Why AI Strategy Feels Like the World’s Hardest Escape Room

ReadFun Yet InformativeStats & TrendsBest Practices
Written By Jeff Winter

I love escape rooms! I love the chaos and panic. I love the false confidence. I love the moment when someone finds a key, holds it in the air like they just discovered fire, and then realizes nobody has any idea what lock it opens. 🤣

There is always a point, usually about ten minutes in, when the room changes. Everyone walked in confident. Everyone assumed the clues would be obvious. Everyone thought they would be the calm, rational genius who sees what others miss. Then the clock starts ticking, someone is crawling under a table, someone else is yelling random numbers from a bookcase, and one very confident person is absolutely convinced the answer is hidden in a painting that is clearly just decoration. That is why I love escape rooms. They are rarely difficult because each individual clue is impossible. They are difficult because the clues are scattered, the sequence matters, communication breaks down, assumptions pile up, and smart people start solving different versions of the problem.

Which feels a lot like AI strategy right now. Many companies have entered what I would call The AI Strategy Escape Room. Some know they are in it. Some are pretending they are not. Some walked in because the board asked about AI. Some walked in because a competitor announced something. Some walked in because a technology vendor showed a demo that made everyone briefly forget how complicated their business actually is.

The good news? AI is awesome.

The bad news? Virtually no one has escaped so far.

That may sound dramatic, but the data keeps pointing in the same direction. McKinsey’s 2025 workplace AI research found that 92% of companies plan to increase AI investments over the next three years, but only 1% of leaders describe their companies as mature in AI deployment. McKinsey’s 2025 global AI survey also found that 88% of organizations are using AI in at least one business function, yet only about one-third have begun scaling AI across the enterprise. S&P Global found that the share of companies abandoning most AI initiatives before production jumped from 17% to 42% year over year, with organizations scrapping an average of 46% of proof-of-concept projects before production. Gartner found that at least 50% of GenAI projects were abandoned after proof of concept by the end of last year because of poor data quality, inadequate risk controls, escalating costs, or unclear business value. MIT NANDA’s 2025 GenAI Divide report found that 95% of organizations were getting zero return from GenAI efforts, while only 5% of integrated AI pilots were extracting millions in value.

So yes, companies are in the room. They are just not finding the exit. And I think the reason is simple: AI strategy is not one puzzle. It is eight puzzles stacked on top of each other.

1. Match the Use Case

This is the wall covered with shiny AI use-case cards. One card actually matches the business problem. The rest are distractions with very impressive names. I see this one so often. AI has been one of the only times in my career where I have seen budgets assigned to a technology before the business problem was clear. Not always, of course. There are plenty of serious companies doing this well. But I have seen enough organizations allocate arbitrary amounts of money to “AI” with an objective that essentially boils down to: go figure out how to apply AI to the business. That is backwards.

Imagine walking into an escape room and being told, “Your objective is to use this key.” Great. Which key? What lock? In what sequence? To open what door? That is exactly how many AI efforts begin. The objective is not “reduce unplanned downtime by 15%,” “cut quoting cycle time in half,” “improve first-pass yield,” “reduce engineering search time,” or “increase service response accuracy.” The objective is simply “use AI.” I believe one sentence explains a lot of stalled AI projects: Vague business priorities will lead to vague AI projects. When the business priority is fuzzy, the AI project becomes fuzzy. Teams start comparing models before clarifying workflows. They fund pilots before agreeing on success. They debate tools before defining outcomes. Then everyone is shocked when the project produces something interesting but not especially useful.

To solve this puzzle, force every AI idea through a use-case card before funding it. The card should answer: What business problem are we solving? Who is the user? What decision, action, or workflow are we improving? What is the current baseline? What metric must change? What data is required? Who owns the outcome? Then ask the most uncomfortable question: if AI were not allowed, would this still be a business priority? If the answer is no, you probably do not have an AI strategy. You have an AI hobby.

2. Crack the Data Lock

This is the combination lock that looks simple until you realize the code requires access, quality, context, ownership, and trust, in exactly the right order. Everyone says they have data. That is not the puzzle. The puzzle is whether the data is usable for the business outcome you care about. I have seen this play out many times. The team finds a great use case. Everyone gets excited. The business value is real. The demo makes sense. Then someone asks where the data lives, and suddenly the room gets quiet.

Some data is in the ERP. Some is in the MES. Some is in a spreadsheet. Some is in a historian. Some is in a supplier portal. Some is in emails. Some is technically available but owned by a different function. Some is accessible but not trusted. Some is trusted but not current. Some is current but missing the context required to make it meaningful. The lock has numbers on it, but nobody agrees what the numbers mean. This is why so many AI strategies turn into data strategies wearing a clever disguise. Gartner predicted that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data. Confluent’s 2026 Data Streaming Report found that 72% of global IT leaders say a lack of real-time data infrastructure is stalling their efforts to scale AI. Source: Confluent.

To solve this puzzle, do not start with “get all the data.” That is how teams disappear into the basement for three years. Start with the minimum viable data product for one priority use case. Define the exact data needed, source systems, data owner, access rules, quality thresholds, refresh cadence, lineage, and required business context. The goal is not perfect enterprise data. The goal is trusted data for a specific business outcome.

3. Untangle the Legacy Systems

This is the box of tangled wires in the corner. Inside the box are ERP, MES, CRM, PLM, spreadsheets, middleware, custom databases, dashboards, old integrations, new integrations, manual workarounds, and one mysterious system everyone refers to by a nickname.

The objective sounds simple: untangle the wires. The problem is that some wires are obsolete, some are duplicated, some are mislabeled, some go nowhere, and some are quietly holding the entire business together through a script written by someone who left the company in 2017. Pull the wrong one and the lights go out. This is one of the biggest reasons AI projects stall between pilot and production. In a pilot, the AI solution can operate in a controlled environment. In production, it has to live inside the real business. That means security rules, integration patterns, exception handling, workflow handoffs, master data issues, technical debt, process variation, and operational constraints.

To solve this puzzle, trace the workflow before designing the solution. Pick one business outcome and follow it end to end. Where does the work begin? Which systems are touched? What decisions are made? What exceptions occur? Who approves what? Where does data get re-entered? Where do people leave the system and use Excel because the process technically works but spiritually gave up years ago? Then define the thinnest valuable integration path. Do not integrate everything. Find the smallest integration path that allows the AI output to enter the real flow of work and create measurable value.

4. Unlock the Approval Chain

This is the puzzle where Legal, Security, IT, Procurement, Finance, and leadership all have to turn their keys at the same time. And, naturally, one key is missing.

This puzzle is real because AI creates legitimate concerns. Data privacy matters. Cybersecurity matters. IP protection matters. Regulatory compliance matters. Vendor risk matters. Cost matters. Model accuracy matters. Accountability matters. The mistake is treating these groups like blockers after the fact. Too often, teams build the pilot, fall in love with the output, and only then bring in Legal, Security, Procurement, IT, and Finance. At that point, every review feels like a delay. Every question feels like resistance. Every control feels like bureaucracy. But many of those teams are not trying to stop AI. They are trying to prevent the company from walking through a door labeled “innovation” that opens directly into a liability conference room.

To solve this puzzle, build a risk-tiered AI intake process. Low-risk use cases should not go through the same review path as high-risk use cases involving sensitive data, regulated decisions, customer-facing outputs, safety implications, or autonomous action. Create clear tiers. Define decision rights. Establish review criteria. Build pre-approved patterns for common AI use cases. Agree on service-level expectations. Good governance should not be the locked door. Good governance should be the map that keeps people from trying to escape through the ceiling.

5. Navigate the Procurement Maze

This is the maze filled with intake forms, vendor reviews, preferred supplier lists, budget approvals, security questionnaires, legal redlines, pricing models, architecture promises, pilot agreements, and one mysterious portal that somehow makes everything take longer.

This puzzle is hard because AI buying is not like buying a normal piece of software. In many cases, the company is not just purchasing a tool. It is making decisions about data access, model risk, workflow integration, intellectual property, cybersecurity, regulatory exposure, operating cost, vendor dependency, and long-term scalability. That means Procurement is not the only player in the maze. Legal, Security, IT, Architecture, Finance, Data Governance, Compliance, and the business owner are all standing somewhere in the hallway holding part of the map. And this is where things get messy. The business wants speed. Procurement wants process. Security wants assurance. Legal wants protection. IT wants architecture fit. Finance wants economic clarity. The vendor wants a signature. The executive sponsor wants momentum. Meanwhile, the project team is trapped between “we need to move fast” and “we cannot accidentally create a data, risk, or cost problem that follows us around for the next five years.”

The AI vendor landscape makes this even harder. Every vendor has AI now. Every product has a copilot. Every roadmap has agents. Every platform claims it can transform the business. Some of those capabilities are genuinely impressive. Some are immature. Some overlap with tools the company already owns. Some solve a narrow problem very well. Some create a new platform dependency before anyone has agreed whether the use case is even worth scaling. This is why the procurement maze is so dangerous. A company can make a purchase that feels fast in the moment but creates complexity later. It can approve a pilot that cannot pass security for production. It can buy a tool that works beautifully in isolation but does not fit the workflow. It can sign a contract without understanding data rights, model training boundaries, usage-based pricing, escalation costs, or exit options. It can also spend six months reviewing a small, low-risk use case as if it were a mission-critical autonomous decision system.

Both extremes are bad. Too little process creates risk. Too much process kills momentum. The real challenge is designing a procurement path that matches the risk, scale, and strategic importance of the AI initiative. A low-risk internal productivity tool should not move through the same maze as an AI system that touches customer data, regulated decisions, safety-critical operations, or autonomous actions. The company needs different lanes, not one giant approval tunnel where every AI idea goes to slowly lose the will to live.

To solve this puzzle, create an AI procurement pathway that starts before the vendor shortlist. First, define the use case, business owner, success metric, data requirements, risk tier, integration needs, and production-readiness criteria. Then evaluate vendors against that specific context instead of comparing generic feature lists. The vendor scorecard should include workflow fit, data access requirements, data usage rights, model governance, cybersecurity posture, integration effort, deployment model, scalability, implementation support, commercial structure, pricing triggers, auditability, human oversight, and exit options. Ask practical questions early: Will this vendor train on our data? Where does the data go? Can we restrict usage? What happens when volume increases? Who owns the outputs? How does the solution integrate into the workflow? What controls exist for hallucinations or incorrect recommendations? What happens if we stop using the product? Most importantly, define the production gate before the pilot starts. What must be true for this to move from experiment to scaled deployment? What business metric must improve? What risk must be resolved? What integration must exist? What adoption level is required? What cost model must be proven? Who has authority to approve the next stage?

6. Fit the Adoption Pieces

This is the table with process, training, incentives, ownership, workflow, trust, and behavior-change pieces scattered everywhere. The catch is that every piece is a different shape, and none of them naturally connect. This may be the most underestimated puzzle in the room. Companies often treat adoption as a communications exercise. Send the email. Host the webinar. Record the training. Add the link to the intranet. Announce that everyone should now use the new tool. Then everyone goes back to the old way of working. Not because they are bad employees. Because the old way is embedded in the workflow, incentives, metrics, approvals, habits, and management system. Adoption is not awareness. Adoption is behavior change, and behavior change requires changing the actual system of work.

To solve this puzzle, do not ask, “Did we train people?” Ask, “What behavior must change on Monday morning?” Then get specific. Which role changes? Which task changes? Which decision changes? Which approval changes? Which metric changes? Which manager must reinforce the new behavior? What happens when the AI recommendation conflicts with experience? How will users provide feedback? How will the solution improve over time? If the AI tool is added on top of the old process, it becomes optional. If it is embedded into the redesigned workflow, it becomes how work gets done.

7. Set the ROI Dial

This is the dial that only opens when the promised value matches something measurable: cost reduction, revenue growth, cycle time, quality, risk reduction, productivity, or customer experience.

This puzzle is harder than it looks because AI value often feels obvious before it is measurable. The demo is amazing. The prototype is slick. The early users are excited. The executive sponsor is enthusiastic. The vendor says “transformative” seven times. But Finance still wants to know what changed, and Finance is right.

One of the reasons AI gets stuck is that teams fail to define the economic logic early enough. They confuse activity with impact. They measure usage but not value. They count pilots but not production outcomes. They celebrate ideas but do not baseline the process they are trying to improve.

BCG’s 2025 AI Radar found that only about one-quarter of executives say their companies have created significant value from AI initiatives, and that 60% of companies are failing to define and monitor financial KPIs related to AI value creation. Later 2025 BCG research found that only 5% of companies were achieving AI value at scale, while 60% were not achieving material value at all despite substantial investment.

To solve this puzzle, involve Finance before the pilot. Define the baseline. Define the value lever. Define what will count and what will not count. Is this hard savings, soft savings, avoided cost, revenue lift, margin improvement, working capital impact, risk reduction, customer retention, faster cycle time, or improved quality? Then define the proof method. Will you compare before and after? Use a control group? Track adoption-adjusted value? Measure time saved and translate only a portion into financial impact? Track error reduction? Tie output to throughput? The CFO does not need every AI project to become a perfect spreadsheet, but the organization does need to know whether it is moving the business or just entertaining the room.

8. Align the Stakeholder Arrows

This is the puzzle where every stakeholder arrow has to point at the same outcome instead of five different finish lines. This sounds simple until you ask five functions what the AI project is for. Operations may want efficiency. IT may want architecture control. Security may want risk reduction. Finance may want measurable return. Legal may want defensibility. Sales may want customer responsiveness. Service may want faster answers. Executives may want transformation. Users may want the tool to stop making their day harder. Everyone is supportive, but everyone is also pointing somewhere slightly different. That is how AI projects spin in place. I have learned that alignment is not agreement in a meeting. Agreement in a meeting is easy. People nod. The slide looks good. The phrase “cross-functional” appears. Everyone says they are aligned. Then they leave the room and optimize for their own incentives.

To solve this puzzle, create an outcome contract for every major AI initiative. It does not need to be complicated. It should define the business problem, target outcome, accountable owner, decision rights, required data, funding, risk tolerance, workflow impact, success metrics, timeline, and what the organization is explicitly not doing. That last part matters. Strategy is not just deciding what to pursue. It is deciding what distractions to ignore.

When the stakeholder arrows finally point in the same direction, everything changes. Decisions get faster. Tradeoffs get clearer. Governance becomes more practical. Adoption becomes more intentional. Investment becomes easier to justify. And, finally, the door starts to move.

Finding the Exit

The best teams will escape. Not because they have the fanciest tools, the biggest AI budget, the most pilots, the most copilots, or the highest concentration of people saying “agentic” in conference rooms.

They will escape because they slow down long enough to solve the right puzzles, in the right order, with the right people in the room. Wild concept, I know.

AI strategy is hard because companies are complicated. The clues are in different departments. The locks are attached to workflows. The keys are owned by committees. The wires run through legacy systems. The ROI dial is stuck between “cool demo” and “actual value.” And every time someone gets close to the exit, another stakeholder arrow points in a new direction. But the exit exists. I really believe that. The companies that escape will not be the ones shouting “AI” the loudest at the door. They will be the ones disciplined enough to match the use case, fix the data, untangle the systems, unlock the approvals, navigate the vendors, fit the adoption pieces, set the ROI dial, and align the stakeholder arrows. That is the real AI strategy escape room. And for many companies, the clock is already ticking.

References:

Jeff Winter
Previous

What Matters Most When Scaling AI in 2026
Next

Stop Blaming the Wave (Copy)