How to Get Your Team Actually Using AI

73% of companies say AI is used regularly across their business. Only 10% say it's core to operations.

That number comes from Publicis Sapient's 2026 enterprise AI survey, and it's the most honest summary of where most organizations actually stand. Broad experimentation. Shallow change. A lot of people who have technically "used AI" and almost no one who has restructured how work gets done because of it.

If you're a COO, a department head, or a founder trying to close that gap, the problem is almost never the technology. Deloitte's 2026 State of AI in the Enterprise, which surveyed 3,235 leaders across 24 countries, found only 25% of them had moved 40% or more of their AI pilots into production. The documented bottlenecks across both reports aren't about model quality. They're about weak operating design: unclear ownership, missing governance, legacy workflows that nobody redesigned, and teams that never developed a shared language for what they're even trying to do.

What follows is about fixing that. Not with inspiration, but with a framework and a plan you can run in the next 30 days.

The Real Failure Mode Is Psychological, Not Technical

Most AI rollouts fail quietly. Tools get purchased. A few enthusiastic early adopters start using them. Leadership declares an AI initiative. Then nothing much changes for everyone else.

Failure starts as a psychological problem before it becomes an operational one. When a small, vocal fraction of a team moves fast with new technology, it opens a gap. For everyone else that gap is intimidating, and intimidated people stop asking questions, because asking one means admitting you're behind. The silence widens the gap. Patrick Gilbert covers this dynamic directly in Never Always, Never Never: the people who fall behind aren't lazy or resistant. They're embarrassed, and embarrassment shuts down learning faster than almost anything else.

A Harvard Business School field experiment run with BCG consultants, co-authored by Wharton's Ethan Mollick, is the clearest evidence that this matters. Consultants using AI completed 12.2% more tasks, finished them 25.1% faster, and produced results rated more than 40% higher in quality than the control group. But the most useful finding for managers isn't the average improvement. It's the distribution. Participants below the average performance threshold improved by 43%. Participants above average improved by 17%.

AI training narrows performance gaps when people are shown how to use it on real work. Laggards benefit more than stars. That's the opposite of the "only the tech-savvy will survive" narrative that makes the gap worse.

Your job isn't to make your best people better. It's to give everyone enough shared language and guided practice that the whole distribution moves.

Give People a Framework Before You Give Them a Tool

Most organizations follow this instinct: buy software, run a demo, encourage usage, measure adoption. The sequence is backwards.

People don't fail to use AI because they haven't seen enough demos. They fail because they have no mental model for when to use it, what to hand over, and where to stay in the driver's seat. Without that structure, every decision defaults to doing the task solo, the same way it's always been done.

In Never Always, Never Never, Patrick Gilbert describes a framework called the 4x2 Model of Work that gives teams exactly this structure. Every task falls into one of four modes: Design, Problem-Solving, Decision-Making, and Building. In an AI-first organization, you no longer have a solo option for any of them. You either Copilot (you stay in control, AI acts as a capable navigator) or you Delegate (you hand the execution to AI and move into the role of editor and quality checker).

Copiloting fits Design and Decision-Making. You don't ask the machine to invent your strategy or make your judgment calls, but you use it to pressure-test assumptions, explore alternatives, and synthesize data into something you can actually act on. Delegating fits Building and repetitive Problem-Solving. If a task is mechanical, predictable, or high-volume, it should move to an AI agent or an automated workflow. Your job shifts from doing to verifying.

Cultural shift happens when every person on your team runs this audit on their to-do list before starting any task: Can this be delegated? If not, can I copilot it? If the answer to both is no, that's worth examining. If the answer to both is yes and you're still doing it solo, that's an operational problem, not a personal preference.

The framework is more useful than any specific tool recommendation because it works regardless of which AI products you've purchased. It gives people a decision rule, not just permission.

To go deeper on how this connects to building organizational AI capability, the AI Double Helix framework covers the two-strand model of Internal Efficiency and External Value that underpins this whole approach.

Where People Actually Are (And Why That Matters)

Not everyone on your team is starting from the same place. Pretending otherwise is the mistake that causes rollouts to either move too slowly for your strongest people or leave everyone else behind.

Never Always, Never Never describes an AI Maturity Ladder with four rungs:

• Dabbler: Uses ChatGPT to draft an email or summarize a transcript. AI is a novelty. Nothing about how they approach work has changed.
• Practitioner: Has moved into the 4x2 Model. No longer works solo by default. Uses AI to solve specific daily friction points and has developed prompting skills that produce consistently useful output.
• Architect: Stops looking at individual tasks and starts looking at systems. Builds automated workflows that institutionalize efficiency across the team, not just for themselves.
• Strategist: Uses AI to build things that create external value, bespoke applications, proprietary data models, tools that help clients do things they couldn't do before.

The ladder matters operationally because it tells you what kind of support different people need. A Dabbler doesn't need advanced prompting techniques. They need shared language and a safe place to practice without judgment. A Practitioner doesn't need more demos. They need use cases specific to their role and clear guidance on where the guardrails are. An Architect needs infrastructure budget and organizational permission to redesign workflows rather than just improve their own.

When you conflate these groups and give everyone the same training, you waste time for the Architects and lose the Dabblers entirely.

Structural changes that come later in this process, once you've gotten the team moving, are covered in the guide to building an AI-first marketing team, which walks through what that looks like in practice.

The Jagged Frontier Problem

One reason AI training fails even when people are willing is that the capability boundary is non-intuitive.

Mollick's BCG research describes it as the "jagged frontier": AI helps substantially on some structured knowledge tasks while remaining unreliable on others that look nearly identical on the surface. A tool that writes an excellent competitive analysis summary can confidently produce a hallucinated statistic in the next paragraph. A workflow that automates client intake forms perfectly might fail badly on an edge case that a junior employee would have caught in two seconds.

Training cannot just teach people how to use AI. It has to teach them when not to trust it. Every team needs a working vocabulary for this: what counts as a "high-stakes output" requiring verification, what categories of tasks need a human approval gate before anything ships, and where the risk of a confident wrong answer is high enough that delegation is the wrong choice regardless of efficiency.

Skipping this vocabulary produces one of two failure modes. Either people become over-reliant on outputs they should be checking, and errors start appearing in client deliverables, proposals, and decisions. Or people become so skeptical of AI errors that they abandon the tools entirely after one bad experience. Neither outcome is adoption. Both are predictable if you skip the judgment training.

This is part of what Patrick Gilbert describes as the human element in building an AI-first culture: the technology itself is rarely the problem. Shared language around it is.

What the Organizations That Scale Actually Do

The same Deloitte survey found a consistent pattern in the organizations that get past pilots: they assign named owners, define the scope of each use case, and redesign the workflow around it. The ones that stall leave AI as a side tool with no clear home and no one accountable for it.

The plan below builds exactly that. It needs no budget, no consultants, and no new software. It needs about four hours of leadership time a week and one person whose job it is.

---

The 30-Day AI Enablement Plan (Copy and Run This)

Week 1: Establish Shared Language

Day 1-2:

• Publish a plain-English glossary of 10-15 terms. Include: prompt, model, agent, workflow, retrieval, hallucination, approval gate, human-in-the-loop, sensitive data, delegation, copiloting. Keep definitions to two sentences maximum. No jargon inside the definitions.
• Do not assume people know these terms. Do not present this as remedial. Present it as the team's operating vocabulary going forward.

Day 3-5:

• Run one 30-minute all-hands demo using a real company task, not a toy example. Show the actual tool on an actual piece of work your team does every week. Walk through one Copilot example and one Delegate example side by side.
• End the demo with this question to the group: "What's one task you do every week that might fit the delegate column?"

---

Week 2: Map the Territory

Task for every team or department:

Each team identifies three things before the end of the week:

1. One repeated task where AI could save meaningful time and the output is low-stakes enough to delegate or copilot with light review.

2. One task or category where AI should never be used without human review or should not be used at all (client-facing legal language, anything involving personal data, high-stakes financial decisions).

3. One task that sits in the middle, where AI could help but needs an approval gate before output goes anywhere.

Collect these from every team lead. This is your actual use-case inventory. It will tell you more about where to focus than any vendor assessment.

Assign two named owners this week:

• One person owns AI enablement (use-cases, training, adoption support).
• One person owns AI risk and governance (data rules, verification standards, approval gates).

These can be the same person in a small organization. Each function needs a face attached to it.

---

Week 3: Role-Based Examples and Safe Use Policy

Build three one-page role profiles:

Each major role type in your organization (individual contributor, manager, senior/leadership) should have documented:

• Two Copilot tasks specific to that role, with example prompts
• Two Delegate tasks specific to that role, with example workflows or tools
• The approval gates that apply to that role's outputs

This is the step most rollouts skip, and it is the one that matters most. People don't adopt general AI capability. They adopt AI for their own job.

Publish a "Safe Use" one-pager. It should cover:

• What data cannot be pasted into any external AI tool (client data, personal data, proprietary financials)
• What outputs require verification before use (anything citing statistics, anything going to a client, anything used in a decision)
• Who to ask when someone isn't sure
• What to do when AI produces something wrong (report it, don't just delete it and move on)

One page. Plain language. No legalese.

---

Week 4: Measure One Workflow

Pick the single most-used delegatable task from your Week 2 inventory. Measure it:

• Time spent on this task before AI, per person per week (survey, don't guess)
• Time spent after AI is introduced
• Error or correction rate in the AI-assisted output versus prior baseline
• Adoption by role: who is using it, who isn't, and what the reported barriers are for the non-users

This is your first real data point. It tells you whether the tool choice and training approach are working before you scale either one. Don't skip this step in favor of moving to the next tool. One measured workflow is worth more than ten untested ones.

---

Manager 1:1 Questions: Placing Someone on the Maturity Ladder

Use these in regular check-ins, not as a formal assessment. The goal is to understand where someone actually is, not to grade them.

• What AI tools have you actually used in the last two weeks?
• Which specific task did you try to improve with AI?
• What part of the output did you trust, and what did you verify?
• Where did AI save you time, and where did it create cleanup work?
• Do you know the difference between a prompt, a workflow, and an agent?
• If someone asked you to design a simple agentic workflow, what would you include?
• What risks would make you avoid AI on your current work?
• What would help you use it more safely or more confidently?
• Which terms are still unclear or intimidating?
• What's one thing you could teach a colleague about AI use this month?

That last question is the most useful one. Teaching is the fastest signal of genuine understanding, and it creates peer-to-peer diffusion that no top-down training program can replicate.

---

What This Is Actually Building

A 30-day plan is not an AI transformation. It's the foundation for one.

Organizations that close the gap between "we use AI" and "AI is core to how we operate" are the ones that treat the two sides of this as inseparable. Patrick Gilbert calls this the AI Double Helix: Internal Efficiency and External Value as two strands that feed each other. Efficiency creates bandwidth. Bandwidth enables experimentation. Experimentation produces new value. New value funds more investment in efficiency. The loop compounds.

You cannot start at the External Value strand. The operational foundation comes first: shared language, clear ownership, redesigned workflows, real measurement. Organizations that jump straight to "what can AI do for our customers" without fixing the internal operating model end up unable to scale anything they build.

Marketing teams will find the AI resource gap framework useful for understanding what this looks like when AI changes the economics of what small teams can actually execute. But the operating principle applies across functions: bandwidth has always been the constraint, and AI changes what's feasible without requiring headcount to grow proportionally.

Gilbert's account of this at AdVenture Media is that the shift from experimenting with tools to building systems that changed the work did not come from the tools getting better. It came from the team sharing a way to decide what to delegate, what to copilot, and what stays fully human. That decision architecture is what the 30-day plan builds.

The post on how to become an AI-first organization is worth reading alongside this one for more on what AI-first looks like in practice versus in a pitch deck.

The One Thing to Do First

Build the glossary.

Not a tool evaluation. Not a vendor demo. Not a strategy document about AI's potential impact on your industry.

A plain-English list of 10-15 terms that everyone in your organization will use when talking about AI, starting this week. Prompt. Model. Agent. Workflow. Hallucination. Approval gate. Human-in-the-loop.

Shared language removes the intimidation that keeps most of your team silent. It converts a black box into something discussable. And it costs nothing except an hour of someone's time to write it.

Every other step in this plan depends on people being able to talk to each other about what they're doing. Start there.