The Real Impact of Agentic AI [Case Study]

Last week, Salesforce founder and CEO Marc Benioff said he wants to deploy 1 billion AI agents within a year. 1 billion. “What if your workforce had no limits?”

Benioff also said there are already “hundreds of customers” using Salesforce’s AI Agent platform, Agentforce, for discrete tasks that take work off the plate of human employees.

I sat down with the Salesforce AI leadership team this week in San Francisco and received a compelling briefing about how they are using AI agents to drive efficiency and the early case studies they’re seeing from their customers.

AI agents are no longer a promised future — they are here. Not only are enterprise tech companies rolling out agents, but so are hundreds of startups.

I want to share the story of one startup, Blitzy.AI, founded by two of my former HBS Students.

But first, I hope you’ll join me for my upcoming Q & A about The Experimentation Machine!

AMA! Join me to discuss the first three chapters of my new book, "The Experimentation Machine: Finding Product-Market Fit in the Age of AI."

Ask me anything you want. Popular topics include:

1. Using AI to accelerate your startup
2. Becoming a 10x Founder
3. Experimentation Mindset

This event is for pre-order customers only.

 Pre-Order your book to read the first three chapters and receive a link to the virtual event.

Now, onto the fantastic future of Agentic AI:

Blitzy AI: Building Your Agentic Software Team

Studies have found that the early AI copilots make engineers about 40 percent more productive. AI agents have the power to make us 100 times more productive. 

Agentic AI represents the shift from copilot to full decision-making and action-taking. We will have general use AI agents that can operate across multiple digital surfaces for our day-to-day lives (e.g., “create a workout plan for the week tailored to my fitness goals and then create a daily 30-minute podcast from a personal trainer to talk me through the program and text me a link to that podcast each morning at 7am.”) and also specialized agents designed to complete specific tasks, series of tasks, or to work on teams with other AI agents. 

Agent teams have the greatest potential impact on the way startups scale. Imagine a team of specialized agents working together to move a project from ideation, through development, through testing, and into production. Hard to imagine? This is exactly what startup Blitzy.AI is doing today to build enterprise software in one-tenth of the time and cost of traditional software development firms. 

“Software development today is AI-supported, where you have engineers using copilots,” said Siddhant Pardeshi, co-founder and CTO of Blitzy. “But what if you could flip that and have AI do most of the work, and only have the humans fill in and do the rest?”

Blitzy’s vision is a world where all B2B software is custom-built to the exact specifications of the customer, and this world is not as far off as you might think. Today, 80 percent of all commercial software is built on open-source technology. That means AI has seen and ingested it all as training data. AI models have already learned how to build good software, they just need the agency to do the job. 

That’s what Blitzy is trying to do. The startup recently won a contract for an enterprise software project that required 30,000 lines of code which would normally take six months to build. They completed the project in just six days with one senior engineer (their next goal is to complete projects like this with a junior engineer). 

Their process is cutting-edge, but it’s a harbinger of things soon to be available to the masses. The key is to think about breaking down complex tasks into smaller, specialized steps that can be handled by different agents while maintaining human oversight for quality control and final refinement: 

Step 1: A specialized agent takes the initial product vision and concept and translates it into a product requirements document (PRD). This is the roadmap for all subsequent development.

Step 2: A second agent ingests the PRD and expands them into detailed technical requirements.

Step 3: A third agent converts the technical requirements into technical specifications. This is the step that senior engineers would traditionally take on. The specs are then reviewed and finalized by human engineers. 

Step 4: A fourth agent translates specifications into system architecture. It maps out how different components of the software will work together and handles complex system dependencies and integrations.

Step 5: Finally, thousands of coding agents (up to 3,000) work collaboratively to build the software. Each agent handles a specific portion of the codebase. Together they can generate 30,000–50,000 lines of code. This is how Blitzy overcomes the token limit of individual AI models (which as of this writing is 8,000 tokens or approximately 6,000 words for ChatGPT). 

The entire development process currently takes about twelve hours. Ten of those hours are spent on “thinking” through the first four steps. Only the final two hours are spent actually coding. Right now, Blitzy’s promises to create about 80 percent of a final project, leaving the final 20 percent for human engineers. 

Blitzy is not alone; Replit and Codeium are startups with similar platforms growing in popularity. We are still in the early innings of the Agentic Revolution.

2025 will be the year of AI Agents. As always, my advice to you is this: Start embracing these tools today so that you can be proficient in them when they become fully viable and mainstream.

I would love to hear from you:

• What are you seeing on the Agentic AI front?

• Any real-world experiences deploying agents in your life or work?

• What companies are exciting you most in this space?

All the best,

Jeff