AI agents: The opportunity and the bull$#!%

AI co-workers, AKA agents: “60% of the time, it works every time.”

Picture this: You stroll into your office, coffee in hand, and your AI co-worker, let’s call him “Data McDataface,” greets you with a personalized marketing report, complete with insights and recommendations tailored to your latest campaign goals. He’s already crunched the numbers, analyzed customer sentiment, and even drafted a few social media posts for your approval. Sounds like a marketer’s dream, right?

Well, hold your horses, buckaroo. Data McDataface ain’t quite saddled up for that rodeo yet.

The vision of AI agents and co-workers is undeniably compelling. It promises a future where mundane tasks are automated, creativity is amplified, and marketing teams are empowered to achieve superhuman levels of productivity. But the reality is a bit more nuanced.

While AI is making incredible strides, we’re still in the early stages of developing truly autonomous and intelligent agents that can seamlessly integrate into our workflows. There are technical hurdles to overcome, ethical considerations to navigate, and a whole lot of hype to sift through.

In this blog post, we’ll explore the current state of AI agents and co-workers, delve into the exciting possibilities on the horizon, and, as always, call out the BS. Get ready for a candid and insightful look at the future of work in the age of AI.

If you feel like listening to someone way smarter than me articulate why agentic systems are important to understand, this interview with Brett Taylor is great!

Let’s get started!

What AI agents can do NOW (spoiler alert: they don’t exist yet…and Salesforce Agentforce definitely isn’t one)

Okay, let’s cut the marketing BS and get real. True AI agents? Those mythical beings of marketing automation? They don’t exist. Not yet, anyway. (And let’s be honest, even with all the glitz and glam of Dreamforce, Salesforce Agentforce is just a fancy wrapper around Mulesoft. It’s like putting lipstick on a pig…a very expensive, enterprise-grade pig.)

We’re still on that long and winding road to Artificial General Intelligence (AGI), that sci-fi dream where AI can finally understand a joke, plan a marketing campaign without screwing it up, and maybe even make us a decent cup of coffee.

OpenAI, has one of the better definitions of what progress towards a truly agentic AI world actually looks like.

Defining a true AI agent (because buzzwords are confusing)

An AI agent is more than just a shiny tool or a bunch of fancy features. It’s an autonomous entity that can:

Understand and interpret natural language: It can actually comprehend what you’re asking, even if you’re being sarcastic or using complex marketing jargon.

• Status: 89% Nailed it! Models are pretty solid at this today.

Access and process information from various sources: It can sift through the internet, your dusty old databases, and even those fancy knowledge graphs to find the information you need.

• Status: Doing alright! This is fairly reliable these days.
• Perplexity is an example of a service crushing this.

Make decisions and take actions based on its understanding: It can figure out the best way to achieve a goal, whether it’s writing a killer email campaign, soothing an angry customer, or (gasp!) planning a multi-step marketing strategy without needing a human babysitter.

• Status: Ehhh…not so much. We’re working on it.

Learn and adapt over time: It can learn from its mistakes (and hopefully not repeat them a thousand times).

• Status: Meh. Deterministic adaptation of feedback mechanisms within LLMs (outside of RLHF) is still pretty underwhelming.

Reason and plan across multiple steps: It can break down complex tasks into smaller steps, like a marketing project manager who actually knows what they’re doing.

• Status: OpenAI’s o1-preview is showing some promise here, but multi-step reasoning is still the AI’s Achilles’ heel.

Self-correct and recover from errors: It can identify when it’s screwed up and fix it, without needing a human to come in and clean up the mess.

• Status: See above. Still a work in progress. Although multi-agent loops are making progress here, AgentGen agent-to-agent feedback is an example.

Utilize tools effectively: It can use different tools and APIs to get things done, like a digital Swiss Army knife.

• Status: We can do this today! (well, I mean it won’t get you coffee, but it can call an API…sooo)

The bottom line: AI agents are still a work in progress (but we’re getting there)

But what is there?

Task automation for fairly straightforward automation is there; think of it as the freshman intern of agents.  I’ve personally been pretty impressed with Zapiers deployment across medium, but there are quite a few decent solutions here, and function calling within jigs can get some pretty agentic-like behaviors, too.  Especially when you pair them with a Slack API…I may or may not have a team of these helping me out today.

So, while we might not have fully realized the dream of AI co-workers just yet, the future is undoubtedly agentic. Stay tuned because the next chapter in the AI revolution is about to begin. (And hopefully, it involves less buzzwords and more actual intelligence.)

Under the hood: How AI agents work (it’s more than just chat)

While the conversational interface is the most visible part of an AI agent, there’s a lot more happening behind the scenes to enable those seemingly intelligent interactions. Let’s break down the key components:

The interface

How can humans interface with this digital employee?

• The OG, chat interface, let’s be honest, Slack is how you chat with your co-workers today anyway.
• Multi-modal interfaces: voice and vision are becoming increasingly common as the bridge between the human world and the AI world.  Image input is a key component to Anthropic new computer use technology.

This is the part you interact with, whether it’s a chat window, a voice assistant, or even a virtual avatar. It’s responsible for understanding your natural language queries and translating them into a format the AI can process. Advancements in natural language processing (NLP) are making these interfaces increasingly sophisticated, capable of understanding complex requests and even picking up on nuances like tone and sentiment.

Tool calling: The AI’s Skillset

Behind the conversational interface lies a powerful arsenal of tools that the AI can call upon to fulfill your requests. These tools might include:

• Large Language Models (LLMs): For generating text, summarizing articles, translating languages, or even brainstorming creative ideas.
• Permissioned Writable APIs – For writing generated data to external systems.
• Data Analysis Tools: These are used to process and analyze large datasets to extract insights and patterns.
• Document Creation Mediums – Antropics Artifacts UX is the best pattern I’ve seen here.
• External APIs: For accessing information or services from third-party providers, like weather forecasts, stock prices, or even booking flights.
• And the huge “maybe might be the path here”, is complete computer use as a tool…meaning you let the agent use your mouse and keyboard as tool, and a multimodal input as the eyes.

The AI agent uses sophisticated algorithms to determine which tool is best suited for each task, then calls upon it to generate a response or perform an action.

Information synthesis and knowledge graphs: The AI’s brainpower

To provide meaningful and contextually relevant responses, the AI agent needs to be able to synthesize information from various sources and understand the relationships between different pieces of data. This is where knowledge graphs come into play.

A knowledge graph is a structured representation of information that captures entities, their attributes, and the relationships between them. Think of it as a vast network of interconnected concepts and facts. By leveraging knowledge graphs, AI agents can:

• Understand the Context: Connect your query to relevant information within the knowledge graph, ensuring that the response is accurate and meaningful.
• Reason and Infer: Make logical deductions and inferences based on the information in the knowledge graph, even if the answer isn’t explicitly stated in your query.
• Provide Personalized Responses: Tailor responses based on your individual preferences, past interactions, and other relevant data stored in the knowledge graph.

Putting it all together: The agentic loop

AI agents’ real power lies in combining these components into a seamless and interactive experience. Here’s a simplified overview of the agentic loop:

1. User Input: You provide a query or request through the conversational interface.
2. Natural Language Understanding: The AI processes your query and extracts its intent and key entities.
3. Planning: Using advanced reasoning to create a stepwise flow to follow to execute complex taks.
4. Contextualization and Information Retrieval: The AI leverages knowledge graphs and other data sources to understand the context of your query and retrieve relevant information.
5. Tool Selection and Execution: The AI determines which tool is best suited for the task and calls upon it to generate a response or perform an action.
6. Response Generation: The AI generates a natural language response that’s both informative and contextually relevant.
7. User Feedback: You provide feedback on the AI’s response, helping it learn and improve over time.

What’s on the horizon (AI agents are about to get real)

We’re on the cusp of a new era where AI agents will become truly integrated into our workflows, acting as collaborators, advisors, and even creative partners. Here’s a glimpse of what’s coming:

1. Multi-step reasoning and planning (no more one-trick ponies)

One of the biggest limitations of current AI agents is their inability to handle complex, multi-step tasks. But that’s changing rapidly. Researchers are developing new techniques that enable AI agents to:

• Break down complex goals into smaller sub-tasks: Imagine an AI agent that can automatically create a multi-channel marketing campaign, including writing copy, generating images, scheduling social media posts, and even analyzing performance data.
• Reason about dependencies and constraints: AI agents will be able to understand the relationships between different tasks and make informed decisions about the order in which they should be executed.
• Adapt to changing circumstances: If something unexpected happens (like a sudden shift in customer sentiment), the AI agent will be able to adjust its plan accordingly, ensuring that the overall goal is still achieved.

Tools and Research Driving This Advancement:

• Open AI o1: This approach focuses on training AI agents to follow a set of principles or “constitution” that guides their behavior and decision-making, enabling more reliable and ethical multi-step planning.
• Google AlphaProof: This LLM has initially been deployed against solving very complex mathematical problems, however, there is rumored to be an internal initiative to pull much of the research into the Vertex calls LLM models with Google; we’ll see if that happens in 2024.
• Meta LLama3.x: While not as robust as o1, it shows glimpses of reasoning, and it’s open source, which in my book is always positive.

2. Enhanced contextual understanding (AI that “gets” you)

AI agents are becoming increasingly adept at understanding the context of your requests and providing relevant and personalized responses. This is thanks to advancements in:

• Knowledge Graph Integration: AI agents are being integrated with more sophisticated knowledge graphs, allowing them to access and process a wider range of information and understand the relationships between different concepts.
• Long-Term Memory: Researchers are developing techniques to give AI agents a “memory” of past interactions, enabling them to personalize responses and provide more contextually relevant information.
• Multimodal Understanding: AI agents are being trained to understand not just text but also images, videos, and other forms of data, allowing them to grasp a more holistic view of the situation.

3. Seamless integration with human workflows (the AI co-worker you always wanted)

The ultimate goal of AI agents is to seamlessly integrate into our workflows, acting as true collaborators and co-workers. This will require:

• Improved User Interfaces: AI agents will need to be accessible and easy to interact with, whether through natural language, visual interfaces, or even augmented reality.
• Proactive Assistance: AI agents will anticipate our needs and offer assistance before we even ask for it, like suggesting relevant data, generating content ideas, or automating routine tasks.
• Collaborative Problem-Solving: AI agents will be able to work alongside humans to solve complex problems, offering insights, generating solutions, and even challenging our assumptions.

Integrated interfaces: we played with this at Cordial quite a bit when doing our own RD on a marketing agent. Integration into natural workflows such as Slack saw much higher adoption and iteration. By the way, our R/D hit the same roadblocks as everyone else. On complex workstreams, “80% of the time every it works every time”

The BS meter (because we’re not falling for the hype)

Okay, let’s inject a dose of reality into the AI agent hype. While the future is promising, there’s a lot of overblown claims and misconceptions floating around. Here’s your BS detector for navigating the AI agent landscape:

1. “AI agents will replace humans entirely” (don’t panic just yet)

Sure, AI can automate tasks, analyze data, and even generate creative content. But let’s not get carried away. AI agents lack the critical thinking, emotional intelligence, and nuanced understanding of human behavior that are essential for many marketing roles.

Why this is BS:

• Strategic Thinking: Developing a comprehensive marketing strategy requires a deep understanding of your brand, your customers, and the competitive landscape. AI can provide insights and recommendations, but it can’t replace the human element of strategic planning.
• Creativity and Innovation: While AI can generate creative content, it often relies on existing patterns and trends. True innovation often comes from human intuition, imagination, and the ability to think outside the box.
• Empathy and Emotional Intelligence: Connecting with customers on an emotional level requires empathy and understanding, qualities that AI is still developing.

2. “AI agents will be perfect and never make mistakes” (in your dreams)

AI agents are still in their early stages of development, and they’re far from perfect. They can make mistakes, misinterpret information, and even generate biased or offensive content if not carefully monitored and trained.

Why this is BS:

• Bias and Fairness: AI models are trained on data created by humans, which means they can inherit our biases and prejudices. It’s crucial to be mindful of this and ensure that AI agents are used responsibly and ethically.
• Error Propagation: As we discussed earlier, AI agents can struggle with multi-step tasks, and errors can propagate through the system, leading to unexpected and potentially harmful outcomes.
• Lack of Common Sense: AI agents often lack common sense and may struggle to understand the nuances of human communication and behavior.

3. “AI agents will solve all your marketing problems” (if only it were that easy)

AI agents can be powerful tools, but they’re not a magic bullet. They can’t replace the need for a well-defined marketing strategy, a talented team, and a deep understanding of your customers.

Why this is BS:

• Data Dependency: AI agents rely on data to learn and make decisions. If your data is incomplete, inaccurate, or biased, the AI agent’s performance will suffer.
• Integration Challenges: Integrating AI agents into your existing workflows and systems can be complex and require significant investment in time and resources.
• The Human Element: Ultimately, marketing is about connecting with people. AI agents can help you do that more effectively, but they can’t replace the human touch.

The bottom line: approach AI agents with a healthy dose of skepticism

AI agents hold immense potential for retail marketers, but it’s important to be realistic about their capabilities and limitations. Don’t believe the hype, and don’t expect AI to solve all your problems overnight.

Instead, focus on understanding how AI agents can augment your team’s strengths, enhance your workflows, and empower you to create more personalized and effective marketing experiences. And always remember AI is a tool, not a replacement for human creativity, intuition, and strategic thinking.

Since you’ve made it thus far, here is something to ponder…what types of business models will Agentic systems open up?  A lot of folks think it’s going to be Service as Software, or charging for outcomes vs tools…what do you think?

Next up, we’ll tackle video and 3d asset creation..which is heating up quickly!