Do I need an agent?

Agentic AI. It sounds scary, but it really just means intelligent, automated systems.

Here’s why they can be scary: where we once used to automate small parts of a system that required someone to step in and tell a machine to move to the next step and prescribe the tools to be used and the order to use them, we can now give machines fairly simple instructions, and let them decide how to execute the instructions, using whatever tools they identify are necessary.

Give one access to your bank account, with “authorize every purchase without asking,” and tell it to book you a vacation to Iceland without giving instructions on budget, and you might find yourself booked on a luxury vacation you didn’t expect, now having to cancel reservations and change to more affordable accommodations and flights (or maybe you decide to just go with the AI and enjoy the trip).

For market research, agentic AI can mean setting up a system that starts with you explaining a business question that you need to have answered, and the machine decides on quantitative survey vs qualitative study, writes the study, decides who should answer the study, recruits the participants by connecting to panel providers, executes quality checks on the data as the study progresses, then analyzes the data, writes a report, and sends it directly to the client - and all you had to do was tell it the business question.

To some, that might sound amazing. I’m going to guess some brand-side researchers might really like that. Research agencies definitely won’t.

This is where the industry gets its existential crisis from. If a machine could handle every step of the research process without anyone’s interference, then why have researchers at all?

Remember how last week, I presented risk as an aide for deciding how much you should probably be checking the AI’s output? The same applies to creating agents (intelligent automation).

Let’s take that Iceland trip as our example again.

Asking AI to look for 3 different hotel options and to create an itinerary for a 4-day, 3-night trip to Iceland is fine. Low risk. You aren’t asking it to book anything for you, no sensitive information is being handed over.

This is the equivalent of an agent that takes a business question and drafts a research plan for you, or takes an already-written study and drafts an analysis plan for you.

Asking AI to email 2 restaurants that only take reservations that you really want to dine at and ask for reservations for a specific day and within a specific time frame on each of those days is medium risk. It’s an email to a restaurant. It’s not entirely low risk, but you also haven’t given it bank or credit card information for it to make the reservations. You’ve given it some clear guidelines, and you’ll be getting the responses from the restaurants to reply to yourself.

This is the equivalent of having an agent that reviews data as it’s coming in for a study and flags suspicious responses for you to review, or an agent that writes a report based on an analysis plan and sends to you for editing. That’s what staying in the lead looks like.

Asking AI to make flight reservations, hotel reservations, car rental, and anything else requiring your credit card information, especially with the “ask for permission for every connection” turned off - that’s high risk. Let’s throw in that you didn’t give it a budget, either. You’re giving it permission to do whatever it deems necessary to meet the instructions. That flight might have 3 connections and take 28 hours, the hotel might be either super low end or super high end, just depends what was available when it was making the reservations, and you might be okay with the car rental because that one might be something you can walk back from when you arrive.

This is the equivalent of the end-to-end research process being handed over to a system that has zero people checking it along the way.

While it might seem like a great idea, don’t go asking AI where you should create automation, then start asking it to build the needed agents for you.

Why not? It doesn’t know whether your current workflow has human elements that, if removed, would dramatically impact the outcome. It also does care where those human elements exist that help run the workflow. It also doesn’t care if the machine ends up doing all the work, either. And that’s the biggest problem: it doesn’t care - about security, quality, output accuracy, or your customers.

It also doesn’t know whether creating an agent would require business sensitive data being shared. It might not flag whether there would be security issues if you connected different tools together. It might tell you that you need to open 5 new accounts on 5 new services to be able to connect those tools together when all you actually need is a couple of settings to be changed and a line of code

I’ve learned that last one the hard way after following AI’s instructions and creating 3 accounts to create what ultimately was just a spreadsheet.

And don’t think you’re behind if you don’t already have agents built.

Instead, map out your workflows, check where low-risk agents might help and where you absolutely need to stay in the lead role, and start small.

No, I don’t think it’s possible for us to automate ourselves out of a job. I think our jobs change over time, and I think we’re still learning what that means.

Ultimately, it comes back to the question of what are you handing off to AI and what are you keeping? There is actually a fair amount of the work researchers have traditionally done that can be automated, but there is also a fair amount of the work researchers have done that can never be automated.

AI will always give you an answer to your prompt, a draft survey and possibly even a recommended analysis plan to go along with it, and AI can certainly draft nice-looking reports from data.

What it can’t do is tell you whether you were asking the right question in the first place; whether the questions in that study are going to actually get you the data you need; whether you already have answers in data you collect and where your data gaps are.

And it will never know which story in the data is the one that matters the most.

The judgement calls throughout the process? That’s still your job.

Found this helpful? Forward this to someone who is still a little iffy on the whole agent thing.

Check out the Insight Career Network monthly AI for Market Research class for 1 hour of hands-on learning on how to prompt better and turn AI into a thinking partner, including building an agent with AI!

And check out MRXplorer.com for upcoming cohorts starting October 2026 to learn about where to use AI in your market research work. Pilot courses will be 50% off. Space is limited, so reply to this email to claim your spot for either the beginner or intermediate cohorts!