For the past 18 months, I’ve watched companies adopt AI.
I’ve seen rollouts that worked, rollouts that stalled, and rollouts that looked good in a board update but changed almost nothing in the actual work.
At Golf, we spent a lot of that time helping enterprises adopt AI securely: permissions, visibility, audit, governance, control over what agents can access and do. For a long time, I thought this was the main bottleneck. If companies could use AI securely, adoption would follow. I don’t believe that anymore.
Security still matters. A lot. If agents are going to touch your CRM, ERP, inboxes, calendars, documents, customer data, policies, code, or internal tools, security is not optional. You do not want agents quietly breaking your business. But security is the prerequisite. It is not the end goal.
The bigger problem is that companies are trying to bolt AI onto old workflows instead of rebuilding the workflows for the technology we now have.
And I think the next real wave of AI adoption belongs to non-software companies.
Energy. Logistics. Manufacturing. Healthcare. Pharma. Industrial companies. Operationally heavy businesses. The companies that actually move the world.
These companies are the core of the economy. They produce, transport, insure, treat, manufacture, repair, distribute, and operate the things most people depend on every day. Meanwhile, most of the AI conversation still happens inside software-first companies, where the workflows are already closer to code, APIs, tools, and experimentation.
Software companies are adopting AI faster because their work is easier to reshape around software. Non-software companies are slower because their processes are messier, more regulated, more physical, more human-dependent, and more deeply embedded in how the business actually runs.
But that is also why the upside is so large.
The common failure mode I keep seeing is simple: a company buys AI licenses, maybe launches an internal chatbot, runs a few trainings, and calls it AI enablement… and then nothing meaningful changes.
Maybe 10% of the company uses it. The curious people. The power users. The people who were already AI-pilled and willing to experiment after work. These people are incredibly valuable. Every company should find them, support them, and learn from them.
Most employees already have a full-time job. They are measured on delivering the same quality of work they delivered before AI. They do not have time, permission, or incentives to discover use cases, evaluate models, build automations, redesign workflows, manage risk, and prove ROI.
So AI becomes extra work.
The company thinks it gave people tools. In reality, it gave them another thing to figure out on top of their actual job. This is why most bottom-up adoption stalls.
If you want people to use AI casually, buy them tools. That is fine. Let them explore. Don’t kill the creativity. Don’t optimize token spend on day one. Give the early adopters room.
But if you want ROI, you need a different approach. You need to rebuild workflows from first principles.
The wrong question is:
How do we add AI to this process?
The right question is:
What would this process look like if human intelligence was not scarce?
Most company workflows were designed around human intelligence being expensive, limited, and trapped inside people’s heads. That is why we have so many handoffs. That is why people read documents, summarize context, copy data between systems, chase approvals, review policies, compare records, remember edge cases, and explain the same things to other humans over and over again.
Those steps were not designed because they are sacred. They exist because, historically, software could not do the judgment-heavy parts. If intelligence, memory, document understanding, tool use, and reasoning are available inside the workflow, you should not ask how to make every existing step 10% faster. You should ask which steps should exist at all.
What is the actual goal of the workflow?
Can we go from A to B directly?
Where does a human truly need to be involved?
Where is the human only there because the system had no context, no judgment, no memory, or no access to tools?
This is where real AI adoption starts. Not with a chatbot. Not with a training. Not with a “top 10 AI use cases” workshop.
With workflow ownership.
Someone senior needs to own the workflow and say: this is the business outcome we want, this is how the work happens today, and this is what we are willing to change. The most successful AI rollouts I’ve seen are top-down in this sense. Not because leadership micromanages prompts, but because leadership makes workflow change legitimate. The company does not just ask employees to “be more productive.” It changes the operating model.
The work gets rerouted. The system handles what it can handle. Humans focus on exceptions, judgment calls, approvals, and edge cases. That is where ROI comes from. And this is also where things get hard.
A serious AI workflow is not just “call an LLM.” You have a model. A prompt. Tools. Memory. Evals. Monitoring. Permissions. Human checkpoints. Business metrics. And all of these parts move.
The model changes. If you depend on a closed model provider, the exact behavior behind the same API name can change over time. Performance can improve, degrade, or shift in ways you did not expect. The prompt changes. A small prompt edit can change how the agent behaves, what it refuses, what it overdoes, and which tools it calls. The tools change. APIs change. Internal systems change. Permissions change. The agent may stop calling a tool correctly because the interface changed slightly. Memory changes. If you want agents to stop repeating mistakes, they need to learn from actual work. But memory cannot just be a garbage pile of transcripts. It needs structure. It needs ownership. It needs a way to capture what mattered and feed that into the next run. Evals change. You are not evaluating a deterministic script. You are evaluating a stochastic system inside a messy workflow. You need enough examples of what good looks like. You need to know if the agent is getting better, worse, cheaper, more reliable, or quietly drifting.
This is why companies need to think much harder about model sovereignty.
If intelligence becomes part of your operating model, you cannot blindly hardcode your company into one model provider forever. You should be able to evaluate multiple models for different workflows. You should be able to switch when a cheaper model is good enough. You should be able to detect when a model degrades. You should not let your workflow become hostage to one black-box brain.
This does not mean every company should train its own model. Most should not.
It means the workflow should be designed so intelligence is a replaceable layer, not a vendor prison.
The same is true for ROI. In the early phase of AI enablement, usage metrics are okay. You want people to try things. You want experimentation. You want to see where curiosity emerges. But after a few months, usage is not enough.
Tokens spent is not ROI. Messages sent is not ROI. Number of active users is not ROI. “We launched an internal chatbot” is not ROI. Real ROI sounds different.
How much does one ticket resolution cost before and after?
How long does this process take?
How many cases can one operator handle?
How many errors happen?
How many exceptions need human review?
How much work is rerouted from humans to the system?
What is the cost per completed task?
Is quality better, worse, or the same?
Can we prove the new workflow beats the old workflow?
That is the standard.
This is what I want Golf Labs to work on.
Golf Labs is not here to run another AI training. It is not here to sell a magic platform. It is not here to bolt a half-baked chatbot onto a broken process and call it transformation.
The point is to go into non-software, operationally heavy companies and understand the real workflow.
Energy. Logistics. Manufacturing. Healthcare. Pharma.
We want to sit close to the work. Understand what people actually do. Understand the tools, approvals, data, exceptions, and constraints. Understand why adoption stalled. Then help redesign the workflow around the primitives we now have: agents, tools, memory, evals, monitoring, permissions, and human checkpoints.
I’m not promising a magic wand.
Your data may be messy. Your systems may not talk to each other. Your process may be undocumented. Your incentives may be wrong. Your team may not trust AI yet. All of that matters.
But I do believe this:
The next wave of AI adoption will not come from asking employees to discover AI on their own.
It will come from companies that are willing to rebuild core workflows for the technology that now exists.
If this sounds familiar, if you tried to roll out AI and it did not change real work, I want to talk.