Dr. Ryan Ries here. I recently hosted a webinar on agentic AI, and the questions kept rolling in after we wrapped. Too many good ones to let them sit in my inbox. So this week, I'm turning the Matrix into a Q&A and tackling the questions we didn't get to.
Before I dive in, one quick announcement. The AWS LA & NYC Summits are coming up. If you’re attending, you’ll want to register for our events those weeks:
- June 10th: LA Happy Hour at Katsuya
- June 17th: AWS Summit NYC Happy Hour at Le Jardin Sur Madison
Alright, let the Q&A begin!
"What's the best architecture for setting up an agentic workflow to automate business processes?"
I think the first thing you have to do is actually map the workflow you're trying to automate. I mean every step, every gate, every decision point. People assume they know their processes. They don't. There are always hidden steps, tribal knowledge, and undocumented handoffs.
Once you have a complete picture, you can define the workflow and the tooling required to execute it. A lot of times when we look at workflows, most of the work isn't the agent itself. A huge portion of any agentic implementation is data preparation, cleaning, and running ETL. The LLM can help with some of that, getting data ready for downstream processes like reporting, visualization, or feeding into other systems. But the agent is only as good as the data pipeline underneath it.
"Do you need to read every line of code AI generates?"
With the speed at which AI writes code now, reading every line isn't realistic. The game has shifted from code review to testing. You have to invest more time upfront in design and architecture, then let the system generate substantial pieces of code, and validate through rigorous testing.
These tools are still fairly naive. They don't always pick the right services. They over-engineer. They'll build a production-grade system when all you wanted was a proof of concept to validate a hypothesis. You need someone in the loop who understands what should be built, which repositories to use, and how to verify the output matches the actual requirements.
"What should young students focus on to get hired in this market?"
Stop thinking about writing code line by line. Start thinking about building products. Understand design structures, system architecture, and how components fit together. The junior role is evolving fast. The people who will get hired are the ones who can manage a team of agents, not just write functions. You need to know when the AI chose the wrong approach, when it's overbuilding, and when it's cutting corners. That judgment is what companies are paying for now.
"What are some use cases where agentic AI delivers real business value?"
We find the best starting point is companies with well-documented workflows. If you can review those workflows and identify steps that involve heavy document review, repetitive analysis, or structured decision-making, you've likely found a candidate for automation.
The key qualifier is frequency. If you're performing that workflow constantly, the ROI math works. If it's something you do twice a quarter, the investment probably won't pay off. Look for the processes where people spend hours every week doing the same type of cognitive work.
"What's the benchmark for ROI on a typical agentic AI project right now?"
For document-heavy review workflows, we've seen teams go from four hours down to thirty minutes (that’s a real number from a real project btw!). Other use cases are less dramatic per interaction but compound fast. A chatbot answering the same question hundreds of times saves minutes per query, but the aggregate time savings for that individual or team is massive.
I wouldn't walk into any project expecting 10x returns. Go in looking for use cases that reduce manual effort, decrease error rates, and free up people for higher-value work. The ROI builds from there.
"Our team is burning through token budgets. How are large companies handling token exhaustion?"
This is a brand new problem and most organizations are still figuring it out. The mistake I see most often is teams measuring ROI by token spend, which is the wrong metric entirely. But the bigger issue is how people are using these systems. If you let an agent run without constraints, just throwing tasks at it with vague instructions, you'll burn through tokens and get very little value in return.
You need to be strategic. Think about the information you're providing, be precise about the output you expect, and design your prompts with cost efficiency in mind. Treat token budgets like compute budgets: optimize, scope, and don't just let the meter run.
"How do you use agentic AI with sensitive data like confidential information, intellectual property, or client privilege?"
Treat your agentic system like an employee with a security clearance. Limit what it can see and what it can do. Most of these systems can be deployed behind your firewalls and within your existing security perimeter, meaning the data never leaves your controlled environment.
Where people get into trouble is giving agents too much autonomy to solve problems creatively. If you've given your agent broad permissions to access systems and create solutions without guardrails, you've handed it more power than you'd give a new hire on day one. Set up proper access controls, scope permissions tightly, and assume the agent will try to use every tool you give it.
Remember when we did a deep dive on Clawdbot? If you are trying to use some systems like Clawdbot that are purposely trying to build things and get past your protocols, then you are asking for problems. You have given the Clawdbot too much power to create its solution.
"What's a practical, low-risk way to get started with agentic AI?"
Same answer as the architecture question: understand your workflows first. Then design a system to automate a specific, well-scoped workflow with proper risk management, governance, and security baked in from day one.
If you're on AWS, building your agents on the AWS Bedrock AgentCore framework is a strong starting point. It lets AWS handle the infrastructure heavy lifting while you focus on configuring the system for your specific use case. Start small, prove value, then scale.
People want to jump straight to the agent. They want the AI doing the things!
But you can’t skip the boring work first: mapping workflows, cleaning data, designing architectures, setting boundaries. The agent is the last mile, NOT the first step.
If you're thinking about getting started with agentic AI or you're already in the thick of it and burning through tokens with nothing to show for it, reach out.
Feel free to follow this link.
Until next time,
Ryan
Now time for this week’s AI-generated image and the prompt I used to create it.
Generate a hyper realistic image of me standing at the front of a classroom of ai agents. I am answering questions about ai agents. The questions are coming quickly and I am trying my best to keep up. Some of the agents have their hands raised. Behind me should be a chalkboard. I should have a frantic look on my face and you should see a bead of sweat on my forehead. Attached is my photo for your reference.
.png)
