The pitch sounds like serious infrastructure: “we’ll build you a custom AI agent for your business.” It’s presented as a forward-thinking investment - your own bespoke AI, trained on your workflows, doing your work for you.
In practice, it’s often the opposite of forward-thinking. You’re commissioning custom software wired to one vendor’s framework, in a field where the frameworks themselves are still genuinely immature and changing fast. The durable investment isn’t the agent. It’s documenting your business well and writing your instructions as plain-text skills, not code.
What an AI Agent Actually Is
A custom AI agent is software built to handle a multi-step process on its own - calling tools, querying data, coordinating steps, and deciding what happens next - all wired into one vendor’s specific framework for how those decisions get made.
Picture a real version of this: a plumbing business gets pitched an agent to handle lead intake end to end, built in a platform like GoHighLevel’s Agent Studio - a drag-and-drop canvas agencies commonly use to build exactly this kind of client workflow. It reads incoming enquiry emails and form submissions, checks the job details against the calendar for availability, drafts a quote using the business’s pricing rules, sends a follow-up SMS, and logs the whole thing in the CRM. Genuinely useful, and a real time-saver if it works.
The Real Lock-In Problem
The catch is where all of that logic actually lives: the calendar check, the quote-drafting rules, the CRM logging - every step is built inside Agent Studio’s own no-code canvas, in a format that’s proprietary to GoHighLevel and doesn’t export anywhere else. Eighteen months later, the business decides to switch CRMs, or GoHighLevel changes its pricing or direction, and none of it transfers. Not a few tweaks - a rebuild, step by step, starting from the calendar integration back up. Rebuilding that same workflow on a different platform later isn’t a configuration change. It’s a development project, start to finish.
This isn’t a small-business-only concern being overstated. At the enterprise level, vendor lock-in migrations average $315,000 per switch. A small business obviously isn’t facing that literal figure - but the same structural problem applies at a smaller scale: you’re not swapping a subscription when you want out, you’re commissioning a rebuild. And the caution isn’t coming from the sceptics on the sidelines - 81% of enterprise AI leaders already say they’re concerned about dependency on a specific AI vendor, with 29% saying they’re very concerned. Only one in five companies has any real governance plan for the agents they’ve already built. That’s an industry full of sophisticated buyers accumulating exactly this risk without a deliberate strategy to manage it - and a small business getting sold a custom agent is being offered the same bet, usually without anyone mentioning the downside.
None of that lead-intake logic had to be wired into a framework in the first place. The same process could have been written down as plain instructions instead - which is exactly what a skill is.
What a Skill Actually Is
A skill is plain text instructions describing how to do a task - not software. Take that same plumbing business pitched the Agent Studio build a moment ago: instead of paying to have that lead intake logic built into GoHighLevel, the skill version reads something like: “When a new enquiry comes in, check the job details against the calendar for availability, draft a quote using our current rate card, apply the standard 10% rush fee if the job’s needed within 48 hours, and send a follow-up SMS confirming next steps.” That’s genuinely it - no code, no drag-and-drop canvas, no API to configure, just clear instructions a capable AI can follow, pulling the actual numbers and availability from the business’s own documented facts.
It’s worth being precise about where things actually live in that picture. The calendar itself isn’t part of the skill or the AI brain - it stays in whatever scheduling tool the business already uses, like Google Calendar or Outlook. A calendar is live, constantly-changing data, not a static business fact, so it was never going to sit in a folder of documents. That doesn’t reintroduce the lock-in problem, because a standard calendar tool is vendor-agnostic in a way GoHighLevel’s Agent Studio isn’t - virtually any AI assistant can connect to Google Calendar or Outlook through a connector - a bridge built on an open standard called MCP (Model Context Protocol), rather than something proprietary to any one vendor - the same way a phone, a laptop, and a smart TV can all check the same calendar without any of them being locked into it. The skill just tells the AI to check it and act on what it finds. Swap the AI tool later, and the calendar doesn’t need touching at all.
There’s no framework underneath a skill to get locked into - nothing built inside a proprietary canvas the way it would be in Agent Studio, and nothing to rebuild if GoHighLevel changes its pricing or direction next year. Nothing breaks when you decide to switch which AI tool you’re using, either. It’s genuinely as simple as it sounds, and that simplicity is the whole advantage - a well-written skill from today will still work perfectly in a tool that doesn’t exist yet.
So Does This Mean Custom Software Development Is Dead?
No - if anything, it’s the opposite.
Take that plumbing business one step further. Their quoting skill emails a quote, and eventually they want that quote to arrive as a proper branded PDF, not plain text. One of our own early AI-assisted builds is a custom PHP CRM with exactly this: quotes and invoices render server-side, on demand, and get downloaded directly. No AI involved in that step at all. The skill’s instruction is one line - fetch the PDF from this URL with these job details - and the CRM does the actual rendering. That’s effectively zero tokens spent, every time, compared to having an AI freehand a formatted document from a long list of layout instructions.
Here’s the part worth being precise about: that CRM doesn’t care which AI is asking. It never has an opinion on Claude versus Gemini versus whatever comes after them, because it was never wired to any of them in the first place - it’s just a web server responding to a request, the same way it would for a human clicking a button in a browser. Swap the AI model driving the skill next year, and the PDF generator doesn’t need touching at all.
That’s the actual dividing line this whole article has been circling, made precise: the risk was never “custom software.” It’s custom AI agent orchestration - the decision-making, multi-step coordination logic - built bespoke and wired into one vendor’s framework, the way the plumbing business’s lead intake logic was wired into GoHighLevel’s Agent Studio earlier. A narrow, boring, deterministic tool like the PDF generator above, that a skill calls through an ordinary interface, isn’t that. It’s not just safe to build, it’s often the better call than making an AI redo the same mechanical task from scratch every single time.
Why This Gets More Relevant, Not Less
The trend in the underlying models is toward taking a goal and plain-language context and simply handling it, without a custom-built orchestration layer standing in between translating everything into rigid steps. As models get better at working directly from instructions and context, the gap a bespoke agent used to fill keeps shrinking.
That’s the part worth sitting with before committing budget to custom agent development: the problem it solves is actively getting solved a different way, by the models themselves, on a timeline much faster than most custom software gets rebuilt.
The Portability Payoff
This is where it connects back to actually having your business documented properly. A well-built AI brain - your business facts, your processes, your decisions, written down in plain files you control - doesn’t care whether it’s fed to Claude, Gemini, or ChatGPT. Pair that with skills instead of agents, and nothing you’ve built depends on any single vendor staying exactly as it is today.
That’s the actual future-proofing. Not picking the framework you’re betting will win. Making sure nothing you’ve built needs it to.
Where This Approach Starts Missing the Mark
Push the plumbing business further still. It’s grown to fifteen vans on the road, and job routing now has to account for live GPS positions, shifting traffic, and emergency call-outs bumping the queue in real time. That looks exactly like the kind of complex, multi-variable problem that needs a custom AI agent built specifically for it. Tellingly, it’s genuinely outside what GoHighLevel itself does, too - real-time GPS-based dispatch routing is a feature its own users are still requesting, not something the platform behind the Agent Studio example earlier actually has.
It doesn’t need one, though - not really. The honest answer is the same one as everywhere else in this article: use an existing dispatch and routing platform built for exactly this problem, rather than commissioning bespoke agent software to solve it from scratch - on GoHighLevel or anywhere else. That’s a standard tool, not a custom one, and reaching for it doesn’t contradict anything argued here - it’s the same principle applied one more time.
The genuine exception - a business running one process at such constant, massive volume that even fractions of a second or fractions of a percent of error meaningfully matter - is real in principle. It describes an operation far larger than most small businesses will ever become. Worth being honest about that rather than manufacturing a “yes, but” just to sound balanced: for a business at this size, the case for bespoke, vendor-locked agent software essentially doesn’t arrive.
What to Build Instead
Invest in the AI brain first - it’s the part that keeps paying off regardless of which tools you use on top of it. Write your recurring processes as skills, in plain language, instead of commissioning bespoke agent software. And treat any specific AI tool you’re using today as replaceable by design, because it probably will be.
Frequently Asked Questions
Is a skill the same as a prompt?
They’re related, but a skill is more durable. A single prompt is a one-off instruction for a single conversation. A skill is a written, saved set of instructions covering how a recurring task should always be done - the kind of thing you’d write once and reuse indefinitely, the same way a staff handbook entry gets reused rather than re-explained from scratch every time.
Will a skill still work if I switch AI tools entirely?
Yes, with little to no modification, because it’s just plain text describing a task - it isn’t wired to any vendor’s API or framework. That’s the entire point of writing it this way rather than building custom agent software around it.
Gmail, Outlook, Google Calendar - aren’t those proprietary lock-in systems too?
Fair challenge, but there’s a real difference. Gmail, Outlook, and Google Calendar are tools the business was already using, and would keep using, entirely independent of any AI decision - it isn’t locking into them because of AI, they’re simply what the business already runs on. The lock-in this article warns about is a different thing: new decision-making logic, built specifically for an AI project and wired into one vendor’s proprietary framework - a dependency that didn’t exist until that project was commissioned. Connecting AI to Gmail or Outlook through an open connector doesn’t create that dependency. It just lets the AI work inside tools the business already knows and already owns the relationship with - enhancing how those tools get used, not changing which tools get used or how.
Is it ever okay to build custom software alongside a skill?
Yes, and often it’s the better call. The distinction isn’t custom software versus no custom software, it’s what that software actually does. A narrow, deterministic tool that performs one task the same way every time - like a PDF generator a skill calls through a simple request - isn’t the risk this article describes. The risk is bespoke decision-making logic built into one vendor’s agent framework.
How do I know if my business genuinely needs a custom AI agent?
For almost all small businesses, it doesn’t. The real case for custom agent development is a business running one process at such constant, high volume that fractions of a second or fractions of a percent of error meaningfully matter at scale - an operation far larger than most small businesses ever become. If you’re unsure, check for an existing standard tool first, because it almost always turns out to be the actual answer.
If you’d like help setting this up properly - the documentation, the skills, and the judgement to know when custom development actually is worth it - get in touch with our AI automation team to find out how YourStuff can help you today.
Photo by Sasun Bughdaryan on Unsplash
