How do you create value with AI?

The gap is real, and it isn't the tech

Most large companies have done the visible work. They funded pilots, trained staff, upgraded data platforms, and tried the newest models. Then the CFO asks what it returned, and nobody has a clean answer.

A new Harvard Business Review piece by David L. Rogers and Krishnan Sankaranarayanan (who leads AI transformation at Eastman) calls this the AI value gap. They cite a PwC global study in which most CEOs reported no measurable benefit from AI yet. Their diagnosis matches what we see in the field: the problem is strategy, not technology or people.

We hear the same thing in marketing every week: “We can see the spend but not what it paid for.” The fix is the same too.

Where AI money goes to die

Low-value AI work tends to look like this:

• Fixing friction leaders can see but that doesn't cost the business much
• Shipping a feature customers don't care about
• Copying a use case because a competitor or vendor already did it
• Building new dashboards that don't change a single decision
• Giving people time-saving tools with no plan to capture the time saved

Each of these gives a small benefit at best. None of them builds a competitive advantage. They also feel productive, and that's why they survive.

The equation: one zero kills the product

• Technology: models, infrastructure, compute, and data.
• Adoption: whether people change how they work.
• Strategic alignment: how closely the problem you picked maps to the balance sheet and the income statement.

The factors multiply, so any one of them at zero makes the whole thing zero. A brilliant model with the wrong data returns nothing. So does an elegant tool nobody uses, or a powerful system pointed at a problem that doesn't matter.

Most companies have learned to manage the first two factors. Alignment is the one they skip, and it gives the most leverage. It usually slips for one of four reasons: AI belongs to IT or an innovation team with no P&L, leaders hope a big enough model will spare them hard choices, AI gets treated as a skill to roll out, or fear of falling behind competitors drives the work.

The one-minute test

This example from the article is the one we'd put on every steering committee wall. Say an AI tool saves a worker one minute on a routine task.

In one business, that minute is worth nothing. It goes back into slack, the task wasn't on the critical path, and nothing downstream changes. In another business, the same minute is huge. The task gates throughput, like fitting one more pallet on a truck. Or it runs tens of thousands of times a day. Or it cuts a minute of exposure to a hazardous environment.

The improvement is identical, but the value is completely different. That's constraint thinking, and it's how we run every Digital Optimus engagement: instrument first, find the constraint, fix that before you buy more of anything. For Future Home Loans, the constraint was page speed, not ad budget. Cutting load time from 5.2s to 2.9s doubled the lead rate in the first week. AI in a plant or a supply network follows the same logic. Improvements away from the bottleneck mostly don't show up in the numbers.

Put your money where your value is: four steps

The article's playbook, in the order we'd run it:

1. 1

   Name the value driver

   A value driver is a lever that changes profitability when you move it. In heavy industry, procurement can be 30 to 60 percent of revenue. HBR reports that AI-assisted negotiation has cut that spend by 10 to 20 percent. Asset uptime is another. Pick the lever before you pick the model.

2. 2

   Know where you sit in the value chain

   Two chemical makers can need opposite things from AI. A commodity producer running capital-heavy plants wins on cost per unit. A specialty producer wins on R&D speed and how tightly it plugs into customer pipelines. The same use case can be worth millions to one and nothing to the other.

3. 3

   Read the income statement line by line

   Go through revenue, COGS, and operating expense, and for each line ask what AI could move and by how much. That list is your hypothesis backlog. Also look for new revenue from capital you've already spent.

4. 4

   Define the metric, and who signs off, before you build

   If you can't say how you'll measure the project and who will sign off on the value, you're not ready to start. This is the rule we apply on every engagement, written down on day one.

We add one rule of our own. Order the list by return, then prove it in 90 days. Pick the cheapest move against the biggest constraint. Measure it against the baseline you already agreed on. Fund the next project from what the last one returned.

What this looks like in a supply chain

Here's a starting map for operations teams across Plan, Source, Make, and Deliver. Use it to seed the hypothesis backlog from step three, then replace our examples with your own numbers.

What anchored AI returns

The companies HBR highlights share one habit. Each aimed AI at something central to how they make money, and measured that thing.

Figures as reported in the HBR article linked below.

The larger payoff is trust. When every project maps to a value driver and a P&L line from day one, the people who own budgets start to believe in AI, and the rest of the organization follows them. It also changes what kind of advantage you get. A generic use case from a vendor is available to your competitors too. AI built on a deep understanding of your own economics isn't. As the authors put it, your strategy is your moat, and AI is what puts that strategy to work.

Where Hoshin planning fits

Hoshin kanri exists to solve this exact problem: making sure the work people do connects to the few breakthrough objectives that matter. The AI portfolio needs the same discipline. Set objectives at the top, pass them up and down through catchball until value drivers are agreed, and hold a short list of projects that each trace to a number.

We built the AI Hoshin intake to do this. Each use case is captured against a supply chain area, a value driver, and a baseline before anyone writes code.