Integrations
Agent Sprawl: Why Your Business Is Building the Same Integration Three Times
Every AI tool wants its own link to your accounts. How to keep one register and one connection
Three tools, three connections to the same ledger: three bills, three sets of keys, three slightly different answers.
On 24 September, data platform Dataiku launched Agent Management, a product whose first job is simply to find every AI agent a company is running, on whichever platform built it, then track what each one costs, what it delivers and how risky it is. Its own research gives the reason: in a Harris Poll survey of 685 chief information officers for Dataiku, 84 percent said employees are building agents faster than IT can govern them.
That is an enterprise problem at enterprise prices. The small-business version is the same shape, just smaller, and it is fixable in an afternoon.
What does agent sprawl look like in a small business?
It rarely looks like a decision. The practice turns on the AI features in its email suite. Someone connects a chat assistant to the CRM to draft follow-ups. An automation tool gets an AI step that reads from the accounting system. Each one asked for access, each one got it, and each one built its own connection.
Three tools and three systems can mean nine separate connections. Each has its own keys, its own permissions, its own bill and its own idea of what the data says. When a client asks why two reports disagree, the honest answer is often that two agents read the same ledger through two different pipes.
What is the fix?
Two things, in order.
First, a register. One row per agent or AI tool: what it is, which systems it can reach, whether it can change anything, who owns it, and when its access was last reviewed. This is most of the benefit on its own, and it is the same list the privacy and AI rules coming into force will ask for.
Second, connect once. Instead of every tool holding its own key to the CRM or the ledger, put one connection in front of each system and let the tools share it. The emerging standard for this is the Model Context Protocol, or MCP, an open protocol introduced by Anthropic in 2024 and now supported widely. AWS published a pattern in August for exactly this: host one MCP server and let several agents reuse the same tools "instead of authoring them from scratch again".
Start where the keys are: your accounting system, CRM and email each list the apps connected to them. That list is your register's first draft, and it usually has something on it nobody remembers adding.
What changes when the connection is shared?
One set of permissions. Read-only for every tool, by default, set in one place.
One log. Every request from every tool, through the same door, with a time and the tool's name on it.
One fix. When the CRM changes its API, or a key needs rotating, it happens once, not three times in three products.
One answer. Every tool reads the same data the same way, so two reports built by two assistants finally agree.
Where to start this week
- Export the list of connected apps from your accounting system, CRM, email and file storage.
- Put every entry in the register, with an owner. Anything with no owner gets switched off for a week to see who notices.
- Change anything that does not need write access to read-only.
- For the one system most of your tools touch, usually the CRM or the ledger, plan a single shared connection next.
Sources: Dataiku announcement, SiliconANGLE on the survey, AWS Machine Learning Blog, Model Context Protocol.
Key takeaways
- Sprawl is the defaultEvery AI product asks for its own access to your systems. Left alone, the connections multiply faster than anyone tracks them.
- Register before you rebuildA one-page list of agents, what they reach and who owns them is most of the benefit, and costs an afternoon.
- Connect onceA shared connection layer, such as an MCP server, gives every agent the same tools, one set of permissions and one log.
Frequently asked
What is MCP?
The Model Context Protocol is an open standard, introduced by Anthropic in November 2024, for connecting AI tools to data and systems. You build or install one MCP server for a system, such as your CRM, and any AI tool that speaks MCP can use it, instead of each tool needing its own custom connector.
Do I need a product like Dataiku's to manage agents?
Not at small-business scale. Products like that are built for enterprises running hundreds of agents across several platforms. A spreadsheet with one row per agent, what it can reach, who owns it and when its access was last reviewed does the same job for a practice with five.
Is one shared connection a single point of failure?
It is a single point of control, which is the point. If a connection breaks, one fix repairs it for every tool. If a key leaks, one rotation closes it. The trade is worth it once more than one AI tool touches the same system.
How do I find the connections I already have?
Start in your core systems, not your AI tools. Your accounting software, CRM, email and file storage each list the apps connected to them. Export or screenshot those lists, and ask of every entry: who set this up, and do we still use it?