Much of the education around AI over the past three years has revolved around how to empower your employees to use AI effectively to prompt the platform to get the highest quality results. As the thinking process of these platforms has improved, a new, parallel path to business evolution has emerged. This is deploying specialized AI agents that work concurrently with their human colleagues. They don’t need a human to enter a prompt to initiate; they act when the right circumstances are met. AI agents are tasked with responsibilities and given tools and rules to carry them out.
Just like the varying expertise and responsibilities of a teammate in our organization, AI agents specialize in carrying out a range of work and don’t need to be overly complex to add value. It could be as straightforward as an HR agent answering employee questions about your organization’s policies and procedures. Or an assistant watching your inbox for new emails and drafting replies based on rules you define, writing in your voice and tone. It can also be a system of multiple agents working together on business leads by monitoring web forms, enriching leads, and creating CRM tasks for your team to follow up on.
Some of the biggest competitive advantages will come from businesses who are able to deploy AI agents in their operations to automate and complete tasks. Advantages can turn into liabilities quickly when agents are introduced without oversight. Strong governance is required. You need to have insight into their level of access and autonomy, understand the value they’re delivering, and they need to be monitored. One approach is to introduce AI agents the way you’d onboard a new hire.
High-stakes autonomous systems have been deployed in validated and supervised scenarios for a long time. Let’s take the commercial airline autopilot as an example. “Autopilot” is comprised of multiple systems in an aircraft that can take over certain aspects of a flight from a pilot and copilot. There’s the plane’s “cruise control” or basic autopilot functions that hold a specific heading or altitude. There there’s the Flight Management System, which navigates a pre-planned route but requires a pilot to manage weather or traffic. Then there’s the most advanced, the Autoland System, which can land a plane in conditions where the pilots have zero visibility, and manage thrust, flaps, and steering simultaneously.
These systems run under the watchful eye of the pilot and copilot. Just like autopilot, business agents should be deployed in levels starting with assistance, then supervised execution, then limited autonomy. Each level should come with monitoring, clear boundaries, and a way to shut it off quickly.
Onboarding AI Agents
Let’s walk through onboarding our email draft agent. First, what is the agent’s job description? “Your role is to monitor an inbox and prepare draft replies to qualifying messages, aligned with the inbox owner’s communication style.” This captures the intent of the agent, and much like a job description, it doesn’t tell you how to do the job.
Next, we need to define who is the agent’s supervisor. This is typically the worker who works alongside the agent and decides how to handle its output. In this example, the responsibility belongs to the inbox owner since they’ll be the ones deciding if the reply draft can be sent. If there is an issue with the performance of the agent, they must escalate issues and provide feedback to management (the technical owner of the agent, typically IT).
Onboarding needs to include an operations guide for the agent. The beauty of leveraging an AI powered agent rather than traditional email inbox rules is we can be more general in our instructions and cover many more cases. For example, my filtering instructions say to exclude “newsletters, bulk marketing, meeting invites, and email that has no-reply senders.” Writing traditional inbox rules for “newsletters” and “bulk marketing” would be tricky. Those messages tend to not include these words in the body. Yet identifying emails with these criteria is an easy task for today’s AI models.
The agent needs to understand how to communicate as we would while keeping the audience in context. AI can sound a little generic if not given direction. Every business has its own customer voice. A new customer service hire must learn it, and your agent does too. To teach the agent, we’ll create instructions based on previous email communications.
Here’s an example prompt: “Analyze my recent emails and chat messages from the last three months to identify my communication style. Summarize key traits, tone, and common phrases into a short guide that helps AI write in my voice for professional and casual contexts.” Now this guide can be called upon by the agent when it needs to draft a reply.
Setting Limits
Many traditional job descriptions have that famous clause at the end that says, “other duties as required.” With AI agents, we don’t want them to do anything extra. Think of the New England Patriots’ past mantra of “Do Your Job.” We want boundaries in place for the agent, so their actions are limited.
With our email agent, this is easily accomplished by only allowing it access for drafting emails but not sending them. We’ll instruct it to never send a reply. We can also direct our agent to avoid including sensitive data in all drafts. We can limit the actions taken and how data is handled to reduce the risk of mistakes.
This agent will need to be held accountable for its performance. In most organizations, IT becomes the agent’s “manager,” monitoring performance, adjusting configuration, and controlling access. They’ll coordinate an initial testing phase with the supervisor since validating the agent’s operation and output will be crucial. This is like a traditional 90-day trial period of a new staff member.
The business will need to understand the value that this agent brings. How many emails are being successfully drafted? What’s the estimated time savings? Not having performance and success metrics in place makes it difficult to understand where investments should be made or redirected.
With that, we’ve onboarded our first AI agent. We’ve chosen a low-risk role that can still save meaningful time across our team. We could even give our agent a promotion, adding new functionality to assist in drafting new messages. From here, you have a repeatable way to safely onboard new AI agents with increased levels of autonomy, handle more complex tasks, and deliver even more value to the business.
Chris Toppin is chief service officer with Mainstay Technologies, an IT firm with offices in Manchester, Laconia as well as Burlington, Massachusetts, that provides managed IT, information security, AI and automation, assessment, cybersecurity maturity model certification and vCIO services. For more information, visit mstech.com or email ctoppin@mstech.com.
