Are You Solving the Wrong Problem With AI?
Can you honestly explain what the AI budget has bought lately?
If you had to pause for more than a second, you’re not alone. Walk into any large enterprise today and you’ll find a pattern that’s becoming painfully familiar. A dozen AI pilots, three chatbot proof-of-concepts, a computer-vision trial in one factory and a procurement team that asked ChatGPT to write a vendor email. And none of it changed the quarterly numbers.
The problem isn’t the technology.
The technology works. Model quality is double where it was in 2022 and open-source weights now humiliate what used to cost millions to run. You can stand up a useful copilot in a weekend.
So why does it so frequently feel like effort going nowhere?
Here’s the kicker: most companies treat AI as a new tool to bolt onto stable processes, when the real game is treating AI as the controlling idea of how the business makes money. Confuse those two things and you fund a year of impressive demos that never reach the P&L. A machine-learning engineer walks into a bar, asks for a drink, and the bar builds a demo chatbot instead of serving them. That pretty much sums up corporate AI today.
The Real Crisis Is Strategic, Not Technical
Let’s unpack the usual failure path.
Someone senior reads a report, gets nervous, sends a memo. An innovation office is created. Data scientists are hired. The enterprise gets an OpenAI contract, a great powerpoint, and six months later somebody says the words “We need to be more disciplined about AI,” which is business-speak for we spent money and feel nothing yet.
Then they purchase an AI platform “to organise everything.” Then they hire an AI consultant to organise the platform. And still the org chart hasn’t budged, the workflow is untouched, and the customer experience hasn’t changed by one percentage point.
This is AI theater. And it has a beautiful, seductive quality. It looks exactly like progress.
There are real applications running, after all. The AI ingestion summaries go out every Monday. The support bot deflects a few tickets. The code-completion tool makes developers a bit faster. Yet all of it sits in the corner of a process built for a pre-AI world. The throughput of the whole organisation remains about the same because no core business decision was re-examined.
Here is a clarifying litmus test for the executive team: does any part of your operating model change if someone flips the AI off tomorrow?
If the answer is “nothing changes, we’d just be slower,” what you actually bought was a helpful utility, not a strategy.
Strategy Means Asking What AI Makes Possible That the Model Could Not
Do you remember life before cloud computing? That’s not a rhetorical question. Before AWS, a business that needed a new server waited six weeks, paid for hardware depreciation and scaled in clumsy, predictable tiers. Cloud computing did not make the server slightly faster. It changed the unit economics of experimentation. Suddenly a startup could test an idea for ten dollars instead of ten thousand. That single change reordered industries.
AI is doing the same thing to areas you don’t yet see. It is collapsing the cost of cognitive labour: reading, drafting, predicting, recognizing, translating, prioritising. If you only use that cheap cognition to do the same work faster, you get a 10% efficiency bump. If you let it change how much cognition you can apply at every decision point, you get a different company.
A call centre that uses AI to give agents real-time suggested replies is doing the first thing. A health insurer that uses AI to scan every incoming claim for the subtle pattern of delayed diagnoses — not just fraud, but the clinical story that would be missed by any single human reviewer — is doing the second.
Neither project is inherently better. But only one changes the business model.
In 2017, John Deere bought a small startup called Blue River Technology for around $300M. At the time, people assumed it was about adding computer vision to tractors to make them fancier. In reality, Deere was repositioning itself from a farm machinery manufacturer into an agricultural intelligence company. Blue River’s See & Spray technology uses AI to detect the difference between a weed and a crop and then applies herbicide only where it’s needed. That’s not just an efficiency feature. It cuts chemical use, lowers input costs for farmers and has meaningful environmental credibility, all from one algorithm. John Deere went on to integrate this into its core product line, and the company has spent years framing itself around data, autonomy and outcomes rather than horsepower.
Deere didn’t add AI to its strategy. It rebuilt the strategy so that AI would be the only way to deliver it.
Most companies doing AI well in 2025 approached it the same way: start from a hard constraint or a persistent inefficiency and then ask what a ten-fold reduction in cognitive cost would do to that friction. Not how AI can give a report, but where the bottleneck is and what decision you would change today.
The Five Decisions That Separate Strategy From Pilot
You don’t need a 100-slide AI charter. The companies that get this right make a handful of stubborn, specific choices.
1. Pick the compounding use case, not the easiest one
Here is where most conversations collapse into enthusiasm. Every function can list fourteen AI ideas. Some of them will be easy wins. The mistake is believing easy wins translate into strategic value.
Strategy is about concentration. Where in your business would a 20% improvement in quality or cost create a spiral of other benefits? Not a point improvement — a spiral. That is almost always where the company has already accumulated data that is proprietary, messy and uniquely available to you.
Walmart doesn’t use AI to suggest shopping lists. It uses AI to optimise the route that replenishes inventory across 4,600 stores, because a small improvement in that system compounds across millions of daily transactions. Pick your compounding point and let other ideas sit on the shelf.
2. Don’t start with AI. Start with the decision.
A strategic frame answers the question “what decision, made better, would alter our trajectory?”
If your company is a freight broker, the premium decision is matching a shipment with a truck at a price that clears the market. If you’re a pharmaceutical company, it could be which drug candidate to advance. You would not believe how many “AI strategy” documents begin with the technology and never name a single decision that will be improved. That inverts priorities. Write the decision down. Describe who makes it today. Describe the information they lack. That paragraph is your strategy; the model is just execution.
3. Design the workflow around the machine, don’t drop the machine into the workflow
Throughput gains happen when human tasks vanish, not when a human becomes slightly faster at a task.
This is a difficult organisational question, because there is emotional attachment to the old flow. Someone built it. Roles depend on it. If the AI handles the first-pass review of a document, what does the junior associate do with the eight hours previously spent on that review? That job, not the AI, is the actual thing you need to manage. Name the new role. Decide where the redundant effort goes. Reconfigure the team around a human who now reviews the machine’s output under uncertainty, rather than performing the drudgery. If you avoid this step, the pilot stays isolated in a corner. It will not feel like strategy because you never let it touch the neck of the organisation.
4. Set boundaries around data fast, but avoid governance theatre
Not all data is created equal. Strategic AI is built on data assets that are hard to replicate. That should focus your internal data strategy on quality, not quantity.
But do not let governance become the velvet rope that keeps everything in the lobby. A risk-averse committee that requires a six-month review for every model update will kill any system’s usefulness, because models need frequent iteration to stay relevant near the business frontier. Get comfortable with three levels of AI governance: a fast lane for low-impact internal tools (translation, drafting, summarisation), a standard lane for customer-facing or regulated workflows requiring validation, and a slow lane for anything affecting employment decisions or safety. If your governance team treats every use case the same, their caution becomes the real constraint. Better to set those boundaries early with clear ownership than to ask a central AI team for permission every time.
5. Change the executive meeting metrics
You cannot manage what you refuse to quantify.
Traditional KPIs like “number of models deployed” or “data scientist headcount” are vanity metrics. Have the courage to ask less comfortable questions. What percentage of transactions passed through an AI-enhanced decision last quarter? Did that number increase from last quarter? What is the cost per successfully processed unit, and how did AI change it? This is less comfortable because it exposes accountability. Executives suddenly need to answer for value, not activity, and many would prefer to discuss the breathtaking sophistication of the prompt engineering. HBR has long argued that AI success comes from treating it as a management problem rather than a technology problem, and the governance of the meeting is part of that shift.
The Real-World Example That Shows the Whole Loop
Let’s come back to John Deere because it illustrates all five decisions in motion.
Deere’s strategic challenge was not “how do we use AI on a tractor?” It was the brutal economics of a dealership network selling an expensive machine used only part of the year. Farmers buy a tractor, and the dealership hopes for parts and service revenue later. But the operating profit for farmers is squeezed by input costs (seed, fertilizer) and thin margin.
The strategic decision to improve? Apply herbicide exactly where it is needed. That single decision touches the farmers input cost, chemical regulatory pressure and environmental brand story.
And the hard data requirement? Years of labelled images of weeds and crops from real fields, plus machine steering data. Nobody else had that distribution. Walmart’s compounding advantage is shelf-space data. Deere’s is ground-truth image data from acreage that you cannot fake in a lab.
The workflow redesign? See & Spray works in real-time while the machine crosses the field. Precision reduced herbicide use substantially in published farmer trials, putting a measurable dollar figure on what had once been vague promises of “smart agriculture.”
The vehicle for delivering it was a pivot: Deere began presenting itself as an intelligence company in its investor materials and made technology acquisition a recurring strategic habit. None of this worked instantly. There were organizational shifts, markdowns and cautious dealer adoption. But the strategic direction remained coherent. Deere did not merely ‘deploy AI’. It made the resolution of every field decision, from planting depth to harvest timing, progressively algorithmic.
Would Deere flip AI off tomorrow and experience only a minor slowdown? No, the whole go-to-market proposition would collapse. That is the definition of a true AI strategy: the strategy and the algorithms are inseparable.
What to Do on Monday Morning
If this feels abstract, make it concrete with one hour of exercise.
Call a meeting with your two best operators, the functional P&L owners, not the technologists. Write down the top three operational decisions that account for most of your annual spend or customer churn. For each one, answer three questions:
- Who makes it today, and how often?
- What information would make the decision 10x better?
- Could an AI system get that information faster and cheaper than a human?
You will immediately see that one of the three decisions has low data availability, one has high regulation and one sits in a sweet spot. Choose the sweet spot. Define the metric you will move within one quarter. Then go build the data pipeline for that decision and leave the wide-eyed exploration of generative AI toys alone.
A few practical notes on avoiding the swamp after you commit.
Don’t staff it with only data scientists. Staff it with the process owners, because they know why the decision happens in sequence and what the exceptions mean. If the data scientists are the only ones who can articulate the value, the operational team hasn’t bought in and the model will be orphaned.
Second, expect the first version to be mediocre. The real strategic work in AI is the feedback loop: the model is wrong, a human corrects it, the corrected value lands back in the dataset and the next iteration improves. If your org has no formal mechanism for that feedback, your model quality will sit at 80% forever. The gap between 80% and 95% is where the unit economics become magical, and crossing it requires human participation.
Third, budget for the organisational quit rate. There will be a quarter, usually in the first year, where the AI output appears worse than the previous manual process or costs more to operate. That is normal. What separates the companies that succeed is not patience with failure but a crisp causal narrative about why the current dip leads to future advantage. When leadership has that narrative, they tolerate the dip. Without it, they cancel the project in month nine and claim that AI was overhyped.
You can be forgiven for thinking we are late to this conversation. There is an entire industry of hype, and every vendor will sell you a confidence you probably shouldn’t buy. But the window for strategic advantage is real and it closes a little every quarter. The companies that started in 2023 have already collected three years of operational data that you haven’t, and their models will be better not because of technology but because of the accumulated human corrections you missed. Start now, or you will compete at a compounding disadvantage in your own market.
And listen: you don’t need a fancy central AI team to begin. You need a sharp conversation about which decision is the keystone of your profit model.
Are you solving the right problem with AI? If you can answer in one sentence, you’re already ahead of most leadership teams. If you can’t, go book that meeting with your best operators and leave the models for another day. Strategy first and always.