By Jamie DeAngelis
Jamie DeAngelis is Head of Brand Strategy & Content at BRINK Interactive and founder of the BRINK Generative AI Center of Excellence. She designs and implements AI systems and operating models that embed strategy and governance directly into workflows — so organizations create durable enterprise value. She is the author of The Savvy Guide for Generative AI Beginners (2025).
Last month I listened to a smart executive explain why their AI rollout wasn’t working.
They’d done everything right. Licensed the best tools. Trained hundreds of people. Built prompt libraries. Celebrated early wins in the all-hands.
And yet: six months in, usage was up but outcomes were all over the map. A few teams were killing it. Most were stuck in what she called “expensive inconsistency.”
She wasn’t wrong about the tools. She wasn’t wrong about the training.
She was asking the wrong question.
The question wasn’t “Why won’t people use AI?”
It was “Why are we asking AI to work inside a system designed for humans?”
The Operating Model Gap
Here’s what I keep seeing.
Companies roll out AI. Adoption climbs. A few power users become legends. Everyone else gets … sporadic efficiency and a growing sense that maybe they’re doing it wrong.
Leaders interpret this as a people problem. “We need better training.” “We need champions.” “We need to drive adoption.”
But when I dig in, it’s almost never about motivation or skill.
It’s that the organization never redesigned the work.
McKinsey put a sharper point on it than most leaders do out loud: in their March 2025 State of AI survey, “out of 25 attributes tested … the redesign of workflows has the biggest effect on an organization’s ability to see EBIT impact from its use of gen AI.”
So the real question isn’t tool access.
It’s architecture.
Why “Just Add AI” Breaks Down
Generative AI arrives with a reassuring simplicity. It’s a chat interface. You type, it responds. Feels intuitive.
That simplicity is deceptive.
Because underneath every “simple” AI interaction is a set of decisions:
- Who owns the output?
- What’s the quality standard?
- What assumptions are we allowed to make?
- What has to be verified, and by whom?
- Where does the authoritative context live?
When those decisions are implicit, AI doesn’t create efficiency. It amplifies ambiguity.
You see it as “inconsistent quality,” but what’s actually happening is that different people are making different calls about what AI is allowed to do, what humans are responsible for, and what “done” means.
One person treats AI like a co-author. Another treats it like a search engine. A third treats it like a risky shortcut.
None of them are wrong. They’re just working in a system that never told them what right looks like.
And here’s the part that stings: training doesn’t fix this.
I’ve watched organizations pour resources into upskilling, and it helps — of course it helps. People need baseline fluency. (I literally wrote a book about this.)
But fluency doesn’t create a system. It just means more people are improvising at a higher level.
Deloitte’s January 2026 State of AI in the Enterprise report describes the same trap: “Companies are focused on building AI fluency instead of redesigning work around AI,” and 84% report they have not redesigned jobs or the nature of work itself around AI capabilities.
If your metric is “usage” and you’re still not seeing durable outcomes, the gap isn’t motivation.
It’s design.
You’re asking individuals to invent the same human–AI collaboration model from scratch, every single time.
Redesign the System, Not Just the Skillset
Most organizations already understand architecture in other domains.
Finance has controls. IT has standards. Legal has review lanes.
Those systems exist because individual competence doesn’t scale. Organizational reliability does.
Now AI is entering those same domains — and suddenly the architecture question becomes urgent. Work that once had clear boundaries now needs new ones. Work that never had formal structure suddenly requires it.
Not because AI is uniquely risky — though it can be. But because AI changes what can be delegated, automated, and scaled. And if you don’t redesign the work to account for that, you end up with an invisible operating model made entirely of improvisation.
Work architecture, in this context, has five components.
1) Roles, boundaries, and handoffs
Where does human judgment end and machine execution begin? And just as importantly — which machine execution?
The answer shifts by function, risk, and maturity. The real mistake is leaving it undefined. The second is assuming one agent should absorb everything a human once did.
Work doesn’t break that way. Research, synthesis, briefing, outlining, drafting, revising — each carries different standards and constraints. When a single generalist agent handles all of it, handoffs disappear, expectations blur, and rework multiplies.
Clarity comes from orchestration: defining roles across human→AI, AI→AI, and AI→human transitions. Distinct scope. Explicit ownership. Structured context transfer between steps.
That’s the architecture. Without it, you’re not scaling AI — you’re scaling ambiguity.
2) Decision pathways
Who decides what, and when?
Not just approvals. Decision ownership.
- What counts as acceptable?
- What requires escalation?
- What evidence is needed?
If every output becomes a one-off negotiation, you haven’t scaled AI. You’ve scaled overhead.
3) Knowledge encoding
If knowledge only lives in people’s heads — or in transient prompts — you can’t compound learning. You can only repeat it.
This is the part most organizations skip, and it’s why AI programs feel like Groundhog Day. Every team rebuilds the same prompts, relearns the same lessons, reinvents the same workarounds.
The value never travels.
The most successful organizations understand this. Microsoft’s 2025 Work Trend Index studied AI-native startups and found them using AI to “fundamentally reshape how work gets done … and build entirely new company structures.” These Frontier Firms are twice as likely to report their companies are thriving compared to their peers.
That’s not a tooling move. That’s a work design move.
4) Governance that enables, not just restricts
Governance gets a bad rap because it’s often designed as a bottleneck. But the goal isn’t to review everything. It’s to design a system where the right work is reviewed in the right way, consistently.
When governance is unclear, teams either move fast and create rework, or they slow down and debate every output. Neither scales.
The best governance I’ve seen isn’t a policy layer. It’s an enablement layer that reduces uncertainty and rework.
5) Context propagation
How does context move across people, teams, and tools?
Without a designed mechanism, context fragments. Teams make the same decisions repeatedly. The same mistakes recur. Every function builds its own private prompt library and its own private definition of “good.”
This is precisely the issue agentic AI brings into sharp relief. As Deloitte observed in their 2026 Tech Trends report, “Many organizations attempt to automate current processes rather than reimagine workflows for an agentic environment.” When AI agents can execute across domains — sales, services, supply chain, engineering — the cost of poor context propagation compounds exponentially. What was once a coordination problem between human teams becomes a coordination crisis between autonomous systems.
What Happens When Work Isn’t Redesigned
Once you see the architecture gap, you start recognizing the failure modes. They show up everywhere.
Prompt entropy: Everyone improvises. Quality forks. A few people become local heroes. Their colleagues copy what they can, but without shared standards, the organization accumulates variation, not capability.
I’ve seen companies with hundreds of custom GPTs, all slightly different, many abandoned. When standards change, nobody maintains the ecosystem. It just decays.
Delegation confusion: Teams swing between over-delegating (“AI can do this, right?”) and under-delegating (“I can’t trust AI with this”). Without explicit boundaries, every person makes that call independently.
The result? Inconsistent trust. And trust, once lost, is expensive to rebuild.
Context fragmentation: Individual wins don’t compound. Each team relearns. Each function rebuilds. Leaders can honestly report both “lots of AI activity” and “no consistent outcomes” because the work isn’t designed to travel.
Governance avoidance: Governance gets treated as overhead, so it’s postponed. Then teams either move fast and break things, or they slow down because every output is a debate.
Workflow friction: AI is inserted, not integrated. The early warning signs show up in plain language: “It’s easier to do it myself.” “I have to redo everything anyway.”
Sometimes that’s avoidance. Often it’s legitimate feedback about a workflow that never defined the handoff, the quality bar, or the review process.
Those are architecture problems. Not attitude problems.
How High-Performing Organizations Design for AI
The organizations that make this work don’t treat architecture as an abstract goal. They make it concrete.
Three shifts tend to matter:
1) They define roles instead of hoping people figure it out
Where should AI always step in? Where should it sometimes step in? Where must a human intervene?
And critically: what is each role responsible for?
That clarity isn’t about control. It’s about consistency. It turns AI from an optional hack into a designed capability.
2) They encode standards so quality is reproducible
When standards live only in someone’s head, every team improvises.
When standards are encoded — captured, shared, versioned — quality becomes easier to reproduce. Governance stops feeling like overhead and starts feeling like enablement.
3) They build systems that learn
Most organizations talk about “scaling AI.”
In practice, scaling isn’t about distributing access. It’s about distributing learning.
If every team is building its own prompts, heuristics, and quality definitions, AI will not compound. You’ll get a hundred local experiments, not organizational advantage.
The organizations that make this work build shared knowledge layers and feedback loops that improve the system over time.
The Real Advantage: System-Level Orchestration
The AI adoption story we keep hearing is a tool story: pick the right platform, train people, drive usage. That story is incomplete.
The advantage doesn’t come from having the tools. It comes from deliberately designing how human reasoning and machine capability work together. That’s a work architecture problem. And it’s a problem most organizations haven’t named yet.
If your AI efforts feel like a patchwork of local wins, uneven quality, and constant reinvention, it’s worth asking: Are we trying to scale AI without redesigning the work?
Because tools will keep improving. But tools don’t create shared boundaries. They don’t create shared standards. They don’t create shared memory.
Work architecture does.
Is your organization exploring where AI fits — or why it isn’t delivering at scale?
At BRINK, we design and implement AI systems and the operating models that make them work.