Executive Summary
In 2024, a Fortune 500 company spent $12 million on enterprise ChatGPT licenses. Six months later, usage analytics told a brutal story: 8% adoption. The other 92% of employees opened it once, maybe twice, then went back to doing things the old way. The CIO’s mistake wasn’t choosing the wrong model. It was assuming that access equals adoption—like buying everyone a gym membership and wondering why nobody’s getting fit. The real problem with enterprise AI isn’t capability. It’s that 99% of people don’t know what to do with it. And betting your future on OpenAI or Anthropic or whoever wins the model wars means you’re locked into someone else’s interaction paradigm when what you actually need is an Intelligent Interaction Field—a model-agnostic system that meets people in their actual work, learns from their context, and compounds value you own. Adoption doesn’t come from better models. It comes from better interaction design.
The $12 Million Gym Membership Nobody Used
Let me tell you about a problem that looks like success until you check the data.
A global financial services firm—call them Apex—did everything right. They read the headlines about AI transformation. They assembled a taskforce. They negotiated an enterprise deal with OpenAI. They rolled out ChatGPT licenses to 15,000 employees with training videos, lunch-and-learns, and executive championship.
The announcement email used words like “game-changer” and “competitive advantage.”
Three months later, the CIO asked a simple question: “How many people are actually using it?”
The answer: 8%.
Not 8% of possible use cases. Not 8% below target adoption.
8% of employees had used it more than twice.
The other 92% opened it once, asked it something generic, got an answer that was… fine… and then went back to doing their jobs the way they’d always done them.
$12 million spent. 8% adoption.
The CIO’s first instinct was to blame change management. “We need better training. We need more use cases. We need to show them what’s possible.”
But here’s what he was missing:
The problem wasn’t that people didn’t know AI existed. The problem was they didn’t know what to do with it in their actual work.
The Salad Problem (Or: Why Capability Doesn’t Equal Adoption)
You know you should eat more salad, right?
You know it’s healthier than the burrito. You know it improves energy, longevity, all of it. You have access to salad—there’s a place that sells it two blocks from your office.
And yet.
You don’t eat salad.
Not because you can’t afford it. Not because you don’t believe in its benefits. But because behavior change doesn’t happen just because the option exists.
You’d have to:
- Decide you’re eating salad today (cognitive load)
- Walk two blocks (friction)
- Choose from 47 ingredients (decision fatigue)
- Convince yourself it’s worth it (motivation)
- Repeat tomorrow (habit formation)
And even if you do it once, that doesn’t mean you’ll do it again.
Access is not adoption.
This is exactly what’s happening with enterprise AI right now.
Companies are buying everyone “salad access”—ChatGPT Enterprise, Claude for Work, Copilot licenses—and then wondering why people keep eating burritos (doing their jobs the old way).
Because here’s what nobody’s solving:
99% of employees don’t understand what AI can actually do in their specific work context.
They open ChatGPT. They see a blank text box. They think: “What am I supposed to ask this thing?”
They try something generic: “Summarize this email.” It works. Cool.
Then they close it and go back to their actual job—which involves navigating seven systems, coordinating with three departments, making judgment calls based on context that exists in their head, not in a prompt.
The problem isn’t capability. The problem is that AI isn’t embedded in the interaction where work actually happens.
The Model War Trap (Or: Why Betting on One Vendor Makes Adoption Harder)
Now here’s where most companies make it worse.
They pick a vendor. Let’s say OpenAI. They go all-in. Enterprise contract. Integration roadmap. Training programs designed around ChatGPT’s interface and paradigm.
And then they’ve locked themselves into someone else’s way of working.
Here’s the uncomfortable question: What if OpenAI isn’t the winner?
What if Anthropic leapfrogs them? What if open-source models (which are already closing the gap) become the standard? What if a model you’ve never heard of becomes the default in 18 months?
You’ve just bet your adoption strategy on a vendor that might not be relevant by the time people actually figure out how to use it.
This is like building your entire gym routine around one piece of equipment—and then that equipment gets discontinued.
But here’s the deeper problem:
When you bet on one vendor, you’re not just betting on their model. You’re betting on their interaction paradigm.
ChatGPT thinks in prompts and responses. Copilot thinks in document assistance. Claude thinks in conversational threads.
But your business doesn’t think in any of those ways.
Your business thinks in episodes and moments:
- A customer calls with a complex request
- A salesperson needs to configure a quote under time pressure
- An underwriter is evaluating risk with incomplete information
- An operations manager is rerouting resources because a supplier failed
These aren’t prompt-and-response problems. These are interaction field problems—where multiple signals, systems, and stakeholders need to coordinate in real-time, and the AI needs to learn from every outcome.
Betting on one model vendor means you’re forcing your interaction design into their paradigm. That’s backwards.
The Real Architecture: Intelligent Interaction Fields
Let me take you back to Apex.
After the 8% adoption disaster, they stopped asking “How do we get people to use ChatGPT?” and started asking a different question:
“Where in people’s actual work does value leak because interactions break?”
And here’s what they discovered:
The problem wasn’t that people needed a chatbot. The problem was that their work happened in interaction fields that weren’t intelligent.
Let me explain what I mean.
What Is an Interaction Field?
An interaction field is where multiple participants—customers, employees, systems, partners—come together to create value in a specific episode or moment.
Examples:
- A loan application (customer, underwriter, credit system, compliance, pricing model)
- A supply chain disruption (supplier alert, inventory system, logistics, customer commitments)
- A customer service escalation (agent, customer history, product data, supervisor judgment, resolution protocols)
These aren’t isolated tasks. They’re systems of interaction where value is created (or destroyed) based on how well information flows, decisions get made, and learning compounds.
Traditional AI asks: “What can the model do?”
Intelligent Interaction Fields ask: “What needs to happen in this episode, and how do we orchestrate all the participants—human and machine—to create the best outcome?”
The Four Components
Apex built what they called an Intelligent Interaction Field for their loan underwriting process:
1. Sense: The system captured signals in real-time—applicant data, market conditions, risk indicators, underwriter workload, regulatory constraints, even patterns from previous decisions.
2. Reason: Here’s the critical part—the system applied Apex’s judgment, not OpenAI’s generic reasoning. It knew their risk appetite, their strategic priorities, their values. And it was model-agnostic—it could use GPT-4, Claude, Llama, or Mistral depending on what worked best for that specific reasoning task.
3. Act: The system orchestrated across participants—pulling credit reports, flagging risks, pre-populating documents, escalating to humans when judgment was needed, coordinating with compliance.
4. Learn: After every loan decision, the system asked: What worked? What didn’t? What should we change? And that learning stayed at Apex—it didn’t leak to a vendor’s next training run.
What Changed
Within 90 days, something remarkable happened:
Adoption went from 8% to 67%.
Not because they did more training. Not because they created better use cases.
Because people stopped needing to figure out what to ask AI. The AI was embedded in the interaction where their work actually happened.
Underwriters didn’t open ChatGPT and think “What should I prompt?”
They opened a loan file, and the Intelligent Interaction Field was already sensing, reasoning, acting, and learning—in the context of their actual work.
The system came to them. They didn’t have to go to the system.
Users stopped navigating tools. They focused on judgment. The Intelligent Interaction Field orchestrated everything else.
Model-Agnostic by Design (Or: Why You Don’t Bet on One Vendor)
Here’s what Apex did that most companies don’t:
They made their Intelligent Interaction Field model-agnostic from day one.
They didn’t build around ChatGPT’s API. They built around interactions—and then let the system choose which model (or combination of models) worked best for each reasoning task.
For simple document summarization? A lightweight open-source model (fast, cheap, good enough).
For complex risk analysis requiring regulatory knowledge? GPT-4 or Claude (deeper reasoning, worth the cost).
For pattern recognition across millions of loan decisions? Their own fine-tuned model running on proprietary data (competitive advantage they owned).
The architecture was technology-agnostic. Because they knew the model wars weren’t over—and they weren’t going to be hostages to whoever won.
As one Andreessen Horowitz survey of 100 enterprise CIOs confirmed: adoption of open-source models like LLaMA and Mistral is accelerating, especially at larger enterprises who realize the capability gap is closing and vendor lock-in is dangerous.
You don’t bet your adoption strategy on one vendor. You build interaction fields that work no matter which model wins.
Why This Solves the Adoption Problem
Let’s go back to the gym membership analogy.
Why don’t people use gym memberships?
Because you have to:
- Decide to go (motivation)
- Drive there (friction)
- Figure out what equipment to use (cognitive load)
- Create your own routine (expertise required)
- Track your own progress (discipline)
Now imagine if the gym came to you. If it knew your goals. If it designed the workout. If it tracked progress automatically. If it adapted based on what worked.
That’s the difference between giving people access to ChatGPT and building an Intelligent Interaction Field.
Access requires them to:
- Remember the tool exists
- Navigate to it
- Figure out what to ask
- Translate the answer back into their actual work
- Repeat every time
Intelligent Interaction Fields remove all of that.
The system senses the episode they’re in. It reasons about what needs to happen. It acts by orchestrating the workflow. It learns from the outcome.
They don’t adopt a tool. They just do better work—and the intelligence is embedded in the interaction.
That’s why Apex went from 8% to 67% adoption in 90 days.
Not better training. Better design.
The Big Tech Dependency Trap
Now let’s talk about the other risk nobody’s discussing.
When you go all-in on OpenAI (or Anthropic, or Google), you’re not just betting on their model. You’re creating another Big Tech dependency.
Remember what happened with cloud computing?
Companies migrated everything to AWS or Azure or Google Cloud because it was easier than managing infrastructure.
And then they woke up one day and realized:
- They couldn’t leave without massive switching costs
- Pricing was whatever the vendor decided
- Their entire operation depended on one company’s strategic priorities
We’re about to do the exact same thing with AI.
Except this time, the dependency is deeper. Because it’s not just infrastructure—it’s intelligence. It’s decision-making. It’s the system that learns from your interactions.
And if that system is owned by OpenAI or Anthropic or whoever, you’re a tenant, not an owner.
You don’t control:
- What they learn from your data
- How they price access in 24 months
- Whether they prioritize your use cases in the next model version
- What happens if they pivot, get acquired, or fail
The smarter strategy is to embrace an interaction-first, model-agnostic architecture—what we call Intelligent Interaction Fields—where you own the orchestration and can plug in whatever models make sense.
You’re not dependent on OpenAI winning. You’re not locked into Anthropic’s roadmap.
You’re building a system that gets better no matter which models commoditize.
Don’t create another Big Tech dependency. Build Intelligent Interaction Fields that you own.
The Trust and Context Advantage
Here’s what Apex realized six months into their Intelligent Interaction Field deployment:
They weren’t just solving an adoption problem. They were building a competitive moat.
Because every loan decision that flowed through their system taught it something that competitors couldn’t access:
- Which risk signals actually predicted default (vs. which ones were regulatory theater)
- How to balance speed and thoroughness for different customer segments
- What underwriter judgment calls created the best outcomes
- How to adapt pricing in real-time based on competitive pressure
This wasn’t generic “AI knowledge.” This was proprietary intelligence about how Apex creates value in their interaction field.
And because the architecture was model-agnostic, they owned it—regardless of which foundation model they used for reasoning.
MIT research on brand trust shows that admired brands “enable, entice, and enrich customers” through repeated, reliable value delivery. But you can’t build that trust by renting someone else’s AI.
You build it by designing Intelligent Interaction Fields that learn your customers, your context, your values—and compound that intelligence over time.
When Apex’s competitor signed an enterprise ChatGPT contract and tried to replicate their success, they got generic AI capability that worked the same way for everyone.
They didn’t get 18 months of compounded learning from Apex’s interaction field.
That’s the moat.
The Interaction Field Economics
Let me show you what happened to Apex’s numbers.
Adoption: 8% → 67% in 90 days (because intelligence was embedded in work, not bolted on)
Cycle Time: Loan underwriting went from 4.2 days → 1.8 days (because the interaction field orchestrated across systems and participants automatically)
Cost-to-Serve: Dropped 31% (because context didn’t leak, rework vanished, and coordination became automatic)
Default Prediction Accuracy: Improved from 83% → 91% (because the system learned from proprietary patterns in their interaction field)
Employee Confidence: Underwriters reported 40% higher confidence in decisions (because the system augmented their judgment instead of replacing it)
And here’s the economic proof:
Revenue per underwriter increased 23% (because they processed more loans with better risk pricing)
Operating margin improved 280 basis points (because efficiency compounded without sacrificing quality)
Customer NPS jumped 18 points (because faster, more accurate decisions created better outcomes)
When you solve the adoption problem by building Intelligent Interaction Fields, economic value compounds—not linearly, exponentially.
Because every interaction makes the next one better. And you own the compounding.
Why Model-Agnostic Matters More Than You Think
Here’s the uncomfortable truth about the model wars:
Nobody knows who’s going to win.
In 2023, everyone thought OpenAI was untouchable. By 2025, open-source models like LLaMA and Mistral were matching performance. By 2026, DeepSeek was beating GPT-4 on some benchmarks—and running on hardware a fraction of the cost.
Capability is commoditizing fast.
So what’s the smart architecture?
Build for interactions, not models.
Apex’s Intelligent Interaction Field doesn’t care whether GPT-6 or Claude-4 or Llama-5 ends up being the best model. Because their value doesn’t come from model access.
Their value comes from:
- Interaction design that meets people where they work
- Context ownership that compounds with every decision
- Learning loops that make their system smarter every day
- Model flexibility that lets them use whatever works best—now and in the future
When the next model breakthrough happens, they plug it in. When a model gets too expensive, they swap it out. When open-source catches up, they switch.
Their competitive advantage isn’t the model. It’s the Intelligent Interaction Field they built.
And their competitors can’t rent that from OpenAI.
The Choice: Adoption Theater or Intelligent Interaction Fields
Let me bring this back to where we started.
Most companies are doing adoption theater:
- Buying licenses ✓
- Running training sessions ✓
- Creating use case libraries ✓
- Measuring usage metrics ✓
And then wondering why adoption stays stuck at 8%.
Because they’re treating AI like a gym membership—giving people access and expecting them to figure out what to do with it.
The companies solving this are building Intelligent Interaction Fields:
- Identifying where work actually happens (episodes and moments)
- Designing systems that sense, reason, act, and learn—in those interactions
- Making intelligence model-agnostic so they’re not hostages to vendor wars
- Owning the learning so competitive advantage compounds
One approach treats AI as a tool people need to adopt.
The other embeds intelligence in the interactions where value is created.
Guess which one gets 67% adoption in 90 days?
The Vivaldi Framework: Intelligent Interaction Fields
At Vivaldi, we call this approach Intelligent Interaction Fields—and it’s built on three principles:
1. Interaction-First, Not Model-First
Design around where work happens, not around what ChatGPT can do. Start with episodes and moments. Then orchestrate intelligence—human and machine—to create better outcomes.
2. Model-Agnostic by Default
Don’t bet on one vendor. Build architecture that can use GPT-4, Claude, LLaMA, Mistral, or your own fine-tuned models depending on what works best for each reasoning task. Your moat isn’t model access—it’s interaction design and owned learning.
3. Compounding Intelligence You Own
Every interaction should improve the next one. But that learning must stay proprietary. Build loops where sense → reason → act → learn happens within systems you control, not in someone else’s API.
This is how you solve the adoption problem and build competitive advantage simultaneously.
Growth Becomes Self-Optimizing (When You Own the Interaction Field)
Let me close with where Apex is 18 months later.
Their Intelligent Interaction Field has processed 340,000 loan decisions. It’s learned from every one.
And here’s what the CEO told me last month:
“We used to worry about AI adoption. Now we worry about competitors who don’t have what we’ve built.
Our underwriters don’t think about ‘using AI.’ They just do their jobs—and the interaction field makes them better at it. Faster decisions. Smarter risk pricing. Higher confidence.
And the system keeps learning. Every day it gets better at sensing patterns, reasoning about edge cases, orchestrating workflows, and improving outcomes.
Our competitors can buy the same models we use. But they can’t buy 18 months of compounded learning from interactions we designed and own.
That’s not a technology advantage. That’s a business model advantage.”
The Question That Actually Matters
So here’s where we are:
The model wars are ending in commoditization. Open source is closing the gap.
The adoption crisis is real. 99% of people don’t know what to do with AI—and access doesn’t solve that.
Betting on one vendor (OpenAI, Anthropic, Google) creates dependency without solving adoption or building moats.
The companies winning are building Intelligent Interaction Fields:
- Model-agnostic by design
- Embedded in where work actually happens
- Learning from every interaction—and owning that intelligence
- Compounding economic value: adoption, speed, quality, margin, trust
The question isn’t “What’s our ChatGPT strategy?”
The question is: “What Intelligent Interaction Fields should we build that we actually own—and that solve the adoption problem by meeting people where they work?”
Because adoption doesn’t come from better models.
It comes from better interaction design.
And growth doesn’t come from AI capability.
It comes from Intelligent Interaction Fields that compound value you control.
Interaction design is growth design. And growth, properly designed and owned, becomes self-optimizing.