Is One Agent Enough?

While AI Agents are unlocking new automation possibilities, not every automation needs to be an agent. In our experience of building agentic systems, setting out to build a single AI Agent to accomplish all tasks is usually not the right approach. With generative AI technologies, the more freedom and creativity you allow a model, the greater the chance to get unexpected results. In some scenarios, this creativity is desirable. However, for enterprise automation use cases — especially in regulated industries — the key is to allow just enough autonomy to cope with the variation expected in the use case, but not so much that you are no longer in control of the overall process.

A pattern that works well is to create multiple agents, each with a discrete role or function to perform. This approach allows specific contextual instructions to be provided to each agent, ensuring it remains 100% focused on the desired goal or outcome. Additionally, each agent is only given access to the tools it needs — such as data sets, documents, or integration connectors — to complete its task. We call these “worker agents.”

And just like in the real world, worker agents can collaborate by talking to each other using a combination of natural language (similar to how humans talk to each other in English) and structured data formats, such as JSON.

We then use a “managing agent” or controlling workflow to delegate specific tasks to the worker agents. The process orchestration capabilities of the platform — including comprehensive audit trails, SLA management, and exception handling — facilitate, govern and monitor the interactions between the worker agents.

The Gift of Knowledge

One highly publicised limitation of Generative AI is the problem of hallucinations — instances where the AI model guesses or makes up facts to fill gaps in its training or knowledge.

This problem is particularly concerning for businesses using AI models to process business documents, where accuracy and data correctness are of paramount importance. In regulated industries or use cases where an incorrect answer from an AI model could result in a large financial or reputational risk, steps must be taken to mitigate these risks.

We use a technique called “grounding” to ensure the AI returns facts or data points that transparently reference a specific source, so their assertions can be cross-checked and verified before use. This involves converting the content into a format that the AI can natively query, enabling it to find relevant sections of content based on meaning or intent, rather than searching for specific keywords or phrases that either return too many results or none at all, depending on how the document was written.

In many cases, providing the user with AI tools to simplify finding the desired data points provides sufficient productivity gains by itself. (For a great example of this in action, see: Advanced PDF Analysis with Power PDF Copilot.) However, what if we want to take it a step further and automate the whole process of data discovery from complex documents?

With the next release of TotalAgility, we are introducing a new knowledge base feature that streamlines the process of converting documents into a format the AI Agents can work with.

Any document, regardless of size or complexity, can be loaded into a knowledge base in a single automation step. The AI Agent can then use this knowledge base in a Retrieval Augmented Generation (RAG) pattern to locate relevant information or sections of the document. By employing strategies like intelligent chunking — dividing the document using chapter or section boundaries as opposed to an arbitrary strategy like number of words or pages — we ensure the AI Agent gets complete pieces of content or tables of data to work with, which further the improves accuracy of the AI’s responses. There are many other techniques to further improve the ways in which content is stored or retrieved from the knowledge base, but these are beyond the scope of this article. Feel free to contact us if you’d like to learn more.  

Automated Knowledge Discovery

The true power of the new knowledge base feature is unlocked when it is combined with teams of intelligent agents working together to solve complex problems as part of a broader workflow or case management scenario.

Imagine the research team at a bank that is tasked with analysing thousands of financial statements and annual reports to identify companies to recommend as investments — whether for pension funds, retail investors, or high net-worth individuals.

Each document has similar information contained within it, but the exact wording, layout, structure and terminology will be different from report to report. Specific information, such as balance sheet data, currency hedging positions, and Environmental, Social, and Governance (ESG) programmes, will be relevant to different types of investors depending on their priorities and investment strategies.

Traditionally, a highly skilled team of analysts would spend 4–8 hours per document to extract relevant data, validate it, and then store it where it can be used for writing reports, making recommendations, or supporting large-scale data analytics activities.

As new reports are published at least quarterly, the size of the available analyst team puts a hard limit on the number of companies the bank can research and cover.

How could we improve the scalability of this team using automation?

Research Automation

By creating a team of intelligent knowledge discovery agents, the bank can automate the ingestion, enrichment, and extraction of data from these reports. Here’s how the workflow unfolds, which combines traditional Workflow Automation and Intelligent Document Processing, with the use of AI Agents and AI Knowledge Bases.

1. Document Ingestion: A classic automation workflow to ingest large, unstructured documents (e.g., 500+ page annual reports) into a knowledge base, using advanced chunking, OCR, and vectorization techniques to prepare the data for analysis.

2. Data Enrichment: Agents pre-process the documents, extracting key-value pairs, identifying entities, analysing tables, and generating semantic representations.

3. Question List: Orchestrates the overall process of delegating individual questions to the knowledge discovery agent to answer.

4. Managing Agent: Delegates tasks to worker agents to locate or cross-check specific data points. The managing agent can decide to rephrase a question or task to a worker agent if the initial response is not sufficient.

5. Worker Agents: Highly focused AI Agents that perform specific tasks:

• Semantic Agent: Leverages the knowledge base to perform semantic searches, find relevant sections of the documents, or locate specific facts.

• Data Agent: Searches for specific tables and charts to understand precise metrics and data points.

• Page Agent: Uses vision-capable AI models to read specific pages of a document page for improved understanding of the visual layout.

• Evaluator Agent: Quality control is critical in automation workflows. Evaluator agents are tasked with verifying and cross-checking the outputs of other agents, ensuring accuracy and reducing the risk of errors.

6. Human-in-the-Loop Oversight: If a question can’t be resolved, the response is ambiguous, the process is taking too long, or an unexpected event occurs, then a human expert can be requested to support or correct the AI, helping it learn and improve over time.

7. Output Manager: The final data is formatted and delivered to the desired destination, such as a data warehouse, content system, or CRM.

The result? A process that once took hours per document can now be completed in minutes, with human comparable levels of accuracy, massively scaling the capacity of the bank’s
analyst team.

A New Era of Enterprise Automation

While the new AI Agent and Knowledge base features in TotalAgility are game changers, it’s just the beginning. As businesses continue to push the boundaries of what’s possible with AI, the following trends are likely to shape the future of enterprise automation and knowledge discovery:

• Knowledge Graphs: Expanding from knowledge bases to knowledge graphs will enable agents to understand relationships between entities, unlocking even deeper insights.

• Reasoning LLMs: As AI models continue to improve, their capabilities in advanced reasoning, planning, and problem-solving are unlocking new automation possibilities.

• Cross-Agent Collaboration: Scaling workflows to include hundreds of specialized agents will enable businesses to tackle even more complex challenges.

• Integrated Human Oversight: Responsible AI practices will remain critical, ensuring that humans stay in control of high-stake decisions.

The new capabilities in TotalAgility mark a significant step forward in the evolution of enterprise automation. By enabling businesses to build intelligent knowledge discovery agents, Tungsten Automation is empowering organizations to solve complex problems, drive greater efficiency, and unlock new revenue opportunities.

Whether you’re a business leader looking to streamline operations or a technologist tasked with implementing cutting-edge solutions, the combination of orchestration, low-code development, and AI-powered agents in TotalAgility offers a powerful platform to achieve your goals. The future of knowledge discovery is here — are you ready to embrace it?

Learn more about TotalAgility here and reach out to our team — we’re here to help you every step of the way!