For the past week, I’ve been working with my Grok Bot team to automate some of the most labour-intensive tasks in my e-commerce business. These were processes I had been putting off for quite some time, partly because I knew how much work it would take just to explain them.
To put it into perspective, imagine bringing in a new employee and expecting them to understand your business, learn several platforms, gather the necessary resources and coordinate an entire workflow, all within a day or two. Even for someone experienced, that would be a rather unreasonable expectation.
Yet that was essentially what I was trying to achieve with my agents.
The difficulty wasn’t necessarily in any individual task. It was in getting all the moving parts to work together, in the right sequence, without constantly needing someone to intervene. And I knew that setting everything up, testing it and correcting the inevitable mistakes would demand quite a bit of patience.
I’d been dragging my feet for precisely that reason. But I also knew I couldn’t keep running the business the same way indefinitely. If I wanted to operate with fewer resources and spend less of my own time on routine execution, I would have to tackle these processes sooner or later.
So I gave myself a week to sort them out.
How far did I get?
After plenty of trial and error, and more wasted AI usage than I’d care to admit, I managed to get roughly 80% of what I wanted working.
The first achievement was automating two new product launches, from the initial preparation through to getting everything ready for release. I still spent a fair amount of time reviewing and approving the work, so I wouldn’t call the entire process autonomous just yet. My aim is to reduce that involvement over the next month, ideally leaving myself with just one final decision: whether to press the launch button.
The second was customer service. We developed a workflow covering approximately 80% of the scenarios we regularly encounter, using responses and solutions that have already been reviewed and approved.
Interestingly, this is one area where I don’t particularly want to remove myself from the process.
Customer enquiries and complaints tell me things about the business that I might otherwise miss. They reveal where customers are confused, what isn’t working as expected and which problems might need a more permanent solution. I want to understand those concerns rather than simply have an agent deal with them and close the case.
So, for now, I’m deliberately keeping the final approval with me. It makes me a bottleneck, certainly, but one that serves a purpose.
The third was a manual fulfilment process. This was relatively straightforward compared with the others, but it involved repetitive work that could be handled more efficiently. That workflow is now fully automated.
Taken together, these are meaningful improvements for a small business. But they also made it quite clear where the limitations still lie.
Creativity is proving more difficult
The biggest frustration has been creative work, particularly copywriting.
I understand that Grok Bot is better suited to orchestrating work than producing every piece of content itself. I’ve tried different models and tools, revised the agents’ playbooks and spent quite a lot of time giving feedback.
The results have improved, but there’s still something missing.
The copy often reads more like a factual description than something written to engage a customer. Even after several rounds of revisions, the agents tend to fall back on familiar expressions and structures. Once they find a formula that appears to work, they start applying it to everything.
That might be acceptable for some tasks, but it’s not what I want for a business where product positioning and presentation matter.
For now, I’m considering keeping closer control over the writing and moving more of that work to ChatGPT. I’ve also experimented with connecting the workflows through Notion, although that arrangement still needs refining.
Design has presented a similar problem.
My designer bot is excellent at adapting existing images, but generating something new has been much less successful. I’ve tested a similar workaround and will continue experimenting with other tools. Fortunately, image adaptation already covers about 85% of what I need from the role, so the remaining gap is manageable.
The more persistent difficulty, however, has been coordination.
Getting individual agents to complete their assigned tasks is one thing. Getting them to communicate, hand work over properly, recognise when something has gone wrong and keep the whole process moving is quite another.
This has probably consumed more of my attention than anything else. It’s also one of the reasons we’re developing CrewBook, a human-agent collaboration and project management application. I’m hoping it will eventually address some of the coordination problems I’ve encountered, although there’s still work to be done.
What I would do differently
Looking back, most of the wasted effort came from trying to get the agents to work before I had properly defined how the work should be done.
It sounds obvious, particularly to someone who has spent years managing teams and business operations. But knowing a process yourself is very different from explaining it well enough for someone else, human or agent, to execute independently.
I’ve come away with a few rules that I would now apply from the beginning:
- Document the existing workflow first. Define the sequence of tasks, the tools required and what each step is supposed to produce.
- Give each task one clear outcome. There should be a definite way to establish whether it has been completed successfully, rather than leaving the agent to interpret what ‘done’ means.
- Assign one owner to each task. Everyone involved should know who is responsible for what, without overlapping responsibilities or ambiguous handovers.
- Set clear boundaries. Specify when an agent can proceed independently, when it should stop, when to escalate to its manager and when human approval is required.
- Make someone accountable for the whole process. I don’t want to coordinate every agent individually. There needs to be one responsible manager who keeps the workflow running and comes to me when a decision is needed.
- Turn repeatable processes into reusable skills. Test them properly before putting them into regular operation.
- Verify the output before requesting approval. An agent reporting that a task is complete isn’t sufficient evidence that it has been done correctly. There should be a checking process, ideally handled by another agent or manager, before the result reaches me.
None of these principles is particularly revolutionary. In fact, most would be familiar to anyone who has managed an operational team.
What’s different is the speed at which agents can execute, and how quickly a poorly defined process can produce a great deal of work that still needs correcting.
That was certainly one of my more expensive lessons this week.
You don’t have to get everything right at the beginning
I have quite a few friends and business associates who are interested in using AI agents, whether for their businesses or personal work. Many have been talking about it for some time, but haven’t quite got around to starting.
The reasons vary. Some are reluctant to pay for tools when free alternatives exist. Others find the prospect of documenting their workflows too troublesome. And some have tried an agent once, found the result disappointing and decided the technology isn’t reliable enough.
I’ve also seen people copy someone else’s setup almost entirely, without considering whether it makes sense for their own circumstances. When it doesn’t work, they conclude that the agents aren’t particularly intelligent.
I can understand the frustration. I’ve had plenty of unsuccessful experiments myself, including several this week.
But I think it’s worth distinguishing between an agent that cannot perform a task and a workflow that hasn’t been designed well enough for it to succeed. Sometimes the technology is the limitation. Sometimes the problem is how we’ve organised the work around it.
The only way I’ve been able to make that distinction is by trying things, observing where they fail and making adjustments.
I wouldn’t suggest spending a fortune or automating everything at once. A modest budget, a manageable process and a willingness to invest some time are enough to begin learning what works for your own situation.
After a week of experimenting, I still have plenty to improve. Some tasks are now running with very little intervention, while others continue to demand more of my attention than I’d like.
But the business is operating a little differently from how it did a week ago, and I’ve learnt considerably more about what it would take to make the next set of processes work.
For me, that’s a worthwhile result.