Denis O'Shea: AI Adoption Strategy - 5 Steps Before You Scale

Imagine this scenario.

Your company buys AI licenses. Employees begin experimenting. Some people save hours. Others barely touch the tools. A few quietly use applications IT does not know about.

Then security runs a scan and discovers thousands of sensitive files are more accessible than anyone realized.

Finally, the board asks two reasonable questions:

Is AI actually creating value? And should we expand the investment?

Nobody has a defensible answer.

This is no longer a hypothetical problem. AI adoption is moving faster than many organizations’ ability to redesign work around it. In July 2026, Gallup reported that 47% of U.S. employees said their organization had integrated AI tools, while 52% said they were already using AI in their own role. Yet Gallup’s earlier 2026 research found that only about one in ten employees in AI-adopting organizations strongly agreed that AI had fundamentally transformed how work gets done.

The gap between using AI and becoming better because of AI may be one of the most important management challenges of the next several years.

Denis O’Shea, founder and CEO of technology services company Mobile Mentor, has seen this pattern before. More than two decades ago, while working at Nokia, a customer confronted him with an uncomfortable question:

“Why would we buy any more technology from you when our customers are not using the technology we bought last year?”

That encounter eventually led O’Shea to build a business around helping organizations extract value from technology rather than simply acquire it. Today, he sees essentially the same problem reappearing with artificial intelligence.

The lesson for leaders is simple: an AI adoption strategy cannot begin with AI licenses. It has to begin with work.

Why AI Adoption Breaks Before It Scales

The temptation is understandable.

A new AI capability appears. Competitors are experimenting. Employees want access. Leadership fears falling behind. So the organization deploys tools first and works out the operating model later.

That sequence is backwards.

McKinsey’s 2025 global AI survey found that 88% of respondents reported regular AI use in at least one business function, yet only about one-third said their organizations had begun scaling AI programs. Just 39% reported enterprise-level EBIT impact. Companies generating the greatest value were distinguished partly by something less glamorous than buying better models: they redesigned workflows.

O’Shea’s company learned that lesson firsthand.

After deploying AI broadly internally, the team discovered that it had not adequately defined who should use it, which workflows should change or how value should be measured. Then a data scan produced the more alarming discovery:

“We found that we had 33,000 sensitive data files that were overexposed and overshared.”

The company had effectively accelerated AI adoption before fully preparing the environment AI was about to explore.

Out of that experience emerged a practical five-part framework.

1. Start With the Work, Not the AI Tool

Before asking which AI platform your company should buy, ask a different question:

Which business outcome needs to improve?

Look for work where people spend significant time searching, summarizing, comparing, documenting, analyzing or repeatedly transferring information between systems.

Then establish a baseline.

For each candidate workflow, determine:

• time required today

• cost per completed task

• error or rework rate

• quality expectations

• cycle time

• business outcome affected

Only then ask what AI can change.

This seemingly obvious discipline is frequently skipped. Even Microsoft’s current AI adoption guidance begins with identifying business problems and use cases before selecting technology.

For a founder, that might mean shortening proposal preparation from four hours to one. For a professional-services firm, it could mean reducing research time while maintaining review standards. For an operations team, it might mean automating a repetitive workflow without increasing exceptions.

AI becomes measurable when the unit of transformation is work, not software.

2. Secure the Data Before AI Finds It

Traditional enterprise search often requires employees to know where information lives.

AI changes that relationship.

A capable assistant can search, connect and surface information far more aggressively. That is valuable—until the information was never supposed to be broadly accessible.

Old folders, forgotten permissions, draft employment documents, customer records, financial information and confidential projects can suddenly become discoverable through an interface that makes retrieval remarkably easy.

The question leaders should ask is uncomfortable but useful:

What would your AI discover today that your employees were never meant to find?

O’Shea argues for building security into the foundation:

“Secure by design, not an afterthought. It’s a forethought.”

That means reviewing permissions, classifying sensitive information, applying appropriate labels and access controls, and understanding which systems AI can reach before deployment expands.

This is not merely an IT concern. It is a prerequisite for scale.

3. Build AI Fluency Around Actual Roles

Giving everyone the same AI training is the organizational equivalent of giving everyone the same job description.

People need different capabilities.

A salesperson may need AI for account research and proposal preparation. A finance professional may need analysis and reconciliation skills. An executive may use it for scenario exploration. An operations specialist may increasingly work with automated workflows and agents.

O’Shea’s team discovered that employees who were comfortable using AI in a browser were far less capable when applying AI to spreadsheets, presentations and other real work environments. The answer was not another inspirational presentation about AI. It was practical training.

Leadership matters here as well. Microsoft reported in 2026 that workers whose managers actively modeled AI use reported greater AI value, critical thinking and trust in agentic AI.

The practical implication: train around jobs and workflows, not prompts.

4. Give Every AI Agent an Owner

Generative AI mostly helped employees produce something. Agentic AI increasingly allows software to do something.

An agent might access a CRM, query a database, send information to another application, update records, initiate a process or interact with several systems autonomously.

That changes the governance question.

An AI agent needs something surprisingly similar to an employee record:

Who owns it? What is its job? Which systems can it access? What permissions does it have? Who monitors it? What happens when its creator leaves? When should it be retired?

O’Shea warns of “orphaned agents”—automations continuing to operate after organizational ownership has disappeared.

The concern is becoming increasingly relevant. Deloitte’s 2026 enterprise AI research found that only one in five surveyed companies had a mature governance model for autonomous AI agents, even as agentic adoption was expected to rise sharply.

An agent without ownership is not simply technical debt. It can become operational authority without accountability.

5. Measure AI ROI at the Task Level First

Eventually, every AI initiative encounters the same question: What are we getting for the money?

O’Shea remembers his board asking whether AI should be expanded across the business or treated as an experiment that had run its course.

His problem was not that AI had produced no value. It was that the company could not objectively demonstrate it.

“I had no way of measuring how much of a return on investment we’re getting from AI.”

That forced the company to measure how individual tasks changed with AI—whether they became faster, cheaper or commercially more valuable.

That is a useful model for almost any organization.

Do not begin with a vague question such as, What is our enterprise AI ROI?

Start smaller:

What did this workflow cost before AI? What does it cost now?

Then include quality, rework, risk, revenue impact and capacity—not merely minutes saved.

Task-level evidence can accumulate into workflow ROI. Workflow ROI can accumulate into business-unit economics. Only then does enterprise AI ROI become something more than a presentation estimate.

The Next AI Problem: Manage the Data, Agents and Spend

Today, organizations debate which AI platform to adopt.

Tomorrow, the problem may be how to control the collection they have accumulated.

Employees already experiment across multiple systems. Agents will multiply. API calls, subscriptions, infrastructure and consumption-based pricing add additional layers of cost and complexity.

O’Shea reduces the emerging management challenge to three imperatives:

“Manage the data, manage the agents, manage the spend.”

That is a useful lens because AI sprawl will not look exactly like traditional software sprawl. An AI system may consume data from one platform, reason through another model, trigger an agent elsewhere and create costs across several services.

Governance therefore cannot simply mean creating an “approved tools” list. It increasingly requires visibility into what AI is doing inside the business.

How Do You Build an AI Adoption Strategy That Scales?

Begin with five questions:

1. Use cases: Which workflows are valuable enough to change?

2. Data: What information can AI reach, and should it?

3. People: Do employees know how to use AI effectively in their actual roles?

4. Agents: Who owns every autonomous workflow and its permissions?

5. ROI: Can you demonstrate the value created?

Notice what is missing from the beginning of that sequence:

Which AI should we buy?

Technology selection matters. But it comes after the organization understands what the technology is expected to accomplish.

What Should Every Leader Do Next?

Choose one important workflow this week. Not twenty.

Document how it works today. Measure its time, cost and quality. Identify the data involved. Define what AI may and may not access. Give the experiment an accountable owner. Train the people involved. Then measure what changed.

If the economics improve, expand deliberately. If they do not, learn before spending more.

The companies that win the AI race may not be those that adopt the most AI fastest. They may be the organizations that become exceptionally good at converting new technological capability into controlled, measurable business value.

Twenty years after a customer asked Denis O’Shea why anyone should keep buying technology people were not using, the question has become even more relevant.

Only now the stakes are higher.

Before buying more AI, ask your leadership team a harder question:

Are we becoming better at the work—or simply accumulating more technology?