In the summer of 2024, Jeremy Kirshbaum, CEO of Handshake, an AI research and consulting company, then running an AI agent course on Maven, told me AI agents would propel a lot of AI adoption forward and would be the new big thing in AI development. He was 100% right.
Recently, I met Himanshu Joshi, who works on Applied AI at the Vector Institute, and in addition to his research work, he works with companies implementing AI agents. He considers himself an “AI connector.” With all the marketing hype and overpromising around AI agents, I wanted to chat with him about what is and isn't an AI agent and get more context around AI's new darling.
One of the things I noticed about AI agent platforms is that a lot of them look like existing automation platforms that already exist. Many of them seem like a harder-to-use Zapier. My first question to Himanshu was, what isn't an AI agent?
"AI agents are not just chatbots or rule-based automation. Remember those days of RPAs and simple predictive models, which we have traditionally been using via traditional machine learning. Unlike traditional applications and software, these are...a series of workflows, algorithms, which have been designed to operate autonomously, they can do a lot of things on their own. They can decide on behalf of the user if you have accordingly programmed them. Another thing about them is they can adapt to the changing environment. Based on the inputs you have programmed in the workflow, they will be able to take those decisions as things evolve. Though many companies label any AI-powered system as an agent, I think true agents should essentially exhibit autonomy."
One of the key aspects of what makes an AI agent an "agent" is its ability to make decisions. This means going beyond the most simple "if-then" logic to determine which action to perform next. Usually, it is something that requires the AI to "think" through the context of the decision to make the right choice.
One example Himanshu brought up was around booking air travel. If you tell an agent to book you a flight to a specific destination on a specific day, it will do that, but it is going to have to make decisions around how much you're willing to pay for a flight. In general, whoever creates the agent would prompt the user for that input once they've found flight options and a price range. But let's say this is not a well-built agent and that feedback loop isn't a part of the flow. The agent should probably assume the user doesn't want the most expensive flight possible. It might balance what it thinks is a good time of day to fly from your departure airport, or a good time to arrive at your destination, and balance that with the low to mid-range priced flights. It takes in a lot more input and requires a lot more context and nuance than an automation that is designed for something like checking if a customer has made a purchase in the last 6 months or not.
While what AI agents are able to do is pretty powerful, especially tools like OpenAI's Operator, where they are today, they're not ready to operate fully independently. OpenAI is betting that they will be pretty powerful given that they are launching AI agents that range in price from $2,000/mo to $20,000/mo. Though, how humans work with them dictates their success or failure. Himanshu told me the first time we talked that the number one reason that AI agents don't end up working out in a company is that the employees didn't trust the AI, or product adoption was low due to friction or lack of alignment with how [human] employees work. In the short term, yes, it means training teams how to work with AI. But the fact teams need to be trained at all shows there is an opportunity for AI agents to be designed to work for humans and with humans in a way that is more aligned with a human flow. The way humans do things may not be the most efficient, but as long as AI still needs to work with humans, it does need to be designed in a way that will encourage adoption so companies and people can get the benefits.
"AI agents may fail if we do not give due focus to human-focused challenges. There could be an expectation mismatch," is how Himanshu framed it.
He added, "We had foundation models. We had GenAI, but it was restricted to certain few members of our teams. The entry barrier to use them was a little high…and now the better the UI/UX is, the better it is for everyone to use." The improvements of the UI/UX over time have helped make these tools more accessible already, even if they aren't perfect.
Beyond the interface and interactions needed for adoption, it also needs to be clear to businesses what AI is capable of so they don't have unreasonable expectations for what they can expect and the timeline they can expect it in. I asked Himanshu what tasks are the best for AI agents.
"Where you have a lot of high-frequency repetitive tasks, which can be offloaded to someone else, so those tasks definitely can be offloaded to the multi-agent system where you can task them to do certain things. But I would once again dwell deeper into the point that at the end of it, there should be a human in the loop who's taking the final decision and who's finally perhaps checking it and making sure it's happening the way you would like to make it happen."
Something he clarified that surprised me a little, but also made a lot of sense, is that sometimes it can take a while for the efficiency and cost savings to kick in. With some of the promises being made about AI, some businesses may feel surprised or disappointed that the benefits aren't more dramatic sooner.
Lastly, I asked what are some good resources for people who want to learn more about AI agents.
"Google, Amazon, as well as Microsoft, they have their courses. All these courses are free. So I would encourage you to kind of enroll for them, use them, play around in their playgrounds and see, get your hand a little dirty. Then you have a very interesting course being run by Hugging Face, which is right now also on. That could be another way of learning. Then there are certain advanced courses which are being offered by a lot of universities."
Now is a great time to learn more about AI agents as they're in the very early stages of development and usage.
Thank you so much to Himanshu Joshi for being my guest on the Product Mind Lab this week!
Stay curious.
P.S.:
Fractional Product Leader: I’m a fractional product leader and consultant. If your business is building its first AI features, I have an AI audit offering to show businesses where their best (and worst) are opportunities to implement AI. You can reach out here.
Jumpstart Your AI Career: Jumpstart Your AI Career is BACK on Pearson/O’Reilly will be April 24. Register here.
Share this post