7 Strategies To Get an Effective AI Agent Up and Running

Luis Lopez Garay
5/5/2025
7 Strategies To Get an Effective AI Agent Up and Running

You’ve probably heard how AI agents can increase productivity. Our strategies can help you go from just talk to true business transformation.

Businesses that want to take their AI technology to the next level are talking about AI agents, which can do much more than give organizations a technology boost. Unlike traditional computer programs that require explicit instructions for every step, AI agents have the autonomy to decide what actions to take based on the information provided to them. Using these agents offers organizations the opportunity to automate their more intricate workflows, but how can companies go from talk to action? Our AI team offers seven strategies to help develop successful AI agents.

Explore AI transformation services 
Boost your business with quick wins and tailored AI solutions from our AI transformation team.

Transfer institutional knowledge: Treat the agent like a new employee

Anyone who has onboarded a new employee quickly realizes how much institutional knowledge is required to effectively perform a role. Current employees – especially those with longevity at an organization or in a profession – inherently understand processes, roles, sources of information, and nuanced industry terminology. Much like training a new employee, building an AI agent requires that same transfer of institutional knowledge, using a two-step process:

  • Create a glossary of terms. Companies should create a list of words they use in their organization and in their industry that are not standard terms – or that are terms used in a nonstandard way – and define those for the AI agent.
  • Document institutional knowledge. Organizations should document the institutional knowledge necessary for performing tasks, especially the knowledge that’s typically passed down verbally. For some companies, it can be helpful to conduct a variety of interviews with different subject-matter specialists to gather as much of that knowledge as possible and then run transcripts of those interviews through an AI tool to summarize the information.

Define scope and objectives: Start small and scale

When building an AI agent, it’s crucial to start with the smallest possible task, adjust for errors, and then build on the successes. The smaller and more specific the initial scope, the more efficiently and accurately the agent can perform. Scope creep is the enemy of effective AI implementation.

How are we doing this at Crowe?

Following are a couple of real-life examples that show how our AI team put this strategy into practice for clients:

  • Order fulfillment. For one client, rather than building an agent to manage an entire order workflow, we started by training it to identify and categorize product stock-keeping units (SKUs). Once the agent excelled at this foundational task, we expanded its capabilities to extract quantities, shipping addresses, and other order details.
  • Contract review and renewal. For another client, we directed the AI agent to identify and flag specific clause types within standard contract templates first, and then, as performance improved, broadened its scope to compare clause language against an approved library and to provide risk scoring.

Choose the right AI technologies: Start with out-of-the-box solutions and max out their capacity

Not every AI agent has to start as a bespoke, custom-built solution to be successful. Starting with low-code solutions allows companies to test – and max out – capabilities before moving to agents that need more complex coding. This way, organizations can validate use cases, identify gaps, and build organizational experience before making larger investments. A variety of tools on the market offer a valuable on-ramp for experimenting with AI agents in a low-risk environment.

How are we doing this at Crowe?

Following are a couple of real-life examples that show how our AI team put this strategy into practice for clients:

  • Product location. We used a client’s existing Microsoft™ technology stack and integrated generative AI (GenAI) to identify, extract, and search product and part information from structured and unstructured documents. Then, we developed a Microsoft Copilot agent to provide users with an intuitive and accessible way to search for products and parts, increasing efficiency.
  • Data extraction and entry. For a real estate client that wanted to streamline the leasing process, we used the Microsoft technology stack to integrate GenAI into the workflow. We built an app using the Microsoft Power Apps™ and Microsoft Power Platform™ software that could extract all text from the lease, process the text through a custom-engineered prompt designed to identify and extract relevant details and terms, and store the relevant data.

Train the AI agent with quality data: Curate data to remove contradictions and limit access

When training an AI agent, data quality is paramount. Throwing a vast amount of data at the agent will just confuse it – much like a human employee. So, how do companies know the right amount and type of data to provide?

  • For unstructured data, including documents, images, and meeting recordings, companies can provide only the most relevant information, screening for conflicts and outdated procedures.
  • For structured data, including databases and systems, companies can give the agent prewritten queries to control access. At best, allowing the AI agent free will to access all structured data can cause confusion. At worst, AI agents could inadvertently leak or publish sensitive data.

How are we doing this at Crowe?

Following are a couple of real-life examples that show how our AI team put this strategy into practice for clients:

  • SKU locator. A client provided an export of all the parts listed in its enterprise resource planning (ERP) system. From there, we created a code base to process and called on GenAI to extract part information from unstructured PDFs, specification sheets, and manufacturer catalogs. Our team then formatted the data sets into structured CSV files and uploaded them to separate Microsoft Azure™ search indexes using Azure AI Studio, which enabled efficient character searching of the structured and unstructured data fields. 
  • Fixed asset manager application. Traditionally, assigning depreciable lives for fixed assets requires significant expertise and manual effort, with tax teams spending hours reviewing raw files. For one client, our solution resolved these challenges by automating key aspects of the process, locating key data using Azure AI Builder and using identified guidelines to classify assets and assign depreciable lives. A portion of user-validated responses are also linked to internal systems and used to improve responses in future runs.

Integrate the AI agent across business systems: The most vital strategy for success

The most crucial aspect of deploying an effective AI agent is integrating it into existing business systems because an AI agent’s true value is in how easily it integrates with emails, documents, databases, customer relationship management (CRM), ERP, and other core platforms. A variety of emerging technologies are simplifying AI system integration by allowing large language models (LLMs) to connect with various data sources and applications, but organizations still need to plan for integration by:

  • Making a comprehensive list of all the systems the AI agent will need to access
  • Developing a strategic integration plan with prioritization
  • Determining if the company has the right people and technologies to enable these integrations – including a new open-source and universal communication protocol, Model Context Protocol, which is designed to improve LLM access to enterprise data

How are we doing this at Crowe?

Following are a couple of real-life examples that show how our AI team put this strategy into practice for clients:

  • AI-driven workflow. A private equity client relied heavily on analysts to review lengthy documents and extract key details and metrics. Our team used the Microsoft technology stack to build an automated workflow that is triggered when a user emails a confidential information memorandum document to a dedicated inbox. The AI solution can extract the document text, input the extraction into a custom prompt, analyze the prediction output by extracting key entities, and contain an integration with the Salesforce CRM platform.
  • Search efficiency. To streamline the part-finding process for a client’s sales team, we developed a Copilot agent that integrates with the Azure search indexes via a Microsoft Power Automate™ flow to process user queries and return relevant parts based on the indexed data. The agent deploys in Microsoft Teams™ software, and it is easily accessible across the organization.

Implement oversight and governance: Have humans in the loop

While AI agents offer immense potential to amplify productivity, it’s crucial to maintain human oversight and implement robust governance controls. Even as AI technology advances, AI agents can still make mistakes, hallucinate, or veer off course, which is why it’s vital to designate a human in the loop to help prevent mistakes and course-correct missteps. Blindly trusting an AI agent’s outputs without review could lead to a variety of issues, including:

  • Costly errors that can easily snowball
  • Compliance violations, especially in highly regulated industries
  • Reputational damage that can create lasting issues

How are we doing this at Crowe?

Following is a real-life example that shows how our AI team put this strategy into practice for a client:

  • Built into the process. As part of creating a fixed asset manager application for a client, our team built in a human-in-the-loop validation mid-run to help determine accuracy and fix any errors before the run continued. This design balances automation with the need for human oversight and helps maintain compliance by allowing tax specialists to review and correct classifications as needed.

Optimize and scale: Consider AI agents an iterative process

If thoughtfully deployed, AI agents have the potential to dramatically amplify human productivity. An iterative approach allows teams to identify and resolve edge cases, expand the AI agent’s scope responsibly, and progressively automate more complex workflows over time. Rather than trying to implement wholesale automation, AI agents are best used to augment human knowledge workers by managing tedious, repetitive tasks at scale, which can free up time for higher-value activities that require human judgment, creativity, and decision-making. To realize these productivity gains, organizations should:

  • Embrace a new workforce paradigm built on human-AI collaboration
  • Allocate tasks among humans and AI agents based on their respective strengths
  • Define oversight protocols, quality control, and higher-level strategic direction

Building an effective AI agent requires an appetite for exploration, a willingness to make upfront investments, and an acknowledgement of where internal knowledge and resources are lacking. For many companies, working with a third party – like Crowe – that offers deep expertise in AI technology and broad experience implementing AI across industries can be a worthwhile investment.

The most important thing companies can do? Get started on your AI journey now and continue to invest as the technology develops.

Contact our AI transformation team


Our AI transformation team delivers tailored AI solutions – including AI agents – that can help companies solve problems and optimize performance. Contact us today to start your company’s AI transformation.

Contact us today

Luis Lopez Garay
Luis Lopez Garay
Partner, AI Transformation

Microsoft, Azure, Microsoft Teams, Power Apps, Power Automate, and Power Platform are trademarks of the Microsoft group of companies.