Buy or configure your first agent, and build only when the software is your product

In the span of four days last week, Salesforce agreed to buy the customer service agent company Fin, formerly Intercom, for $3.6 billion. SpaceX agreed to acquire the AI coding company Cursor for $60 billion in stock. And Snap spun its internal AI video team out into a separate company called Dotmo, citing the high cost of doing that work in-house.

Earlier in the month, OpenAI told developers that its visual Agent Builder is being deprecated and will shut down on November 30, 2026, pointing them to its Agents SDK or ChatGPT Workspace Agents instead.

If you run a company of 20 to 200 people and you are deciding how to get your first agent working, these stories are worth a minute of your attention. They bear on the first decision every owner faces with AI: buy something, set something up on a platform, or build your own.

My answer is short. Buy if the tools you already pay for can do the job. Configure if they can’t. Build only when the software is the thing you sell.

TimelineThree weeks of AI vendor news

Jun 3OpenAI deprecates Agent Builder
Jun 15Salesforce agrees to buy Fin
Jun 16SpaceX agrees to buy Cursor
Jun 18Snap spins out AI video team

What the three options actually mean

People use these words loosely, so here is how I use them.

Buying means turning on an agent that comes inside software you already run. Your help desk has a resolution agent. Your CRM has a prospecting agent. Your accounting software reconciles transactions. You pay for it, you set the rules, and the vendor maintains it. Fin is a good example of how this is priced now: about 99 cents per resolved conversation, and nothing when the agent fails to resolve it.

Configuring means taking a general platform, such as a workspace agent from one of the big model companies or an automation tool, and setting it up for your workflow. You write the instructions, connect it to your documents and systems, and decide what it may touch. Nobody on your team writes code, or very little.

Building means engineers, code, hosting, monitoring and tests that you own and maintain. It gives you the most control and it costs the most, every month, forever.

Most companies should start by buying

For a first agent, I want the shortest path to a real result in one department. That is almost always the tool you already pay for. The data is already there. The permissions are already set. Your team already knows the screens. The vendor carries the cost of keeping up with new models.

The common objection is that the built-in agent is generic. That’s true, and for a first workflow it rarely matters. Answering order status questions, drafting follow-up emails and summarizing calls are generic tasks. You win on how well you write the rules and how carefully you check the output. A built-in agent with sharp instructions beats a custom one with vague ones.

Snap’s decision is a useful reference point. This is a public company with its own generative AI research operation, run by its chief technology officer. It still decided that carrying a dedicated AI video team was too expensive for work outside its core business. If that math is hard for Snap, it is harder for a 60-person distributor.

Configure when the job crosses systems, and build only when it is the product

Buying stops working when the job spans tools that don’t talk to each other. Say you run a 40-person agency and the task is a weekly client report that pulls from ad platforms, a project tracker and a spreadsheet of retainer hours. No single vendor’s built-in agent sees all of that. That is when you configure: one platform, your instructions, connections to each system, and a person who reviews the output before it goes to the client.

Configuring has its own risk, and OpenAI’s announcement is the clear example. If you set up workflows inside a product that the vendor later retires, you rebuild them. November 30 is a hard date for anyone whose agents live in that builder. Platforms are still the right call for most cross-system work. Keep your instructions and examples in your own files, outside the platform, for the reasons I come back to below.

Before you configure anything, decide who owns it. A configured agent needs an owner the way a shared spreadsheet does: someone who updates the instructions when a policy changes, checks the output every week and notices when a connected system changes a field name. If nobody has that job, the agent quietly gets worse and people stop trusting it.

I have built software myself. I built Track Raider, a music discovery and set-building app for DJs, on my own, and that was the right call because the app is the product. Nobody was going to sell me a better version of my own idea.

That is the test I use. If the agent is how you compete, and a customer would pay for what it does, building can make sense. If the agent does internal work that every company in your industry also needs, someone is already selling it, and they will maintain it better than you will.

Own the parts that move with you

Whichever path you pick, the vendor can change. Products get bought, renamed, repriced or shut down. Fin changed its name from Intercom on May 12 and agreed to a sale five weeks later. You cannot control that, and you don’t need to. You need to own the pieces that let you move.

ChecklistOwn these whatever you choose

The written instructions
Twenty real test cases
Your data in your own systems
A log of what a person corrected

The instructions are your standard operating procedure, written so a new hire or an agent could follow it. The test cases are real examples with the right answer next to each, so you can check any new tool against them in an afternoon. The correction log is the record of what your reviewers changed, which tells you where the agent is weak. If you have those four things, switching vendors is a week of work. If they live only inside the vendor’s settings screen, switching starts from zero.

What to do this week

Pick the one department where you want your first agent and write down the single task it should take over.
List the software that department already pays for and check whether any of it includes an agent for that task.
If it does, turn it on in a test setting and run your real examples through it.
If it doesn’t, write down every system the task touches. That list tells you whether you need to configure a platform.
Start the instruction document and the test cases now, before you choose anything, and keep them in your own files.