Autonomous AI agents

The next frontier: Autonomous AI agents

At Implement Consulting Group, we conceptualise autonomous agents in terms of human analogues: the heart, mind, hands and legs, each representing different capabilities.

The heart

Just as the heart pumps life into the body, large language models are the engines that power autonomous agents. These agents process complex inputs and generate sophisticated outputs thanks to continuous advances in models like GPT-4o, LLaMA 3, Mistral, Gemini and Claude.

LLMs can be given different roles and personalities as part of the “system prompt” and different instructions. This allows for multiple agents with different specialisations and personalities to collaborate, which is controlled by giving different instructions to the LLM.

The head

A major breakthrough in the practical implementation of LLMs has been in how agents process logic and reason. The mind of the agent ensures that complex problems are broken down to simpler problems, solved one by one until a solution for the main problem is found. Perhaps even with a quality and sensemaking check when each task is completed.

Early LLMs often struggled with even basic logic questions because they were not able to properly break down tasks and reflect on and quality-assure their own answers.

For instance, when asked:

“Which weighs more: 1 kg of feathers or 1 kg of stones?”

The model might have instinctively answered “stones” based on an associative bias rather than understanding the equality in weight. To overcome this, users employed prompt-engineering techniques to help the model think step by step, but this was only a partial fix.

The latest models – one such example is OpenAI’s o1 model – have introduced a more refined approach: chain-of-thought reasoning. In this method, the model breaks complex tasks into smaller logical steps, solving each incrementally. This not only boosts accuracy but also fosters deeper problem-solving capabilities.

The hands

The hands of an autonomous agent refer to its ability to use external tools and “grab” information from outside the LLM. For example, instead of calculating numbers based solely on word predictions, the agent can use a calculator. Similarly, it can leverage APIs to fetch real-time data or integrate with other systems, extending its knowledge base and enabling more precise actions.

One way of doing this is called retrieval-augmented generation (RAG), where the model combines internally generated content with externally retrieved information, often from a company-specific and maintained knowledge base, ensuring both accuracy and relevance in the specific company context. At Implement, we have previously explored this concept in detail (source: Building high-quality RAG systems | Implement).

The legs

The legs of an agent allow it to act in the real world, react based on the new information and collaborate with other agents. Unlike humans, agents are not dependent on solving one subtask at a time. Instead, multiple agents can be deployed to solve a problem. For instance, a financial advisor agent can launch one agent that analyses the market and another to evaluate the portfolio simultaneously.

By working in tandem with other digital entities, agents can also execute roles that require multi-tasking or cross-functional collaboration, much like how humans rely on teamwork to accomplish broader goals.

Making agents collaborate to solve complex problems is currently being implemented in many different forms and frameworks, like OpenAI’s experimental programme “Swarm” (source: https://github.com/openai/swarm), LangChain (source: https://www.langchain.com/agents) and CrewAI (source: https://www.crewai.com). Common for these frameworks is that each agent is assigned a role, such as “researcher” or “data analyst”, which defines their function within the team.

In CrewAI, each agent is also given a goal that drives their decision-making, and a backstory that provides context, making their interactions with other agents more dynamic. Each agent is equipped with a set of tools – for example, web search capabilities or data analysis functions – that help them complete their tasks. A task specifies what needs to be accomplished and assigns an agent responsible for it, along with clear expectations for the outcome. For instance, one task might involve a researcher gathering market data, while a data analyst processes and organises the information for a final report. These tasks are carried out step by step, with agents passing information along the workflow.

Real-world use cases

Now that we have discussed how autonomous agents work, let us look at two examples of how agent solutions are being applied in practice today.

Use case #1: Leveraging autonomous AI agents for master data quality management

In today’s data-driven organisations, maintaining high-quality master data is crucial for operational efficiency, compliance and strategic decision-making. However, master data is often scattered across various systems, incomplete or corrupted due to inconsistent entries or outdated records. In this case, autonomous AI agents were deployed to clean, enhance and maintain master data quality in a cost-effective and efficient manner.

In this use case, AI-driven autonomous agents interact across multiple stages of the data management life cycle to ensure that master data is accurate, consistent and reliable. The agents utilise advanced language models to assess, clean and enrich master data autonomously, reducing the burden on human teams and ensuring continuous monitoring and maintenance of high data standards.

Key agents in action

1. Data assessment agent (“The Analyst”)

• Role: This agent is responsible for the initial evaluation of master data quality. It identifies gaps, inconsistencies and missing information in the data. By comparing the records to predefined business rules and external datasets (e.g. industry standards), the agent assesses the overall health of the data.
• Example: If product descriptions are incomplete, the agent flags them, identifying fields that require updates, such as missing dimensions, colours or product specifications.
• Output: A detailed report highlighting areas that need improvement along with recommendations for further actions.

2. Data cleansing agent ("The Cleaner")

• Role: The cleansing agent automates the process of correcting errors in the master data. It corrects duplicate entries, resolves inconsistencies and standardises the data according to organisational rules.
• Example: If multiple records exist for the same product but with slight variations (e.g. different spellings or formats), the cleansing agent will merge them into a single, accurate record. For instance, it might combine different formats of “KG” and “kg” in product weights into a single standardised unit.
• Output: A clean and unified data set that meets the business’s consistency requirements.

3. Data enrichment agent (“The Enhancer”)

• Role: The enrichment agent augments master data by pulling in external sources of information. This might include adding detailed product descriptions, pulling in colour data from images or even generating dimensions based on drawing files.
• Example: The agent scans product images and uses AI to detect visual attributes such as colour or shape. It then updates the master data records with this new information, enhancing product descriptions and ensuring that they meet all necessary specifications.
• Output: Enhanced data records that are more informative and useful for downstream processes such as supply chain planning or customer service.

4. Data monitoring and maintenance agent (“The Guardian”)

• Role: Once the initial clean-up is completed, the maintenance agent continuously monitors the master data to ensure that it remains accurate over time. This agent alerts the team to any data degradation or emerging issues and can autonomously fix minor errors as they occur.
• Example: If a user mistakenly inputs a new record that duplicates an existing one or uses incorrect formatting, the agent immediately flags it or corrects it autonomously, ensuring that the master data remains pristine.
• Output: Constant data health surveillance, ensuring that the data quality is sustained over time.

Agent collaboration

These agents collaborate in real time to achieve the goal of clean and accurate master data.

For example:

• The data assessment agent identifies gaps in the data set.
• The data cleansing agent then standardises and corrects the identified issues.
• Once the data is clean, the data enrichment agent adds value by enhancing the data with additional context or attributes.
• Finally, the maintenance agent ensures that the data remains in good health by continuously monitoring it for issues.

This interaction between agents ensures that no part of the data management process is overlooked, creating a comprehensive solution for maintaining data integrity.

Use case #2: SOP complexity reduction at a pharmaceutical company

A major pharmaceutical company was struggling with a highly fragmented and complex set of standard operating procedures (SOPs). This complexity accumulated over time as various teams and departments created their own procedures, leading to redundancy, inefficiency and increased compliance risks.

To tackle this challenge, the company employed a generative AI solution, leveraging the power of autonomous agents. These agents were tasked with streamlining, reviewing and refining SOPs, resulting in a more cohesive and standardised set of documents.

The solution

1. Iterative document merging and refinement

• Autonomous agents analysed the existing SOPs, identifying overlapping or redundant procedures. Through natural language processing (NLP), the agents merged similar documents and refined them into a unified version that retained the necessary compliance and operational requirements while eliminating unnecessary complexity.

2. Agent-based collaboration

• The solution relied on multiple AI agents working in parallel to perform distinct tasks. One agent focused on identifying document overlap, another on revising and rephrasing the content to ensure clarity and consistency, and a third agent checked the refined SOPs for compliance with regulatory standards.
• These agents worked independently yet collaboratively, much like a team of human experts, but at a significantly faster pace and with a higher degree of accuracy.

3. Improved compliance and reduced process complexity

• By deploying AI agents, the company reduced the overall number of SOPs and simplified the remaining ones. The new SOP framework became 15-30% less complex, depending on the operational area, resulting in easier navigation for employees and smoother adherence to compliance standards.
• By using AI-driven document analysis and refinement, the company not only achieved immediate gains but also set a foundation for continuous SOP improvement going forward.

Taking the first steps towards deploying agents in your organisation

At Implement Consulting Group, we advocate for a gradual, structured approach to integrating AI as well as autonomous agents into your business processes.

Our 3T’s Framework – tools, training and task force – ensures a seamless transition:

1. Tools: Start by equipping your organisation with the right platforms. Start automating simple tasks, and as the technology matures, expand its role to more complex areas.
2. Training: Employees need to understand how to collaborate with agents effectively. Meanwhile, agents should be fine-tuned to work with your organisation’s unique workflows and data systems.
3. Taskforce: Create a dedicated team to oversee the deployment and evolution of agents, ensuring that they align with business goals and continuously improve their performance.

By following this framework, businesses can unlock the full potential of autonomous agents, driving not only productivity gains but a fundamental transformation in how work is conducted.

Conclusion

Autonomous agents are not just a technological innovation – they are a paradigm shift. Deploying them is not science fiction. Leading companies are already deploying AI systems in this way. By understanding their capabilities and integrating them into your workflows, your organisation can stay ahead of the curve in this rapidly evolving digital landscape.