AI agents are the hottest topic in technology right now. But those eager to experiment face an overwhelming maze of hype, conflicting ideas, competing platforms and tricky technical and ethical challenges.
Over a series of articles, I’ll define the questions, practical steps, and decisions that need to be addressed to embark on this journey.
Perhaps the most exciting aspect of agents is that although they represent bleeding-edge technology, they’re simple enough for anyone to understand and start working with.
This means you don’t need to be an AI specialist, data scientist or software engineer to solve problems with AI, which just a few years ago would have taken some serious hard coding and technical skills.
AI agents provide the opportunity to automate complex tasks and even entire workflows, but also raise important questions and critical risks, and I’ll cover all of this, too.
But let’s start with a quick refresher in case anyone isn’t entirely sure how agents are different from other automation software or ChatGPT-style chatbots.
What Are Agents And What Can They Do?
All software applications essentially automate tasks by following fixed, coded instructions. So while they can be (and often are) used to deliver automation, they aren’t truly autonomous in their own right.
AI Chatbots (like ChatGPT, for example) go a step further. We give them prompts that they interpret in order to work out what we want and find solutions. Generally, however, they process one instruction at a time, then wait to be told what to do next. This means they’re great at answering questions or generating content, but they can’t plan and execute complex multi-step processes.
AI agents, on the other hand, go a step further still. They understand goals, work out how to achieve them, and work towards them with minimal human intervention.
Like software apps, they can be used to design complex, automated workflows. And like chatbots, they’re powered by large-language models, and we instruct them in natural human language.
However, they’re also capable of connecting with external tools, data sources and systems. They’re also “always-on”, and run continuously to manage ongoing processes, rather than handling prompts in a “one and done” in a chatbot-like “one and done” way.
The potential of agents is virtually limitless, from managing industrial processes like production lines to handling customer service calls or acting as “always-on” personal assistants, proactively organizing information and executing tasks as needs arise.
Sometimes they’re described as “virtual workers”, although personally I am cautious of this line of thinking, with its implication that they are in some way interchangeable with human workers, or a tool that can be used to replace them (more on this very important point later!)
So, How Do I Get Started?
With the possibilities being almost unlimited, simply getting started can seem daunting—there are so many things agents can do, and so many ways to go about getting them to do it. How do we decide which is best?
And be warned, there are also many ways things can go wrong, and a technology intended to save time and effort can end up costing you both, as well as leaving you open to some unpleasant risks.
So here’s my simple five-step guide to getting started, covering the basic practical steps anyone can take if they want to start experimenting and learning today.
This is only meant to be a high-level overview; think of it as an action plan for learning the basics. I’ll be covering every stage in more depth in later articles, as well as taking a deeper look at the safety issues and ethical concerns that are only briefly highlighted here.
1. Pick A Use Case
The very first step is to have a clear idea about what you want to use AI agents for.
Start by thinking about simple, repetitive, regular or daily tasks that take up human time that could be better spent elsewhere.
In business, this could include:
- Compiling weekly sales reports to understand business performance
- Summarizing meeting notes and sending follow-up emails with next-step actions to attendees
- Responding to customer support requests and solving customer problems
- Analyzing competitor activity and understanding its implications for your business
I also believe that many of us will increasingly use agents in our personal lives for day-to-day activities where our attention can be better spent elsewhere. This could include:
- Managing our calendars and schedules
- Managing personal finances, such as budgeting or tracking income and expenditure
- Making travel arrangements, planning itineraries and searching for the best deals on flights and transport
- Researching buying decisions in complex markets
The key here is to start small. Pick a simple, repetitive process that you understand well, and where it will be easy to measure the agent’s performance, speed and accuracy against your own.
If there’s a task that you realize you spend too much time on, which requires a relatively simple decision-making process, and you’ve often thought it would be great if you could just get a computer to do it for you, then it’s probably a good choice!
2. Choose Your Tools
Once you have a clear use case in mind, the next step is to pick your tools.
If you’re setting out to do this yourself to learn the ropes, you’ll probably want to start by using a platform that lets you build agents using a simple, language-based interface. Some examples include OpenAI’s AgentKit, Microsoft CoPilot Studio, Salesforce Agentforce and Amazon Bedrock Agents.
They all differ in terms of interface and what jobs they’re suitable for; GPTs are great for quick, lightweight automations like drafting emails, Copilot Studio is great for workflows involving commonly-used Microsoft business tools like Teams, Word or Excel, and Agentforce is designed for working with sales and customer data. I’ll dive deeper into the specifics of choosing a platform in a later article.
For those with technical and coding skills, or working with developer support, there are also a number of low-code options that give a little more control over how agents will work and behave. These include Autogen, Langchain and CrewAI.
As far as these go, the principles of using them are much the same, although they may allow more granular control over your agent’s behavior if you’re willing to get your hands dirty writing a little code.
But for the purposes of the simple, quick-win use cases, I’d recommend starting with one of the more straightforward, visual platforms mentioned above, which is often more than adequate.
3. Prepare Your Data
This means making sure that all the information your agents will need to complete their task will be accessible, and that it’s clean, up-to-date, accurate, and free from bias.
You’ll need to make sure that allowing your agents to access it isn’t going to cause any data protection issues, and that all the correct safeguards and permissions are in place.
Start by identifying where the data that the agent will need lives; this could be in spreadsheets, Cloud drives, your sales or customer data records, or, in the case of third-party data like market data, social sentiment, demographic or meteorological information, in third-party services you can subscribe to and connect to your agent via an API.
Then, ensure your agent can reach it. This means telling your agent-building platform where it is, configuring APIs if necessary, and setting up relevant access permissions.
Next, clean up the data. Ensure it’s in a format the agent can read, remove duplicate or outdated records, and make sure labelling, date formats and naming conventions are standardized.
Agents are often intelligent enough to work around issues like unevenly formatted dates or non-standardized labeling. But as your projects get more sophisticated, this can start to cause problems, so following principles of good data stewardship from the start is a wise move.
AI can help with this, and there are ways to automate data preparation and cleansing that will certainly save time when you’re working on larger projects.
But for the purposes of learning, I’d recommend working with a simple, sampled dataset where you’re sure everything is in order, and where no one will get upset if you make a mistake or misconfigure something, resulting in data being used in a way that it shouldn’t.
4. Define The Workflow
The next step is to define your agentic workflow. This is the sequence of actions that your agent will take to accomplish its task.
Most platforms let you do this by defining basic parameters such as inputs, tasks and outputs, although the terminology here might vary depending on the platform you’re using.
Inputs can be triggers, such as information in a database changing, or an event being detected, that tell the agent that the task needs to be carried out. This could be a customer service ticket or sales record being created, a mention of your company on social media, or practically anything else.
Tasks are specific actions that the agent should perform, such as generating a report, sending an email, conducting sentiment analysis or summarizing an activity that has taken place.
Outputs are the result we expect from the action. This could mean categorizing a support ticket, analyzing a sales event to mine customer behavioral data, or responding to a social post.
The key here is to start by defining the high-level workflow, which is the general sequence of actions that can be applied to the process every time it runs.
5. Iterate, Refine And Scale
Provided you’ve picked a simple, repeatable use case, it should be straightforward to compare the results to those you’ve achieved via manual processes in the past.
Starting on a small scale makes it easier to test, learn and improve in a safe way. Remember, the goal at this stage isn’t perfection; it’s to understand the process behind building an agentic workflow and use it to achieve a small gain in efficiency.
If it isn’t working as expected, make adjustments to the prompts and parameters, or have it start working with a smaller data sample in order to narrow down the potential for mistakes to be made.
Use the activity logging functions of the agent platform you’re working with to understand exactly what it’s doing, and where it might be going wrong.
Once you’ve got the hang of using agents to complete small, simple tasks, you can think about scaling up. This might mean giving it a larger dataset to work with. Or it could mean building more agents to work on separate but related tasks, interacting and collaborating with each other to accomplish bigger jobs.
Rewards, Risks And Challenges
Agents offer enormous opportunities but also create risks. They can work 24/7 without getting tired or making mistakes, and their real value lies in freeing up valuable human resources to work on higher-value and more rewarding tasks.
But they can also pose security threats if we don’t thoroughly understand the data they’re using and what they’re doing with it.
There’s also a risk that human users could over-trust them, failing to provide proper oversight or forgetting how to carry out essential tasks themselves.
Over-reliance on agents could also cause workers to feel devalued or at risk of losing their livelihoods, and careful thought needs to be put into ensuring tasks are delegated and distributed in a way that doesn’t negatively impact the human workforce.
The key here is to always think of them as tools, or assistants, not autonomous employees in their own right. Keep humans in the loop at all times and make sure your agents are augmenting rather than replacing human qualities like creativity, problem-solving and strategic thinking.
I’ve touched on a lot of subjects here, and this article is only intended to serve as a high-level overview of the steps and considerations that should be taken by anyone wanting to make a start on working with AI agents.
I’ll be covering each step of the process, both technological and human-centric, in further articles.
But make no mistake, the agentic revolution is underway, and there’s never been a better time to start experimenting, building and learning if you don’t want to get left behind.
