WTF are AI agents?

Despite so much use of the A-word, it’s still early for AI agent adoption, meaning marketers should ask what agents are for, how they’re made, what they do, what they might do — and what they can’t do — including potential reputational risks.  

As tech titans build autonomous bots to tackle the mundane, marketing teams must weigh up how to integrate agents into their existing processes in order to better convert ‘prospects’ to paying customers.

What are AI agents?

Unlike chatbots, which are conversational, AI agents can take action on users’ behalf. They can complete tasks, interact with other software systems, make decisions, and act independently.

Companies building AI agents include Accenture, various advertising agencies and Talkdesk, which debuted new AI agents for retailers during NRF to help with personalization. Another example is Oracle, which just announced a new sales AI agent geared toward removing the admin-laden complexities of the sales process so its IRL team members can focus on “meaningful customer communication.“

Clarifying the difference between AI agents, copilots, LLMs

Agents and copilots are both getting a lot of hype, but they’re not the same. Copilots can collaborate with users but don’t make decisions on their own.

However, agents act autonomously on behalf of users.

Meanwhile, large language models are the foundation for agents; they’re just part of the process. LLMs can generate text, translate languages, provide information, and brainstorm topics. They are also used to power AI agents and complement additional software tools, which give AI agents their actual agency.

What agents can do — and can’t do

David DiCamillo, CTO at Code & Theory, described three different “buckets” for AI agents, see below.

• Overt agents help make decisions and power tools for chatbots.
• Passive agents work behind the scenes to understand data sets.
• Data activation agents then help take insights to help make actions based on them.

DiCamillo further observed how implementing them in safe, accurate, and beneficial ways is a key consideration. For example, where is the data coming from, who controls it, and how do we ensure its accuracy?

He added, “The agent side now has a whole other can of worms for clients: Who’s monitoring these things? What if they cost [clients] dollars? What’s the business impact of this? Then the conversation becomes, ‘Who’s policing these agents?'”

What are some other types of AI agents? 

• Goal-based agents evaluate different types of data and compare approaches based on goals.
• Utility-based agents evaluate actions based on potential options and outcomes.
• Learning agents learn based on various inputs, feedback, and past results.
• Search agents explore data sets and destinations to find information.
• Shopping agents, such as Google’s recently previewed Project Mariner, help people buy stuff.

How do AI platforms enable agents?

Major cloud and AI providers like Amazon, Google, and Microsoft have recently announced updates to help companies make agents — and help multiple agents interact.

In December, AWS updated its Amazon Bedrock platform to help power multi-agent collaboration, improve accuracy, increase speeds, and reduce costs. Earlier this month, Google debuted new agent features for retailers while Microsoft introduced new “autonomous actions” capabilities, gave details for calculating agent costs, and debuted a way to “chat” with AI agents using natural language.

Agents can also help extend the knowledge of AI models, using techniques like retrieval augmented generation to go beyond an LLM’s pre-training data. Agents powered by different LLMs might also collaborate on various tasks, said Paul Roberts, director of technology, strategic accounts, at AWS. He gave an example of a consumer using an AI agent to research products, compare reviews, and find alternative options based on various criteria, such as energy efficiency.

Roberts said, “Imagine a world where there becomes an agent marketplace where all these agents out there doing different tasks, and you start pulling them in to get interesting pieces of content for various use cases.”

Potential challenges and risks?

Building AI agents requires companies to rethink their data strategies. That could include everything from how they collect, clean, and structure data to building new infrastructures to help with real-time data flows and feedback loops. Companies will also have to figure out how to integrate agents into existing processes, whether to change current workflows and how to integrate agents with other AI tools.

Just like with other types of generative AI, concerns include inaccuracies, inconsistencies, and unproven ROI. There are also worries about how agents could create new risks for users, corporate data, biases, and the chance that customer-facing agents might say something that could harm a brand’s reputation.

Are agents proving themselves yet?

It’s still very early, but some companies say their AI agents already are showing results. One example is Twilio, which built a customer support agent to test its own AI assistant framework. The agent, named Isa, has now scaled from handling 2% of marketing leads to more than 50%. It’s also helped improve marketing metrics three-fold, based on how likely customers are to become paying customers after talking with Isa.

“Don’t just go for automating XYZ and looking at the impact to the bottom line,” said Kat McCormick Sweeney, who leads Twilio’s go-to-market team for emerging tech and innovation. “What is the best customer experience? What’s a customer journey that wouldn’t be possible before because we didn’t have unlimited human resources? Now you actually can infuse agents along your customer journey to create an experience that’s better.”