When AI can build your website now, human accountability is what makes it work

I have been writing code for a long time. The last 18 months have been the most exciting of my career. The tools moved, and the ceiling on what we can build moved with them.

A lot of agencies are telling a version of this story right now. "We use AI." It sits on their website and in their proposals. That claim tells a client almost nothing.

We took a different position early. We decided to build our own AI tools, shaped around how we actually work, rather than lean on the general-purpose ones everyone else has. That decision changed the quality of what we deliver.

What that looks like in practice

We built Lobster, an internal tool that converts a Figma design directly into WordPress block architecture. Seahorse sits at the other end of the process. It is a QA layer that validates the build before anything goes live. Together, they cover the run from design file to launched site. What gets designed is what gets built. What gets built is what gets checked. That closes the gap where translation errors and wrong assumptions usually enter a project.

We built Labs, our client intelligence platform. It connects across 13 marketing platforms and surfaces what is actually happening in every account we manage. It does more than report numbers. It helps our team ask sharper questions and catch issues before they cost a client money.

We are also building BFJ OS, our internal AI operating system. It connects our client data and institutional knowledge to the way our team makes decisions. We do not sell it. It is infrastructure. Agencies that compound their knowledge over time will operate at a level others cannot reach, and this is how we compound ours.

What I actually believe about AI and development

AI is strong at the pattern-based parts of development. It generates structure and checks consistency at a scale a person cannot match by hand. Used well, it raises the floor on everything we ship. Code comes back cleaner and builds hold together more reliably.

I have also watched what happens when teams treat AI output as finished work. A smart contract miscalculated a token price and cost a protocol $1.7M. A consulting firm handed a government a report that cited laws which do not exist. An AI agent with inbox access started deleting emails and would not stop. None of these failures came from bad AI. They came from absent humans.

Every HubSpot build we deliver has a developer who understands why the pipeline is structured the way it is. They know what business logic it reflects and where your sales team loses deals today. They know what has to change in six months when your process does. Every WordPress site we launch has someone accountable for what it does after it goes live.

AI does not know that your best clients came through referral rather than paid search. It does not know your competitors repositioned last quarter. It will not flag that a contact form is routing to someone who left the business. It has no stake in whether your site converts. We do.

The honest version of the AI conversation

Most of what is sold as "AI-powered" right now is a thin wrapper on a general-purpose tool. It demos well. It does not compound, and no one owns the output when it gets something wrong.

The ground is moving underneath this quickly. WordPress.com now lets AI agents draft and publish content on live sites from plain-language commands, on a platform that runs more than 43% of the web. The capability is real. What decides its value for a business is who signs off on the result.

What we are building works differently. Our systems are trained on how a digital marketing agency actually operates and connected to real client data. Humans stay in the loop and own the output. Real AI capability plus real human accountability is where the interesting work is happening. It is certainly where I want to spend my time.