Start in the tools you have, and build when you know why

Most of the organisations I speak to are already using Claude, ChatGPT, or Copilot in some form. Plenty of them are also weighing up spending real money on an AI product, an agentic platform or something built specially for them. What tends to get skipped in between is finding out what the tools they already have can do, and most people have barely scratched the surface.

I see AI as a journey, and a lot of this series has been about the parts of it you can’t shortcut: learning, iterating, and getting people to use what you provide. General platforms like Claude and ChatGPT are powerful, capable and easy to get hold of, which makes them the natural place to do that work. 

The question I find more useful is this: how do you know when the general tools are no longer enough, and what has to change when you want to become an agentic organisation? For us, much of the answer came down to who the agents were working for and what they needed to do.

We did eventually move our first agent out of the Claude app and onto our own platform, which has since become the Vaul Agent Hub, where our organisational agents now run. Four things drove that move, and the first was about who the agent was really working for.

From personal assistant to how the business works

In the setup we started with, agents in the Claude app ran on your own credentials. They could see what you could see, acted as you and fitted around how you liked to work. That made them brilliant personal assistants. It also meant the way your organisation worked with agents ended up being decided by each individual and how they happened to use them.

We wanted to decide, as a business, how Vaul works with agents. Left to each person, the ways of working vary, and so do the benefits: some people get a great deal out of it and some get little. Moving the agent onto our own platform meant it could work for the team, rather than for one person.

Your own harness, built for how you work

The second reason is the harness: everything around the model that lets it do useful work, which I wrote about in an ++earlier piece++. Claude and ChatGPT both come with excellent harnesses. You can customize them in some ways, but you’re still using theirs, and they’re built to be general enough for everyone (a bit like living in a beautifully furnished flat you’re renting: comfortable, but you can’t knock down a wall).

Building our own harness let us shape it around what we need and how we work: which steps it takes, where a person checks, what it records and how it reports back. That’s much of what made the narrow, traceable approach in the last piece possible.

Any model, for any job

The third reason is avoiding lock-in. Our platform lets us choose the model for each job. If we find better pricing across model providers, we can easily switch to a cheaper model that performs just as well. If we wanted to run our own hosted model, we could. If a better one comes along, we can switch quickly, and one agent can use different models for different tasks. Switching is the easy part. Knowing the new model does the job as well still takes testing.

Vendors change their plans too, which is fair enough, but it’s worth planning for. OpenAI, for example, plans to ++retire custom GPTs++ on 11 December 2026 with a migration path to plugins.  Custom actions built into those GPTs won’t transfer automatically. If your agents only exist inside one vendor’s product, their roadmap becomes your roadmap.

Keeping your context your own

The last reason is context. Keeping it somewhere you own, rather than inside any one tool, is something we already treat as best practice, and the move really put that to the test. Brian described the risk well: “you’re kind of locked in, right? Because you’ve got your context in there and you’ve taken the time to configure it.” Moving forced us to check that all our context (our ways of working, our templates, our knowledge of how we run engagements) wasn’t locked inside the Claude app, and could be used by any agent on any platform.

Build when you know why

None of this means everyone should rush off and buy or build their own platform. Building your own means owning it: the hosting, the monitoring, the fixes and the upkeep when models change. We took all of that on when we moved, and it’s real, ongoing work. For most organisations, the right first step is still the tools you already have, used well, with a small group, on a real problem.

The time to build is when you can say clearly why. For us, that came down to four reasons: agents that work for the organisation and not only the individual, a harness shaped around how we work, the freedom to choose and switch models, and context that belongs to us. If you can’t yet name your reasons, stay where you are a little longer. You’ll learn more there than by building a platform you may not need.

Next in the series: Measure business outcomes, not agent counts.