Witness why James Ball benched automations for now and started talking to agents to solve complex work management logic and improve efficiency.
If you’d told me a few years ago that I’d soon be having a conversation with my work management platform — genuinely reasoning with it, debugging alongside it, teaching it where it is and what to do — I’d have probably raised an eyebrow. But here we are, and honestly, I think we’re just scratching the surface.
Jellyfish is a global digital marketing agency, sitting on the edge of AI. We look after some of the world’s most loved brands, supporting them via earned media, paid media, creative, analytics, strategy, the lot. Our inventory, if you like, is our time and expertise. So anything that helps us understand how efficiently we’re using that time, how effectively we’re delivering for our clients, how big our thinking is, and where we might need to improve is enormously valuable to us.
We’ve been using an intelligent work management platform, Wrike, for a few years now to manage delivery, collaborate with clients, and measure the output of our teams. And when AI agents arrived on the platform, I’ll be honest, I wasn’t entirely sure where to start. But I’d been wrestling with a problem that felt like it needed something more than traditional automation rules — those preset, if-this-then-that workflows — could offer…and that turned out to be the perfect entry point.
Starting with a real problem, not a theoretical one
Our growth marketing team needed a way to calculate delivery success rates. Sounds straightforward enough, but the logic behind it is actually quite involved. You’re looking at start dates, target due dates, and actual delivery dates, then working out the delta between them to generate a percentage score. Historically, someone would have had to build that calculation offline, probably in a spreadsheet, and maintain it manually. It’s the sort of thing that works for a while and then quietly falls by the wayside.
Now, with new advancements in AI, I was able to build an agent that handles all of this automatically. When a task transitions to a completed status, the agent examines the delivery date against the due date and applies a success rating. I’ve then got averages rolling up at the folder level, so I can see at a glance where we’re delivering efficiently and where we’re not.
Simple, effective, quite complex logic behind it, but this for me is something I’ve never seen before in any platform or system. And the fact that I could have a working prompt and iterate from there to get it tested and running within twenty-four hours? That’s the sort of speed to value that actually keeps teams moving.
The reasoning is the thing
What really shifted my thinking, though, was understanding how agents differ from automations at a fundamental level. With an automation, it either works or it doesn’t. And when it doesn’t, you’re often left guessing why. Agents give you visibility into their reasoning. They tell you what they tried and why they failed, which allows you to course-correct in a way that feels genuinely collaborative.
I had a brilliant example of this just recently. I’d built an agent to categorise tasks based on the folder they live in, assigning goals, KPI types, and other custom fields contextually, depending on location. But it wasn’t working, and I couldn’t figure out why. I stripped the prompt back to a single sentence, and started from scratch. The first thing I did was simply tell the agent: figure out where you are. Once it understood its context, everything else clicked into place.
That was a real learning moment for me. With automations, you build everything in situ with the end goal considered. With agents, you can let them assess context and behave dynamically. If a task moves from one folder to another, the agent recategorises it automatically. You’d need a significant number of automation rules to achieve something similar, and they’d need constant maintenance.
Reducing the overhead, not replacing the foundations
I want to be clear. This isn’t about ripping up everything you’ve already built. Automations are still doing essential work across our setup. They’re reliable, fast, and for straightforward trigger-and-action tasks. I’d say to anyone considering agents: think of them as complementary to your existing setup. The groundwork you’ve done with automations, request forms, and workflows still matters. But where you’d normally reach for yet another automation rule, just pause and consider whether an agent might handle it more elegantly.
One area where I’ve found this particularly powerful is request forms. At Jellyfish, we’ve had forms that are genuinely twenty pages long with conditional logic branching everywhere. I’ve seen it across the industry. What happens when a client opens a form and sees “page one of twelve?” They check out. An agent lets us replace all of that with a single conversational input. The client describes what they need in natural language, and the agent handles the categorisation, the routing, the assignment. Still structured output, but from a much more human input. It’s not perfect yet, but the reduction in friction for the client is already significant. That’s really quite important.
Where I’m headed next
I’ll be the first to admit that I haven’t wrapped my head around everything yet. The analysis piece, such as having agents proactively surface insights, flag bottlenecks, and do my weekly checks for me, that’s where I want to get to. I’ve also been thinking about whether an agent could help me identify technical debt: automations that have broken because someone deleted a custom field or request forms that no longer trigger properly. The sort of housekeeping that’s important but never urgent enough to prioritise.
Agent chaining, where multiple agents pass context and outputs between each other to complete more complex, multi-step workflows, is going to make all of this even more powerful. I’ve currently got thirteen agents running across one project space, and I suspect that number will consolidate significantly as chaining matures.
My advice? Don’t be afraid of the baby steps. Start with something you probably could build in an automation, but resist the urge to fall back on what you know. Experiment with natural language. Test, iterate, ask the agent why it did what it did. Those small steps are genuinely how you’ll progress towards something far more sophisticated than automation alone could ever deliver.
We’re just getting started, and I’m really excited about where this is going.
Explore AITechPark for the latest Artificial Intelligence News advancements in AI, IOT, Cybersecurity, AITech News, and insightful updates from industry experts!