How We Use AI Agents at Walr

By Walr

• article

• DIY Surveys
• Survey Research
• Data Processing
• Data Visualisation
• AI
• Artificial Intelligence
• AI Agents

People tend to interpret the word “agentic” in different ways. For some, it’s a technology story. For others, it’s a threat to jobs. For us at Walr, it has been something more straightforward: a way to do the work we were already doing, faster and more accurately, while keeping human judgment at the centre of the process.

I want to explain how we have actually implemented agents in our operations, because much of the conversation around this topic still feels quite abstract. The reality, at least in our case, is more practical and grounded than the headlines suggest.

Learn more by watching or listening to Lewis on the Founders and Leaders Series podcast here:

What Walr Does

Walr provides an end-to-end solution for enterprise online quantitative data collection. We build surveys, we access audiences in over a hundred markets, and we structure data. Our customers are market research agencies and research teams that want to get to their data accurately and quickly. A number of them do not particularly want to learn another platform. They just want their data done well.

That context matters because it shapes how we have thought about agents. The goal was never to automate for its own sake. It was to take away friction so that our customers can spend their time on the work that actually requires their expertise.

The Survey Programming Pipeline

The most developed use of agents in our business is in survey programming. Historically, this was a multi-step human process where a researcher would send a Word document with their questionnaire, and a scripter would work through it to produce a functioning online survey. It took time, involved a lot of repetitive interpretation, and, like any repetitive human task, it wasn’t always done perfectly.

We now run that process through three layers: two agentic, one human. The first agent takes the questionnaire and converts it into a script for our survey-building platform. The second is a verification agent that checks the first agent’s work. Its job is to ask: how confident am I in what has been produced? Where has the first agent done something that warrants a human look?

The output of the second agent isn’t a finished survey, but rather a prioritised list for the human reviewer. Instead of going through all 40 questions before finding the three that need attention, the human can go directly to the parts of the survey that require their judgment because the rest have already been verified as accurate.

What this creates is not less human oversight, but a more efficient way of applying it. The human is still closing the loop, and the trust within the process is the same. We’ve just removed the parts of the job that didn’t require a person.

In terms of measurable impact, the number of projects managed per person has roughly doubled in the last six months. That is not a projection, but what we are already seeing in practice.

Why Accuracy Holds Up

The question I get asked most often about this approach is: What about accuracy? People have heard about agents hallucinating or making decisions they shouldn’t have made, which are reasonable concerns.

The key for us has been not asking our agents to think freely. Instead, we are training them on our own proprietary platform, using data from over 1.5 billion questions asked through Walr since we launched. When an agent encounters a question type, it has almost certainly seen it before. As such, we are relying more on memory recognition than on creative problem-solving.

We also re-educate agents when something goes wrong. If an agent produces an incorrect output, that becomes training data, and the system improves over time.

The other thing I would say to anyone nervous about agents is that the task we are automating is one where right and wrong are clearly defined. If I have a Word document with questions and need an online survey that represents them exactly in the structure the document intended, that is much more straightforward to validate than, say, deciding what insight is most relevant for a particular stakeholder. We operate in a part of the market research flow where the definition of success is unambiguous, which makes agents well-suited to the work.

Who is Building the Agents

One thing that has honestly surprised me is how agent building has evolved within our organisation. We have three core areas where we formally use agents, but we now have well over 100 agents in use across the business for various purposes. Fewer than half of them were built by our product and engineering team.

The rest were built by people in our operations, sales, and other teams who asked themselves: Can I automate this task to make myself more scalable? They didn’t need a software engineering background to do it, just an understanding of their own workflow and the curiosity to try.

Our sales team runs agentic workflows that govern how work gets booked in and pushed through our systems. The innovation team doesn’t hand down these tools; the sales team built them themselves to solve problems they understood better than anyone else.

That technology development is no longer confined to engineering teams has been one of the more significant shifts for me. It’s genuinely new and changes how research businesses need to think about the role of technology in their organisations.

What Governance Looks Like

None of this happens without a certain amount of oversight. When people across the organisation are building their own tools, you need to make sure everything behaves correctly and that data is properly looked after. That governance layer is something we have had to build alongside the agent capability itself.

But the requirement doesn’t undermine the broader point that governance is a normal part of running a technology-enabled business. The fact that non-technical teams can now contribute meaningfully to how the organisation operates, through tools they have built themselves, is worth the governance overhead.

The Human Element

I want to be clear that none of what I have described is about removing people from the process. At Walr, we firmly believe in the value of human involvement, and closing the loop with a human is critical to delivering the right product to our customers.

What agents give us is not a replacement for human judgment, but rather a more effective way to apply it. In our survey pipeline, the human reviewer is still essential, but they can now focus on the work that genuinely needs them. The same is true across every part of our business where agents have been deployed.

Speed and accuracy are not in tension when agents are set up correctly for the right tasks. The tension people experience tends to come from deploying agents on tasks where the definition of success is unclear or where the agent has not been properly trained. Get those two things right, and the benefits flow through consistently.

Learn more by watching or listening to Lewis on the Founders and Leaders Series podcast here:

Author

Learn more about