Connector sprawl shows up before the bill does

TL;DRWhen your AI assistant is connected to everything, the first cost is speed, not answers. Watch the clock and ask why a simple question took minutes.

A few weeks ago we standardized our team at AI House on one AI harness. Then we did the obvious next thing and connected it to everything we use: email, calendar, Drive, Slack, Notion, our CRM, our events platform, our meeting notes, our database, our expense and accounting tools, and a few more. That came to seventeen connectors for a small team. Nobody decided we needed seventeen. Each one solved a real problem the day we added it.

The trouble shows up when a single question needs several of them at once.

Here’s the one that got my attention. I asked who I should invite to an upcoming event. A person who knows where things live could answer that quickly. The model had to check the events platform for who’d registered, our database for who’d said yes before and then not shown up, the CRM for who we know best, and Notion for how I write, so it could draft the invitations. That’s five systems for one question. It took about nine minutes and 1.9 million tokens to come back with fifteen names.

The answer was fine. That’s the part I didn’t expect, and I think it’s why this problem is easy to miss.

The answer still looks good

Today’s models are smart enough that if you give them enough material, they’ll sort it out. The invite list I got through five connectors probably looks a lot like the one I’d get from a single clean source. What’s different is everything around the answer.

Most connectors sit on top of APIs that were designed for code. A program calling an API doesn’t care if the response is long. It pulls the two fields it needs and ignores the rest. A model reads all of it. When I asked Claude why another query had taken about eight minutes, it told me one lookup had returned around thirty records at a few thousand tokens each, when the question only needed a single field. In my experience a lot of connectors work this way, because they weren’t built with a context window in mind.

And it compounds. Whatever lands in the context stays there for the rest of the conversation, so every follow-up question carries the noise forward and pays for it again. Anthropic’s engineers have described the same pattern from the other side: tool definitions and results can sometimes use up 50,000 tokens or more before an agent has even read the request.

Watch the clock, not the answers

So the first symptom isn’t a wrong answer. It’s a slow one, and a usage meter that moves faster than it should. Even if your team is on pricing generous enough that you don’t feel the cost yet, you’ll feel the wait. Nobody wants a simple question to take nine minutes.

What I ended up doing was building our own connector over a weekend. It sits in front of the systems we use most, replaces several of those individual connections, and returns only what a question needs. For events, for example, it now returns the counts we actually want instead of pulling every guest record (and everyone’s personal details) just to count them. It also lets me decide what the team’s assistant can reach, which was part of why I built it.

I don’t think most teams should have to do this. I’m an investor first, and writing this kind of tooling on a weekend is a side project, not my job. The more useful habit is cheaper. When a simple question takes minutes, ask the model why. Look at how many tool calls it made and how much came back from each. Turn off connectors nobody uses, and favor tools that summarize over tools that hand back everything.

Where this stops applying

We’re a small firm with a handful of heavy users, so I’m describing what sprawl costs a team like ours: time and usage. At larger scale I’d expect it to cost accuracy too. Anthropic says the most common failures in tool use are picking the wrong tool and passing the wrong parameters, and a longer list of tools gives the model more ways to pick wrong.

It’s also new. Six weeks ago I wouldn’t have recognized any of this, because we hadn’t picked a harness or wired anything up yet. The problem only appears after you’ve standardized and connected everything, and a lot of teams are about to get there.

A year ago the AI question most teams asked was which assistant to buy. The next one is what that assistant should be allowed to see, and how much of it at once.

Sources