How to start transforming your company with AI

Executive summary: How to adopt AI without overcomplicating it

1. AI fails when organizations overcomplicate it AI initiatives fail when organizations overcomplicate before they learn. Avoid ambitious architectures, long evaluations, and hypothetical future needs. Learning beats certainty.

2. Start with people and reward the ones who lean in Give broad access to approved AI tools. Treat usage as enablement, not automation. Identify and empower employees who explore AI deeply; they are your early advantage.

3. Discovering use cases through bottom-up usage Bottom-up usage exposes high-frequency cognitive work, information bottlenecks, and inconsistent practices. These signals are more reliable than workshops or strategy exercises.

4. From tools to workflows Once patterns are clear, connect AI to real systems with small, replaceable applications. Narrow scope, clear ownership, minimal abstraction.

5. Choose vendors you can afford to be locked into Leverage large vendor ecosystems where lock-in is survivable. Be cautious with generic AI platforms that add limited value on top of existing systems.

6. Data foundations compound everything AI compounds with better data access. Prioritize read-only access and clear ownership over perfect data quality. Do not wait for ideal foundations.

7. Start with quick wins and place one deliberate bet Build momentum with obvious internal use cases. In parallel, run one higher-risk initiative to learn under different constraints. Kill weak initiatives fast.

8. Work in short cycles, not projects AI work is probabilistic and social. Requirements emerge through use. Every initiative should be killable within ninety days.

9. Do not over-engineer for today’s AI models Model capabilities are improving faster than organizational change. Avoid heavy infrastructure built to compensate for short-lived limitations.

10. Keep the organization ready to compound Avoid premature standardization and centralization. Let structure emerge from repeated success. Optimize for learning that makes the next initiative easier.

Bottom line AI advantage comes from starting early, staying simple, and preserving the ability to learn repeatedly without resetting.

Introduction: Why this matters now

Most companies already agree that AI matters. That is no longer the question.

The real problem is that most of them will still fail to get meaningful value from it. Not because the technology is immature, but because organizations respond to uncertainty by either moving too slowly or overcomplicating too early.

Large Language Models are not incremental automation. They change how knowledge work is prepared, executed, and reviewed. Searching, summarizing, drafting, and reasoning can now be done faster and at a higher baseline quality. Employees see this immediately. Boards ask about it. Executives want to invest. The value is no longer theoretical.

What remains unclear is how to act without locking the organization into fragile decisions.

This challenge is especially acute in non-tech companies. They are not built to absorb fast-moving technology, unstable best practices, or unclear tool ecosystems. In an attempt to reduce risk, many default to complexity or caution. Both slow learning.

This essay is about how to start transforming a company with AI in a way that is practical, low risk, and able to compound over time.

1. AI fails when organizations overcomplicate it

For non-tech companies, AI initiatives rarely fail because the technology is not ready. They fail because organizations start with approaches that are too complex for their current maturity, and because they try to reach certainty before gaining experience.

The pattern is consistent:

• Ambitious architectures designed upfront
• Long evaluation phases meant to reduce risk
• Heavy platforms selected before real use cases are understood
• Decisions driven by people who sound confident but lack hands-on AI experience

This behavior is understandable. AI is a new and fast-moving domain. Uncertainty feels uncomfortable, especially in organizations that value predictability. Complexity becomes a way to feel safe. In practice, it backfires.

Trying to be right before being informed leads to over-engineered solutions, slow progress, and fragile systems. Limited internal technical knowledge often makes this worse. Complexity becomes a substitute for understanding.

AI rewards a different approach. Simple systems, obvious applications, and fast feedback outperform elegant designs built in isolation. Quick wins are not a lack of ambition. They are how experience is built. Complexity should be earned through use, not assumed upfront.

What actually breaks AI initiatives in non-tech companies:

• Starting with complex architectures instead of simple use cases
• Designing for hypothetical future needs
• Treating AI as a planning exercise instead of a learning process
• Optimizing for certainty instead of feedback

Rule: If your first AI initiative is designed to feel certain rather than to generate learning, it will be too slow, too complex, and too fragile to matter.

2. Start with people and reward the ones who lean in

Employees are already ahead of the organization. Many are experimenting with AI tools in their daily work, often without guidance, shared standards, or visibility. The fastest way to start is to acknowledge this reality and channel it deliberately.

Give broad access to a small set of approved tools such as ChatGPT or Copilot. For most non-tech companies, this is the highest ROI step available. Not because it automates workflows, but because it improves how people think, write, analyze, and prepare work. The gains show up immediately in clarity, speed, and quality of output.

This step is about enablement, not automation.

Access alone, however, is not enough. Some employees will go much further than others. They will test limits, develop repeatable workflows, and discover where AI meaningfully changes how work is done. These people are disproportionately valuable. They are not just more productive; they surface opportunities the organization would not have identified top-down.

They should be actively empowered:

• Recognize their effort
• Give them time and visibility
• Involve them in shaping next steps
• Treat them as part of the journey, not as shadow users

Adoption follows incentives. If using AI feels risky or invisible, people will stay quiet. If it is encouraged and rewarded, learning compounds quickly.

At this stage, guardrails should be light:

• Approved tools only
• Clear rules on what data must not be uploaded
• Internal use first

The goal is not control. It is to create a visible learning loop.

This only works with explicit executive cover. Bottom-up learning is uncomfortable in organizations optimized for control. Without clear permission, people self-censor and experimentation stalls. Enabling AI adoption is not neutral: leaders must protect early experimentation and accept that some initiatives will be visibly killed. Without that protection, the rest of this approach remains sound in theory but blocked in practice.

3. Discovering use cases through bottom-up usage

Bottom-up enablement is not just about productivity gains. It is a discovery mechanism.

When employees use AI freely, patterns emerge:

• Repeated prompts across teams
• Friction in accessing information
• Manual reasoning steps that appear again and again
• Work that is critical but poorly documented

These signals are difficult to surface through workshops or strategy exercises. They emerge naturally from behavior.

Bottom-up discovery is especially effective at identifying:

• High-frequency cognitive tasks
• Information synthesis problems
• Decision preparation work
• Inconsistencies in how similar tasks are performed

This provides grounded input for deciding what to build next.

At the same time, not all valuable use cases will emerge this way. Anything that requires shared data access, system integration, or automated actions needs deliberate design. Giving someone access to a chatbot will not produce a system that reads internal data or takes operational actions.

That is expected. Bottom-up discovery is not meant to surface every opportunity. It is meant to prevent you from building the wrong things too early.

4. From tools to workflows

Once usage patterns are clear, the next step is unavoidable. AI must interact with real data and real systems.

This is where many non-tech companies hesitate or overcorrect. Either they delay integration indefinitely, or they jump straight to large platforms and complex architectures.

In many cases, lightweight workflow or agent platforms that sit close to existing systems are enough. Tools embedded in ecosystems like Microsoft, Salesforce, or similar platforms can support narrow, well-defined workflows with relatively low friction. When they match the use case, they are often the fastest and safest way to move forward.

Problems arise when these tools are stretched beyond what they are designed for, or when their constraints start to shape the workflow more than the business need.

When that happens, custom software becomes a viable alternative earlier than most organizations expect.

With AI-assisted programming, building small, focused software applications is faster and cheaper than it has ever been. What once required large teams and long timelines can now be done quickly, as long as scope stays narrow.

Default position: Build the smallest possible workflow that connects AI to existing systems and can be replaced later.

This can take different forms. What matters is not the tool choice, but the shape of the solution:

• Narrow, well-defined use cases
• Clear operational ownership
• Minimal abstraction

Custom solutions make sense when:

• There is a clear owner responsible for outcomes
• The application is a thin layer, not an internal platform
• Replacing it in six months would be acceptable

Anti-pattern: Designing a generalized AI architecture before you understand where the value is.

At this stage, the goal is learning at the system level, not optimization.

5. Choose vendors you can afford to be locked into

Most non-tech companies are already deeply dependent on a small number of vendors. Adding AI capabilities on top of those ecosystems can be a reasonable trade-off, especially early on. Platforms like Microsoft, Salesforce, or Google control distribution, will track model evolution closely, and are unlikely to disappear.

This is very different from building core workflows on top of agent platforms from startups or consultancies. These platforms often add limited value compared to AI capabilities embedded in established vendor ecosystems, especially when they position themselves as generic solutions.

There are exceptions. Specialized tools that are clearly bounded and easy to replace can make sense. With AI, switching costs are often lower than expected. Avoid lock-in where it is unnecessary.

6. Data foundations compound everything

AI is only as useful as the information it can reliably access. This does not mean you need perfect data to start, but it does mean that well-organized data compounds every AI investment you make.

Companies with clear sources of truth, consistent structure, documented ownership, and accessible data move faster and get more value from the same models.

Early on, prioritize access over cleanliness. Read-only integrations are usually enough. Narrow scopes reduce political friction. Waiting for ideal data quality is a reliable way to stall progress.

It is also worth stating explicitly: improving data organization pays off far beyond AI. It improves reporting, decision-making, and operational clarity even without advanced models.

AI simply makes the benefits unavoidable.

Anti-pattern: Treating data work as a prerequisite instead of a parallel effort.

7. Start with quick wins and place one deliberate bet

Early momentum matters.

The best early use cases tend to be obvious:

• Extracting structured data from unstructured documents
• Summarizing internal knowledge
• Supporting decisions with better information access
• Reducing time spent on preparation and review

These are rarely glamorous, but they are measurable and repeatable.

In parallel, it can be valuable to place one bigger, riskier bet. Often customer-facing. Often with clear upside. The goal is not guaranteed success, but learning under different constraints.

Measurement philosophy:

• Directional metrics are enough
• Time saved, volume handled, cycle time reduced
• Avoid vanity dashboards

Important discipline: Kill initiatives aggressively.

AI progress comes from accumulated learning, not from protecting sunk costs.

8. Work in short cycles, not projects

AI work does not behave like classic software projects.

Requirements are not fully knowable upfront. Output quality is probabilistic and contextual. What is useful only becomes clear once people interact with the system in real work.

The way AI interacts with your data is also unpredictable. Issues with context, semantics, and assumptions surface only through use.

Finally, value is discovered socially. When AI is introduced, people change how they work. Processes evolve. New workflows appear. Classic project planning assumes stability. AI breaks that assumption.

These characteristics make long, rigid plans fragile.

Short cycles are the correct response. They allow teams to test assumptions, observe behavior, and adjust direction before mistakes compound.

Working in short cycles is not a lack of ambition. Large outcomes are built by stacking validated steps, not by committing to fragile plans upfront.

Rule: Every AI initiative must be designed so that its value can be evaluated through real use within ninety days.

9. Do not over-engineer for today’s AI models

AI models are improving faster than most organizations can adapt. Designing heavy systems to compensate for current limitations is rarely a good investment.

For internal tools, imperfections are often acceptable. For external systems, guardrails matter, but heavy infrastructure designed around short-lived problems usually does not.

The goal is not to perfect systems for today’s models. It is to stay positioned so that better models translate quickly into better outcomes.

Anti-pattern: Spending months engineering around limitations that will soon disappear.

10. Keep the organization ready to compound

Long-term AI advantage does not come from early sophistication. It comes from the ability to learn repeatedly without resetting.

What breaks this ability is premature institutionalization. This does not mean that structure, standardization, or platforms are wrong. It means they should be introduced only after experience makes their role clear.

In practice, premature institutionalization shows up in familiar ways:

• Standardizing tools before patterns exist
• Centralizing control instead of enabling teams
• Freezing architectures and roadmaps too early
• Treating early decisions as permanent

Over time, repeated initiatives reveal what is stable and what is not. Only then does it make sense to introduce shared components or additional structure.

Real progress happens when each initiative leaves behind more than working systems. It leaves behind reusable understanding: what worked, what failed, where the limits are, and why certain choices were made.

To preserve this:

• Keep ownership close to the teams using the systems
• Let central functions enable access and guidance rather than dictate solutions
• Allow standards to emerge from repetition and reuse, not policy

Early on, simplicity beats sophistication. Learning beats optimization. Flexibility beats efficiency. Organizations that protect their ability to build on what they learn benefit disproportionately as AI capabilities continue to evolve.

Conclusion

AI transformation is not a one-time initiative. It is a capability that improves as organizations learn through real use instead of committing to plans too early.

For non-tech companies, the goal is not to be early or sophisticated. It is to be practical, deliberate, and adaptable. Start with simple tools, learn from real use, build only what earns its place, and avoid choices that are hard to reverse.

The companies that benefit most from AI will be the ones that keep their ability to learn, change direction, and stay positioned to take advantage of continued improvements in AI capabilities.