There’s one thing I particularly like about running open-weight models locally:

You can actually try things with real, mission-critical data without the dissonance that typically comes with sending that data to a 3rd-party API.

And I've increasingly come to believe that this is probably one of the (if not THE!) key advantages of local AI, especially in larger enterprises.

Local AI gives you both capability and freedom to actually do things.

Today, I’d like to share why I think a relatively small investment in local AI can solve a much bigger organizational problem.

Let's dive in!

Black Friday offer

The Everything
Bundle 2026

Every paid workshop from 2026, plus 12 AI Skills you can run immediately inside Claude or ChatGPT – at 91% off.

(Limited one-time offer. Expires on Nov 26 and won't come back after.)

The infrastructure trap

For large enterprises – or any company with more than a couple hundred employees for that matter – the biggest roadblock to getting ROI from AI usually isn’t technical.

The problem lies in the organization's ability to move quickly, try out new things and take risks without spending months going back and forth in internal approval chains.

Large organizations usually don’t lack technology.

I’d argue that in many cases they’re actually over-equipped. SAP for ERP, Salesforce for CRM, Snowflake for warehousing, Databricks for analytics, etc.

The irony is that having access to all this technology doesn't necessarily mean you have the freedom to actually use it.

Rather than enabling rapid innovation and progress, these heavy systems often end up locking teams in. Because these environments are so tightly guarded that getting even the most straightforward AI use cases off the ground becomes incredibly difficult – especially if they touch data from multiple systems.

The compliance and integration overhead alone can make simple ideas ROI-negative before a single prompt is written. Not because the value isn’t there, but because the cost of even trying is too high.

Let me give you an example:

I’ve worked with a finance team that had a great idea to automate a repetitive task with AI – account reconciliation at month’s end.

But the data and context necessary for account reconciliation live in several different systems, including SAP for the ledgers, another system for transaction data, a group inbox for communication, and SharePoint for supporting documents.

Just giving an AI agent access to even one of these systems required weeks of internal reviews and waiting for available integration capacity from IT. In a regulated industry, these weeks can quickly turn into months.

What happens if you can’t iterate quickly

When you are trapped in this kind of approval chain, your team loses one of the most critical assets in the AI age: agency.

Automating work with AI can be hard, and a lot of ideas you have at the beginning will not work out.

So you have to be able to prototype your ideas quickly, before you ever think about scaling AI projects.

You cannot solve most AI challenges simply by thinking through them in committee meetings and meticulously planning things out. You need a gardener mindset.

If every minor experiment requires weeks of vendor negotiations or IT sign-offs, your flexibility dies and so does your AI ambition – one meeting at a time.

You end up blindly asking vendors to build what you think you need, without actually having a solid understanding of what the technology is realistically capable of achieving.

And there's another problem:

You cannot afford to find out which AI ideas are actually worth pursuing.

If testing an idea already requires months of coordination and substantial investments, you'll naturally become much more selective about what you're willing to try.

How to solve this problem

I believe one of the most effective ways to break through this organizational gridlock is to create a separate experimentation environment for AI.

And one of the most practical ways to do this is by investing in local AI hardware to "buy" agency and autonomy.

I’m not talking about six-figure infrastructure investments. Getting a relatively cheap (in enterprise terms) computer like a Mac Mini / Studio or DGX Spark provides you with a nice sweet spot for local AI workloads.

Selected Local AI hardware options

You simply pull exports from your main systems – like a CSV or an Excel spreadsheet – drop them onto your local hardware, and build a system around it independently.

(Pro tip: Have a local AI agent build this system for you.)

Of course, this assumes you have permission to work with the data in that environment. Local doesn't automatically mean compliant or secure.

But that's precisely why I would invest in getting such an environment approved once, with clear boundaries around what data can be used, rather than going through a new approval process for every little experiment.

Coming back to our account reconciliation example:

Instead of building a deep system integration first, we decided to work with exported Excel files, as well as other information needed like selected emails, PDF documents, and some other spreadsheets. We put them onto the machine, and ran them through a local agentic system:

Local account reconciliation input

Within hours, we had the first working prototype ready that could reconcile a subset of transactions and identify exceptions:

Local account reconciliation demo output

Of course, this wasn’t something to release into production the next day.

But you now have a reference point for what AI can actually do for your particular problem, and a better sense of what you’re asking for.

You might discover that AI works surprisingly well.

Or that it doesn't.

Or that the real problem isn't the reconciliation itself, but the fact that information flows poorly through your organization.

All three outcomes are valuable because they help you decide what to do next.

Now, instead of asking permission to explore an idea, you can go directly to your vendors or internal IT, show them the working system, and confidently articulate your needs to them.

Or if everything works nicely, you could even challenge them: "Why can't we get this from you?"

Having a proof of work under your belt is always more convincing than a nice PowerPoint.

(And sometimes, these local prototypes can even become the starting point for production systems if you can manage to bring them under appropriate IT and security controls.)

Key benefits

To sum it up, here’s what a local AI setup gets you when developing AI use cases in the enterprise:

  1. Faster experimentation: Go from an idea to a prototype in hours or days instead of weeks or months.

  2. Lower cost of failure: Test more ideas at a fraction of the usual organizational cost.

  3. Data control: Experiment on actual data without sharing it with third-party AI providers.

  4. Fewer dependencies: No need to involve IT every time. Come back to them once you really understand what you need.

  5. Better investment decisions: Explore dozens of potential AI use cases before committing to full implementation projects.

Conclusion

I’m not saying that getting a local AI setup approved in your organization is easy.

But it might turn out that fighting this battle once allows you to avoid many other battles down the line – or fight them with much better arguments.

Right now, organizations I work with are looking at dozens or sometimes even hundreds of engineered AI use cases.

You can’t take each of these through a planning committee. At the same time, you also can’t just give Claude or Copilot access to all your production data and hope for the best.

That's why I think we need to become much better at separating AI discovery from AI delivery.

Discovery is about figuring out what's worth doing. Delivery is about doing these things reliably.

Both phases operate under very different processes, resources, and rules.

In the end, the biggest advantage of local AI in the enterprise is not the technology itself, but having both the capability and the freedom to actually do things with AI.

And in large organizations, that combination can be surprisingly hard to come by.

See you next Saturday,
Tobias

PS: Ahead of Black Friday, I’m launching my Everything Bundle 2026. It gets you immediate access to all 15 paid live events I hosted in 2026. And if you don’t have time to watch all this, the bundle also comes with 12 AI Skills – ZIP files you can easily drop into Claude or ChatGPT to bring the knowledge from these workshops into your AI. → See the video demo here

Reply

Avatar

or to participate