Aller au contenu principal

Contactez-nous
View from horseback of a rider’s hand on the saddle and the horse’s windblown mane against an open grassy plain under a blue sky.

AI agents as workhorses: taking time-consuming tasks off your hands

Nov 19, 2026

Some tasks in your organisation have always been done the same way. Nobody questions them, so nobody improves them. They're hardly glamorous, but they take up valuable time that your people could be spending on more important work.

And these are exactly the kinds of tasks you can already reliably hand over to an AI agent. Think of it as a workhorse: it gets on with the job, completes the task and delivers the results, without you having to oversee every step. Two real-world cases show what this looks like in practice and what it takes to get it right.

Case 1: AI-agents in content migration

Content migrations tend to follow the same steps: taking inventory, analysing and cleaning up content, designing and mapping, agreeing on requirements, enriching content and migrating it. Most of the time goes into the first two stages. And that's precisely the kind of work AI agents are well suited to. In practice, this means running crawl analyses, checking which published URLs appear in analytics data and investigating what's stored in the CMS database. Essential work, but repetitive. We're now handing these tasks over to AI workhorses, fundamentally changing the role of the analyst. 

The real benefit goes beyond saving time. It's about the depth of analysis. Previously, analysts often had to stick to the bigger picture simply because there wasn't enough time to dig deeper. An AI workhorse can analyse content at individual page level. Even two business analysts working together would struggle to achieve the level of detail AI models can now provide. That frees up time for what really matters: working with the client to decide which content is worth migrating. 

Close-up of a cluttered computer desktop filled with overlapping folders and spreadsheet file icons.

Start the conversation with suppliers

Deeper analysis also opens up access to information that was previously difficult to obtain. CMS vendors can be reluctant to share data model structures or source code. But when you have a clear business case and can demonstrate how that information will be analysed securely, they're much more willing to have the conversation. 

The key is having a platform with the right guardrails: rules that determine what AI can and cannot do with your data. We use iO Bonzai, our unified AI platform for integrations, governance and AI orchestration. Bonzai acts as a harness for different large language models (LLMs), depending on the task, and is compliant with the EU AI Act and ISO 27001 certified. This helps remove a major barrier for vendors: the security questions have already been addressed before they even come up. 

Close-up of a blue human eye seen through a round magnifying lens, showing detailed iris patterns and eyelashes.

Human-in-the-loop: keeping people in control of AI agents

Even with the right guardrails in place, human oversight remains essential. At iO, that's a deliberate choice. Take content mapping, for example. Once you've provided the components and content, it's tempting to let AI handle the mapping entirely. But that doesn't always go to plan. Give an LLM free rein over a large volume of content and it may fill in any gaps with information it thinks should be there. 

The solution is straightforward: assign a unique ID to every piece of content. Let AI map only those IDs, then restore the original text afterwards. That way, you get exactly the content you intended, without anything the AI has made up. Other agents also review the results before a person checks whether the page actually works for its intended audience. 

And human involvement goes well beyond reviewing outputs. Writing prompts, installing agents and configuring how they work still require people. That's exactly how we want it. Build human oversight into the process from the start, and you won't have to fight to regain control later. 

Case 2: AI agents for architecture documentation and threat modelling

The same principle applies to software security: agents take on time-consuming work while people remain in control. But here, the stakes are different. 

Threat modelling is an exercise in looking at a system through an attacker's eyes, before a real attacker gets the chance. You examine the architecture, the data the system processes and the trust boundaries between its components: the points where one part of the system cannot automatically trust another. You then explore the different ways an attacker might compromise the system. 

It's a labour-intensive process that brings together experts from development, the business and systems engineering. And it's becoming increasingly urgent. Attackers are using AI too. To stay ahead, organisations need to understand how resilient their own systems really are. 

Documentation: the missing foundation

To do that, you need architecture documentation. A clear description of how the system works, including its components, integrations, data flows and the decisions behind its design. Unfortunately, that documentation is often missing. Poor documentation has been a longstanding problem in IT. There might be something available at go-live, but it rarely gets updated afterwards. New functionality tends to take priority over basic maintenance and housekeeping. Without that context, meaningful threat modelling becomes almost impossible. Garbage in, garbage out. 

For one client whose architecture documentation hadn't been updated for years, we built a team of AI workhorses to reconstruct it retrospectively. A planning agent directed an orchestration agent, which deployed other agents to analyse the source code, commit history and Jira tickets. Those tickets revealed the actual business requirements behind each change. The result? 400 pages of detailed documentation, complete with diagrams and in-depth technical information. And an unexpected bonus: 11 security issues, ranging from low to high risk, uncovered in an existing system. A system originally built by humans. 

There was still one gap, though. The documentation didn't cover the application's functionality from a user's perspective. So we introduced a functional analysis agent that used a demo account to navigate the application in an isolated test environment. An agent reconstructs what's there. A person spots what's missing.  

Colorful sticky notes with handwritten reminders overlap on a wall, creating a cluttered, organized planning space.

A turning point for security

An initial threat-modelling exercise for an existing system can easily cost more than €10,000. That might sound expensive. But for systems where security is critical, it's simply part of good security practice. With AI workhorses taking care of the documentation beforehand, the barrier to getting started is now considerably lower. 

And we're approaching a point where AI agents may outperform human developers in certain areas of security. That's because agents can consistently identify patterns that people might overlook. 

That's why we deliberately build peer review into our agent workflows. Different models analyse the code, another model reviews their findings, and a security expert validates the final results. We never rely blindly on a single model. People remain in control, with human oversight built into every stage. AI allows us to multiply what people can achieve. 

Empty theater stage with a single bright spotlight shining through haze onto the wooden floor, viewed from the dark audience seats.

Which part of your process could benefit from an AI workhorse?

The best candidates are rarely the most visible tasks. They're the ones that simply need doing, day after day, without anyone questioning whether there's a better way. Handing that work over to an AI agent frees up your people to focus on where they add the most value. 

Want to find out which tasks in your organisation could be handled reliably by an AI workhorse? Let's talk.