Originally published on IT Assure.
An overdue website refresh changed how I think about AI—from a tool that gives me instructions to something I can delegate work to, with direction and review.
Over one weekend, I used ChatGPT with connected tools to carry out most of a refresh of IT Assure’s website. The project had been moving slowly because I was trying to fit the work around running the business.
What surprised me was not the quality of another set of recommendations. I had been getting useful recommendations from AI for some time.
This time, it could carry out the changes.
My role shifted from following instructions inside WordPress to defining the outcome, reviewing the plan, approving work, and checking the results. It felt much closer to working with a web-development team than asking a chatbot for help.
But I also discovered something less obvious: once AI could execute, deciding which work was worth doing became more important.
I had already been automating work for years
I run a small IT services business with about ten people. For years, I have looked for ways to reduce repetitive administrative work—the tasks that are tedious, easy to overlook, and rarely anyone’s favorite responsibility.
One challenge has been that our work spans multiple applications. Information might be in a meeting transcript, a document library, someone’s email, or our customer relationship management system. Bringing it together often required a person to move it between systems.
No-code automation platforms such as Zapier helped us connect those steps.
For example, we use an automation to organize meeting information. Transcripts and summaries are stored in SharePoint, participants are matched with CRM records, and relevant summaries are added to the client’s record. Action items become tasks. AI helps distinguish a client meeting from a prospect, internal, or vendor conversation.
The purpose is practical: information from a client conversation should be available to the people who need it, rather than remain in one person’s inbox or memory.
That workflow already combined AI and automated execution. What I wanted to explore next was different: could AI help carry out a project where I had not already configured every step?
The website was a good test because I was the bottleneck
We use WordPress and host the site ourselves. A development agency had worked on the original site. Other providers and our own team had made changes over time.
The result was inconsistent. Some links went to the wrong places. Navigation was confusing. Colors and page layouts varied. Mobile presentation worked better on some pages than others. Calls to action did not follow a consistent visitor journey.
I work in technology, but I am not a web developer. Fixing those issues was taking more time than I could comfortably give it.
My usual approach was to find a problem, describe it to AI or upload a screenshot, and ask how to fix it. AI would explain the steps. I would make the changes and test them. When something did not work, I would return with more information and try again.
The advice helped, but I was still doing the implementation.
I estimated about twenty important pages needed attention. At roughly two hours per page, that was around forty hours of my time. With perhaps four hours available in a typical week, I was looking at roughly ten weeks of work.
That was my estimate based on how slowly I was getting through the pages—not a professional developer’s estimate.
The change was who carried out the next step
I decided to use the refresh to test AI-assisted execution.
Getting the necessary connections and plugins ready took me about thirty minutes. I also had an advantage: I had already prepared documents explaining our positioning, services, website direction, and design requirements. The AI had something concrete to reference.
I asked it to audit the website against those materials and propose a refresh plan.
At first, the experience was familiar. It returned findings and recommendations.
Then it asked whether I approved the plan and wanted it to proceed.
After I approved the work and the necessary website access, it began making changes. I could see the results in the browser. When it encountered something unexpected, it came back with an explanation and asked for direction or approval.
The difference was not that tools and integrations had disappeared. Those connections were what made execution possible. The difference was that I no longer had to translate every recommendation into a sequence of actions inside WordPress.
With my earlier automation workflows, I had configured the path in advance. Here, I was reviewing a proposed path and letting AI perform the approved work, including responding to issues encountered along the way.
Usage limits interrupted the work. Scope creep complicated it.
The first interruption was usage.
At one point, the work stopped. I learned that I had reached the usage allowance for my plan. I moved to a higher tier and purchased additional credits so I could continue.
I was still learning how usage was calculated. What became clear was that a substantial execution task could consume enough allowance to interrupt the project.
The second problem was less technical: the AI kept finding more work to propose.
Some findings were important. Others needed to be weighed against the purpose of this particular refresh. A legitimate issue did not automatically mean I should expand the current project to address it immediately.
The interaction started to become a cycle: recommendation, approval, work, another issue, another recommendation.
I realized that I was approving individual tasks without always stepping back to ask whether the overall plan was still the right one.
The ability to execute did not resolve the question of priorities. It made that question harder to ignore.
A clearer brief made the process work better
I paused and wrote a more complete set of instructions.
The main objective was to make the website accurately represent what we do and give visitors a consistent, understandable experience. I wanted that objective to remain visible as the work progressed.
I also set boundaries around the design and implementation. The main marketing pages should follow our established dark theme. I wanted to avoid unnecessary changes to CSS code. Proposed work needed to respect the requirements already agreed.
Four principles became especially useful:
Keep the goal attached to the plan. I asked the AI to retain the objectives and requirements with each revised plan and check proposed actions against them. The latest issue should not quietly become the new purpose of the project.
Eliminate first, then simplify. Before adding work, consider whether it directly contributes to the intended outcome. When work is necessary, look for the simplest way to complete it without creating more maintenance.
Break execution into reviewable parts. I asked for batches intended to take no more than thirty minutes. That gave me opportunities to review the results and monitor usage before authorizing more work. The batches often finished in about fifteen to twenty minutes, but the time boundary was a planning instruction, not a guaranteed runtime.
Review before continuing. The working pattern became straightforward: agree on the next part, let AI execute it, inspect the result, correct anything necessary, and approve the next part.
The process became smoother after that. I still had to make decisions, but I was making them within an agreed plan rather than responding to an expanding stream of suggestions.
The project fitted around my weekend
I started late on Friday. Over the weekend, I checked the work between ordinary family activities.
I reviewed a section before lunch. I approved another before taking my children to the park. Later, I checked progress before dinner and again before bed.
I moved between the desktop app, browser, and mobile app as needed.
That flexibility was a meaningful part of the experience. I did not have to spend the whole weekend sitting in WordPress, carrying out each change myself. I could review a completed part, give direction, and return when the next part was ready.
It was not unattended work. I was still involved in content, decisions, corrections, and approvals.
But the balance had changed. I was spending more of my time reviewing work and less of it figuring out how to perform each step.
By the end of Sunday, most of the planned refresh was complete.
What the result cost—and what the numbers mean
My rough estimate is that AI performed about 80% of the refresh tasks, with me handling the remaining work around content, direction, requirements, review, and some manual changes.
That is an estimate of the task split, not a measured share of total effort. It also does not mean AI built 80% of the website from scratch. We were refreshing an existing site.
I spent approximately five hours interacting with the system, learning how to work with it, reviewing results, and doing the manual work that remained. That compares with the roughly forty hours I had estimated for doing the cleanup myself using my earlier approach.
There is an important qualification: I already had strategy and reference documents available. The five hours should not be read as the total effort required to develop a website’s positioning and content from nothing.
I also spent about $200 on subscriptions and additional credits during the experiment. That was money paid, not a separately measured figure for the credits consumed by this website project alone.
I had not obtained an agency quote for the same refresh, so I would not claim a precise saving against hiring a developer. My previous agency experience involved different work and is not a like-for-like comparison.
What I can say is that a project I expected to work through over many weeks moved substantially forward over a weekend, with much less hands-on implementation from me.
The result still required checking and correction. When I found a mistake, I pointed it out and worked through the fix.
What changed was more than speed
What interests me about major inventions is that they can change our relationship with a task. The benefit is not always doing the same thing faster. Sometimes, something that previously felt out of reach becomes practical.
That is how I think about the progression from personal computers to the internet and now AI.
For a business owner like me, the internet made technical instructions, documentation, and other people’s experience accessible. I could look up how to maintain a website without first training as a web developer.
But finding information and applying it were still different things. I had to determine which instructions matched my situation, understand them well enough to act, and work through the problems myself.
Conversational AI helped close some of that gap. I could describe the problem, ask follow-up questions, and get explanations tailored to what I was trying to do.
In this project, AI went further. With the necessary tools, access, and approvals, it carried out many of the changes.
My progression was from finding instructions, to getting help applying them, to delegating parts of the work.
I did not become a web developer over the weekend. I gained a way to complete work that would otherwise have required considerably more of my time or outside help. The change was in what I could accomplish with the support available to me.
The website project also showed me that access to capable AI was only part of the answer. The work still needed context, priorities, permissions, and a clear definition of an acceptable result.
That is where much of my contribution moved.
I spent less time figuring out which buttons to press and more time deciding what the website needed to accomplish, which changes mattered, and whether the result was good enough.
For another business owner, I think that suggests a useful question:
What worthwhile work have we been postponing because the time or specialist effort required made it difficult to complete?
That is where I would look for the next bounded experiment. Define the outcome, supply the relevant context, agree on the boundaries, and review a manageable piece of completed work. Include the time spent directing and correcting AI when judging whether the approach actually helped.
This project does not tell me that every workflow can run reliably without supervision. It gives me a reason to test more—and a better way to conduct those tests.
AI expanded what I could accomplish. I still had to define what good work meant.
I will continue testing that boundary and sharing what I learn from the work itself.
To discuss this experience, contact Mark at marks@itassure.com.
