Why a leaner tool stack outperforms a maximally connected one, and how to trim yours
In June, David Cramer did something that looks like a mistake. He nearly halved the number of tools his company’s AI could see at once.
Sentry’s AI agent used to see 14 tools every time it ran. Cramer’s team cut that default down to 8 and tucked the rest behind a layer the agent only reaches for when a task actually needs them.
The agent did not get weaker. It got sharper, with fewer wrong calls, faster responses, and more reliable output.
Most mid-market teams are running in the opposite direction. Marketing and ops leads are wiring their AI into Salesforce, HubSpot, Slack, the CMS, the analytics dashboard, and the project tracker, working from the idea that more connections means more capability. The research keeps landing on the other side of that bet.
Why more tools make your AI worse
The numbers are blunt. Independent analyses of production agents point the same way. They run best with three to five tools, and quality drops sharply once the count passes twenty. A separate test watched accuracy fall off a cliff rather than slide down a ramp, holding strong at ten and twenty tools, then collapsing around a hundred. In one experiment, a strong model’s accuracy on a scheduling task dropped from 43 percent with four tools to 2 percent once the count climbed past fifty. Somewhere along that curve, adding capability stopped helping and started breaking the thing you built.
Picture a sharp new hire on their first day. Give them one clear task and the two tools they need, and they move fast. Hand them that same task plus a hundred-page binder describing ninety other tools they could theoretically use, and watch them slow down. They read the binder. They second-guess. They reach for the wrong thing. Your AI behaves the same way. Every connection you add is another page in the binder it has to consider on every single request.
So the real question for a team lead stopped being how much you can connect. The better question is which connections actually carry weight in the work you do. That sounds abstract until you run the audit.
Run the audit before you add anything else
Open a document and list every AI integration you currently have live. Next to each one, write the specific task it performed in a real workflow this week. Real tasks only. “Pulls the deal stage from Salesforce into the Monday pipeline summary” counts. “Connected to HubSpot” just tells you a wire is live somewhere in your account.
Most teams find half their list falls into that second group. The connection exists because it was available and easy to switch on, never because a repeatable workflow leans on it.
Decide what is load-bearing
A load-bearing integration is one where a real process breaks if you remove it this week. That is the whole test, and you run every item through it.
If pulling an integration would stop a report from generating, block a handoff, or leave a customer waiting, it stays. If pulling it would change nothing you would notice by Friday, it goes on the cut list. Be honest about the gap between a connection you use and a connection you simply like having around. Your AI does not care that the Slack integration feels modern. It sees one more option competing for attention every time you ask it to do something.
Most stacks have three to six integrations doing genuine work, plus a long tail doing nothing but adding noise.
Trim, then retest
Turn off everything that failed the load-bearing test. Do it in one pass rather than gradually, so the difference is easy to feel. Then run your normal tasks for a week and watch response speed and error rate. Both should improve. If some workflow you forgot about does break, you have just learned that integration was load-bearing after all, and you switch it back on with evidence instead of habit.
Cramer’s team did not guess their way to a leaner setup. They measured, cut back what the agent had to weigh, and watched the output improve. You have the same option, and it costs you nothing beyond the comfort of a long integration list.
The instinct to connect everything comes from a good place. You are trying to build leverage, to get the AI doing more of the work so you can do less of it. A bloated stack hands you the feeling of leverage while taxing every output behind the scenes. A lean one hands you the real thing. The fastest, most reliable AI setup in your business is almost certainly smaller than the one you are running today, and getting there is mostly a matter of deleting.
. . .
Want to save hours each week by turning work into repeatable AI workflows?
The Fortune 100 AI Skills Library™ includes plug-and-play prompts built to save leaders time and money. Copy, paste, and edit in 60 seconds, then apply them across planning, execution, and reporting.
Top comments (0)