I have been working with a lot of companies that are trying to become AI-first.
At first, that usually means helping individuals get much better at using AI. They build workflows, create agents, learn how to give the AI useful context, and figure out where it is actually saving them time.
Then the workflows and processes start going beyond the individual. An AI workflow takes messy documents from a partner every morning and turns them into the format the company actually uses. A shared dashboard combines data that lives across five SaaS products. Someone builds a company-wide “research agent” that researches new markets using the company’s latest definitions of a good prospect, a real competitor, or an important signal.
Soon the AI knows that when someone says MRR, they do not just mean the generic finance definition. They mean the exact definition this company uses, where the number comes from, which accounts count, which adjustments matter, and who should review it when the number looks wrong.
It is no longer just automation. It is institutional memory - the company’s knowledge, history, processes, workflows, and way it uniquely does work.
And it likely will be the new backbone and most significant software infrastructure in the AI era.
Twenty years ago, CRM and ERP systems became the backbone of company data. Companies installed systems like Siebel and SAP on their own servers and then moved to cloud SaaS systems like Salesforce, HubSpot, and NetSuite. These systems did not just store contact records and transactions. They became irreplaceable parts of businesses, generating trillions of dollars in revenue for their providers.
And AI company brains will become the next version of indispensable software.
The next generation of company knowledge will not only store what happened. It will store how the company responds when something happens. It will hold a living workflow for cleaning a partner feed, the reasoning for how to interpret a metric based on past history, the trusted sources for research, the historical context behind a customer escalation based on dozens of previous examples, and the scheduled tasks that keep all of it current.
More importantly, it can help improve those processes. A workflow can collect its exceptions. An agent can notice that the same type of document keeps getting classified incorrectly. It can self-review that pattern, update the instructions or data mapping, and have the next run work better.
Unlike a static Zapier, Clay, or N8N recipe, the company AI brain self-improves. The system is not only running the company’s playbook. It becomes part of how the company learns to make the playbook better.
Once a company starts working with a particular AI brain infrastructure, it becomes deeply entrenched in the information that makes the company function. It will know more about the company than any individual person does. It will hold the edge cases, corrections, history, and little decisions that never make it into a process document.
At that point, moving is not like migrating CRM data. Depending on where you put it, this can become close to a one-way door. You might be able to export the data, but extracting the accumulated context, workflows, agent behavior, and learned ways of operating the business, then moving all of it somewhere else with different behaviors will be incredibly hard.
And with this comes a new strategic question for companies: where is all of this going to live?
Companies need to realize they are making a real decision now, before they wake up with 100 daily workflows, company dashboards, and years of accumulated context inside a system they never seriously chose.
There are four main ways companies can do this. They can push the easy button with a model provider. They can use an existing SaaS incumbent like Salesforce, HubSpot or Notion. They can bet on a new closed AI startup. Or they can build on open source, either managing it themselves or using a managed open-source provider.
There is no “always correct” answer. I will go through each, then come back to the real question: what kind of company are you, and which processes do you actually need to win on?
The easy button: OpenAI or Anthropic
Most companies are heading towards the “easy button” of letting OpenAI and Anthropic own and manage their company processes.
To help customers work more productively, OpenAI and Anthropic are building workplace products with shared projects, memories, scheduled tasks, easy-to-create shared dashboards, and even shared agents like Claude Tag. Companies without IT can go from zero to a surprisingly useful company AI system with almost no IT support. Someone on the operations team can build an intake workflow in the morning and have a SOC 2-compliant brain running that afternoon.
From a company’s perspective, it seems like an easy win. The company gets AI infrastructure up and running with basically no work. And for the providers, it is a double win. OpenAI and Anthropic give customers an incredibly useful platform while binding those companies more deeply to their products.
And this is happening at just the right time. Open-source models are showing that selling model APIs will become a commodity. The model providers’ biggest chance to maintain a sustainable advantage is not just getting you to use their models. It is getting you to run your business and all of its processes through them, with them helping you manage it. And the more company processes, history, and context that live inside an AI workplace product, the harder it is to leave.
That is the secret sacrifice most companies will make. In exchange for this amazing, easy access to an AI brain, most companies will bind themselves to one provider or the other. This is not unlike Salesforce, NetSuite, or any other major provider.
For a huge percentage of companies, that is fine. If these processes are important but not how you win, the speed and quality of the easy button are probably worth paying for.
The incumbent play: Salesforce, HubSpot, Notion, and everyone else
The existing SaaS companies see the AI brain opportunity too.
Their pitch is obvious. Your data is already in their product. Why not add the AI agent, workflow, dashboard, and research layer there too?
For many companies, this is a reasonable choice. It is easy and familiar. It keeps fewer systems in the mix, especially for companies that have not invested much in Claude or OpenAI.
But this is where the model providers have an advantage. OpenAI and Anthropic are building AI-first products from the ground up, while big SaaS companies are adding AI to architectures, product lines, and business models built years before this shift. Many have also spent decades focused on one department, such as sales, marketing, or finance. That makes them strong within those departments but weaker at connecting information across them, like linking specific product changes to related customer support requests.
The SaaS AI add-ons will work and create real efficiencies in the short run. But they will keep companies working the way they currently will as opposed to taking more radical organizational changes that could have bigger impacts long term.
That may be fine if this is not where you need to win. But companies that rely on the AI layer of an existing SaaS vendor risk in the medium to long term being lapped by competitors who build on more AI-first infrastructure.
The beautiful closed startup
There are also a lot of sharp founders looking at this opportunity and building beautiful closed SaaS products. Think of these as focused, opinionated versions of the OpenAI and Anthropic tools. These beautifully designed AI-brain and operational tools are trying to become your new company hub. They will work better in some areas and have features and focuses that Claude and OpenAI do not. They are trying to become the next Salesforce or Workday.
Some will be genuinely useful. A lot will receive VC funding. And a few will win.
But I am skeptical that this is the best long-term answer for a company putting its crown jewels into AI workflows. The company will still be locked in and reliant on one provider. The purchasing company is also betting that the startup they choose to store their AI brain will keep its focus, raise enough money, survive the market, and stay ahead of the model providers and incumbents who can copy much of the product. In such a fast-changing area, this becomes a big risk.
That may be worth it for a specific problem. It is a harder bet when the product becomes the place where your institutional knowledge lives.
Open source and managed open source
The final option that is emerging is open source and managed open source. A lot of developers and companies are realizing the difficulty of creating a SaaS company and are moving down the stack to build open-source stacks either for pride, personal fulfillment, or to make money in adjacent channels.
Open source has exploded in popularity with hundreds of new solutions. Thousands of AI projects like OpenClaw, OpenWork, GBrain, Cognee, and Graphiti have emerged this year, with many quickly becoming some of the most popular open-source projects on GitHub.
Like Linux for operating systems, Magento for e-commerce, and Postgres for databases, open source is becoming a viable choice for AI brain hubs.
Open source has a lot of benefits. Companies can keep control of their data, workflow definitions, code, prompts, and integrations. They can customize the system for their industry, change models, and move to another host. They are not asking one vendor for permission to shape the part of their company they think is most important.
And companies can customize it. Unlike an existing SaaS product, a company can completely customize an open-source system to its business practices. If your company needs something unique, like special per-region customizations for each US state or zip code, it may be much easier to add a custom plugin to an open-source system than to cobble together the pieces from the Anthropic or OpenAI offerings. You will have many more options to experiment and grow with new AI-first ways of doing things if you build on open source than if you are constrained by what the model providers offer.
But open source comes with its own downside. Someone has to run it. Someone has to keep the stack current. Someone has to make sure it is locked down and secure. Someone has to make sure a collection of agents, plugins, models, and data sources does not become a hard-to-maintain spaghetti mess.
There is also a hybrid option: managed open source.
Think managed Postgres or managed WordPress. A vendor hosts and maintains the open-source system, but your company retains its data, code, workflows, and definitions. If it stops being a good partner, you can move your stack. You are paying for operational help, not handing over the actual brain of your company. However, even with managed open source, you will still need resources to improve the open-source system, build plugins, and make sure the data is backed up appropriately. You will likely need to build more yourself than if you rely on model providers, who will continue to build at a rapid pace.
Fully managing your own open-source stack makes sense for companies that are on the bleeding edge, have a strong technical team, have very strict security requirements, and really believe their AI processes are core to the company’s advantage.
Managed open source makes sense for companies that want to push the boundaries of AI-first, fully control their data, and are willing to put in the extra work and resources to grow it.
The choice depends on your company’s focus
When I consider where to store your AI brain, I keep remembering the framework we used at Shop It To Me called Win, Play, and Lose. The idea was simple: no company can be the best at everything.
You choose a small number of things, perhaps 15 percent, where you need to win. You choose a larger set, perhaps 60 to 70 percent, where you just need to play, meaning you need to be competitive. And you consciously decide which 15 to 25 percent you are willing to lose at, even if competitors make those things look good on a feature list.
Companies should apply the same framework to their AI infrastructure.
If your unique differentiator is manufacturing amazing physical products and your internal sales reporting just needs to work, use the easy button. Let OpenAI, Anthropic, Salesforce, or HubSpot handle it. You do not need to own every workflow. Honestly, this will be most companies.
But what if your ability to learn faster from customers, partners, operations, or data is actually how you win? What if the processes being encoded into the AI are the key things that make you better than your competitors?
Then you should be very careful about putting all of that into a black box you cannot control or port out of.
The companies that win in the next few years will not be the companies with the most AI tools. They will be the companies that decide which AI processes are generic, which ones are strategic, and where their company’s AI brain should live.
If you start to split the components - intelligence, context, tools - this problem looks different. You don’t have to lock all your context and tools into an intelligence provider. You can build a modular system and have multiple types of models working off of the same context, serving different use cases as you see fit.