AI Didn’t Shrink My Company

It gave me a 100,000 sq ft Office

Ask what AI does for a business and it seems like the answer always includes something about “fewer people”, “slower hiring”, and “a leaner org chart.” While this is a real thing for some companies, this is not the interesting part. For Net Friends, AI has moved our walls.

History rhymes

In 1937, Ronald Coase asked a question nobody had bothered to ask in his paper The Nature of the Firm: if markets are efficient, why do firms exist at all? His answer was transaction costs. A company grows until organizing one more thing inside it costs the same as buying it on the open market. That line is the boundary of the firm.

Then Coase wrote something that reads like it was published last week. He observed that inventions which reduce the cost of organizing across distance tend to increase the size of the firm, and he named the telephone and the telegraph specifically.

Communication technology moves the boundary. Every communication shift since has confirmed it. The telephone. The personal computer. The internet, which is the closest corollary to where we are standing right now.

We are in the boundary shift phase with AI. What matters is what comes after it.

The second wave is where the value shows up

Here is the part people forget about the widespread advent of the Internet: companies poured money into IT through the late 1990s, and for years the returns seemed meager. Then productivity in IT-intensive industries surged in the early 2000s, well after the spending. Research from the San Francisco Fed points to the reason: getting value from a general purpose technology requires intangible organizational investment. This requires reworked processes, retrained people, and rewritten documentation. None of it shows up cleanly on a balance sheet, and all of it takes years.

Brynjolfsson, Rock, and Syverson named this the Productivity J-Curve. Early on, a general purpose technology looks like a cost center. Later, when the complementary investments get harvested, the curve turns up, and turns up in a BIG way. Once they accounted for software intangibles, their adjusted measure of total factor productivity ran roughly 16 percent above official figures by 2017.

Read that as both a warning and an instruction. The gains do not come from buying the tool. They come from the unglamorous work around the tool to integrate it in hundreds of small but meaningful ways.

What actually happened at Net Friends

We are three-plus years into internal AI adoption. That is long enough to be past the first-order effects and into the second, third, and fourth.

Picture moving from a 10,000 square foot office to a 100,000 square foot office all at once and without any desire to be in a bigger space. Same team, ten times the room, and overnight. That is what AI did to us.

I will be honest that all that extra space initially made me deeply conflicted. I knew I should not hire into it. I also knew I should not lease it out. So at first I largely let it sit for a bit. But it wasn’t long until I let people wander around in this extra business space.

Phase one is what I think of as that initial period of discovery and disorientation. Phase two was widespread exploration. I encouraged the team to probe, test, and temporarily relocate into unfamiliar corners of the new footprint, just to learn how the business looks and sounds from a different vantage point. That produced some of the most formative conversations I have had with people at every level of this company. Exploration has real value…but it also has a shelf life.

Phase three was the build. We stood up our own in-house AI tool. We ran multiple training sessions. We restructured our documentation so AI could actually consume it. We put AI to work in marketing and sales, not to replace our voice or cadence, but to deepen our research and business intelligence. We give a lot of that intelligence away free, through public channels and with prospects. Our customers get it deeper, richer, and sustained.

Phase three is over a year old now. I am proud of the multitude of things we accomplished in this short time. I am also clear-eyed that outside our virtual walls, almost nobody knows we built it, what we are doing with it, or where we are taking it.

There were real distractions along the way, and many twists and turns as the rapid evolution meant AI seemed to be something different each week. We started with Microsoft Copilot, moved to ChatGPT, landed on Claude, ran all three at once for a stretch, and dabbled a bit with Gemini, Perplexity, and others. There are genuine quality and use case differences between them. Most of it normalizes once you focus on what actually matters: skills, instruction sets, and constraints.

We also felt the pull of building and managing agentic AI. I found agents to be a real force multiplier. However, I did not find that they changed the fundamentals: governed usage, and outcomes you can measure and demonstrate.

Phase four: two parallel tracks

We are right at the cusp of it of what I see now is our fourth phase of AI. And this phase involves 2 parallel tracks:

• Track one. Weave AI management and consulting directly into our core contract service items. Every solution has to serve the customer first. Their data. Their security posture. The value they are trying to create in their own business.
• Track two. Use AI internally to deliver services that protect the human touch.

That second track will have a clear set of rules associated with it.

Count the machine-to-machine interactions in a normal workday. Computer to computer. Device to network. Cloud to endpoint. Application to application. Millions of them, every second, with no human anywhere in the loop. That is where AI belongs, working inside clear instructions and constraints on your behalf.

The moment a human is on either side of an interaction, ours or the customer's or both, AI moves backstage. Support role only.

One concrete example: MDR

Routine account compromises and device intrusions happen millions of times a day. Each one carries small nuances. However, the underlying pattern is consistent enough for a well-trained system to recognize most common attack sequences as they unfold. A well-instructed AI is ideal for these sorts of scenarios.

The speed argument is no longer arguable. CrowdStrike's 2026 Global Threat Report (which landed in my inbox on Monday) put average eCrime breakout time, meaning initial access to lateral movement, at 29 minutes. The fastest breakout they observed was 27 seconds. AI-enabled adversary activity rose 89 percent year over year. ReliaQuest's 2026 report found lateral movement in as little as four minutes, and containment in roughly four minutes with automation compared to as long as 16 hours without it.

No human answers a page in 27 seconds. No human wins a footrace against an automated attacker. The attackers are using AI. Defending without AI is a decision to lose.

There is a second benefit that gets less attention. When AI absorbs the well-understood attacks, your human security experts get their time back for the novel ones. That is where human judgment earns what you pay for it.

What comes next

In future articles, I plan to share more real-world examples of what we’re doing in this fourth phase of our AI build-out. What worked, what did not, and what it cost us to find out.

I am writing this for anyone who is trying to incorporate AI into their business in as meaningful a way as the internet and personal computers are. Some of you are further along the AI journey than I am, and my hope is that by sharing openly, you will too. Some of you are hesitant for philosophical or circumstantial reasons I respect, and my hope is that you find something practical here anyway.

Coase's boundary is moving right now, for every one of us. I hope these articles I’m regularly posting on LinkedIn will help everyone reading navigate these changing boundaries more effectively and efficiently.