhancements

UiPath also unveiled several platform-level enhancements to support agentic automation. One of the most interesting was ScreenPlay.

ScreenPlay is a new capability enabling users to describe UI automation needs in natural language. ScreenPlay then translates descriptions into executable UI interactions; for example, users can request a bot to extract information from a website, then the bot will use the LLM to read the website, identify where the information is, and extract it for a variable.

The users requests are transformed into actions using one of the LLMs offered through the platform such as ChatGPT, Google Gemini, and Anthropic, and UiPath intends to support organizations in a bring-your-own-model capability.

ScreenPlay aims to enable users to make more stable bots as the elements are not coded into the bot and therefore are less brittle should an application or a website change its UI. Even with additional unit costs, and being slower to identify objects than selectors, ScreenPlay could be a nice fit in processes that are brittle due to UI changes, and should result in less development and maintenance time spent by automation developers.

UiPath also launched UiPath Labs, a set of experiments and capabilities previews. The labs launched with three tools for users to begin experimenting with and providing feedback:

• Agent Sandbox: a low-code playground for building and testing agents. Consider it “Autopilot for Studio”
• Project Delegate: a personal agent that learns repetitive tasks and executes them, positioned as “Autopilot for everyone”
• Enterprise Knowledge Graph: a feature that connects and maps an enterprise's data to provide context for its AI agents. By understanding the relationships within an organization's data, the knowledge graph allows UiPath agents to provide more accurate and context-aware responses and actions. 

Similar to ScreenPlay, Agent Sandbox and Project Delegate aim to use LLMs to create automations – for automation developers automating processes and end users automating tasks, respectively.

The Enterprise Knowledge Graph is now available for preview, and the Project Delegate and Agent Sandbox will be available soon.

We hope that Project Delegate and Agent Sandbox can fulfil the earlier promises of Autopilot for Studio and Autopilot for Assistant (later renamed Autopilot for Everyone), which also aimed to support users in creating bots through natural language. These did reduce some development efforts but often left extensive last-mile work for developers, such as creating variables to be used in the task. We were told that Project Delegate and Agent Sandbox take these operations a step further, hopefully once again reducing the energy spent developing new automations.

Summary

All in all, the FUSION event highlighted that UiPath is elevating its game using GenAI: going from automation in a BPM orchestrator that combines agents, humans, and automation, and hopefully from semi-manual automation design.

This levelling up is demonstrated by the solutions and the intention to develop a solution marketplace – a levelling up from the company's marketplace of prebuilt RPA content, which covered a much smaller percentage of end-to-end processes.

Having UiPath at the center of these agentic processes as their orchestrator is a smart fit. UiPath founder Daniel Dines was clear in his keynote: many enterprise GenAI initiatives are failing because they lack integrated automation, and a tendency to deploy GenAI in narrow, standalone use cases results in low ROI. This is similar to what we see at NelsonHall, with our research showing that a proportion of GenAI projects require upfront services to ensure data readiness. Processes that have already adopted RPA automation will already have a level of readiness. Organizations and processes with some RPA deployment are more likely to have this data readiness, and platforms such as UiPath Maestro can support the case management and auditability of processes that GenAI solutions cannot do alone.