The AI Adaptability Accelerator: Using AI to Compress Learning Cycles

Seventy-eight percent of organizations now use AI in at least one function. Almost everyone is in. And yet only a sliver, around five percent, qualify as high performers, seeing real bottom-line impact. That number comes from McKinsey's 2025 State of AI survey, their annual read on how thousands of companies are actually putting the technology to work. Two-thirds of those companies are stuck in what the report calls pilot purgatory: experiments that never graduate into anything that matters.

That gap is the whole story. The problem isn't access. Your competitors have the same models you do. The gap is between having a capability and actually getting value out of it. Economists Wesley Cohen and Daniel Levinthal gave that gap a name in a landmark 1990 paper: absorptive capacity. They were trying to explain why some firms turn new knowledge into advantage while others sit on the exact same information and do nothing with it. Their answer: it isn't access that separates the winners. It's the capacity to absorb what's available and turn it into something you actually do differently. The bottleneck was never the technology.

Which raises the real question. Most teams point AI at execution. Write the email faster. Generate the code faster. Summarize the deck faster. All useful. All real. And all of it misses the thing that actually compounds.

The teams pulling ahead aren't using AI to do their existing work faster. They're using it to learn faster. There's a difference, and it's the difference between a tool and an advantage.

The speed of doing is not the speed of learning

Most teams fall into the same trap. You drop AI into a workflow, every task gets quicker, and it feels like progress. But faster tasks don't automatically make a faster organization.

Every organization runs on a learning loop, whether anyone calls it that or not. You sense a signal from the market or a customer. You make sense of what it means. You decide what to do, try something, then read the result and figure out what it taught you. Then you go around again, a little smarter than last time. That cycle is what I mean by the loop. If only one step in it got faster, the loop itself barely moved.

John Boyd, the fighter pilot who built his career around this idea, called it the OODA loop: observe, orient, decide, act. His whole argument was that you don't win by being bigger, or even by being faster in a straight line. You win by cycling through that loop faster than the other guy. Get inside their decision cycle and they end up reacting to a world that's already changed.

Adaptability isn't a personality trait or a poster in the break room. It's the speed and quality of your learning loop. David Teece, the Berkeley economist who has spent decades studying why some firms outlast disruption, gave that idea a name: dynamic capabilities. His research lands on the same point Boyd did. The firms that survive are the ones that can sense, seize, and reconfigure faster than their environment shifts.

So the real question isn't "how do we use AI to work faster?" It's "where in our learning loop is AI actually tightening the cycle, and where is it just making one task quicker while the loop stays the same length?"

The trap inside the trap: faster and dumber

This is where many teams can quietly hurt themselves.

Point AI at execution carelessly, and you don't just fail to learn faster. You can learn slower. The work still comes out, maybe quicker than before, but the thinking that used to happen while you did the work doesn't happen anymore. Researchers have started calling the output "workslop" or "thinkslop": plausible, polished, and hollow. The mental muscle that used to get worked by wrestling with the problem goes slack. You've outsourced the doing, and the learning went with it.

Think of AI as a mirror, not a genie. It reflects and amplifies whatever judgment you bring to it. Bring sharp questions and hard-won context, and it makes you faster at getting smarter. Bring none, and it makes you faster at producing confident nonsense.

And there's research behind this, not just a hunch. When Anthropic studied how people actually use AI for real knowledge work, the people who got the most out of it weren't the most fluent with the tool. They were the ones with the deepest domain expertise. Humans were still making roughly 70 percent of the "what should we do" decisions. AI was handling about 80 percent of the "how do we execute" work. The judgment stayed human. The payoff came from pairing that judgment with cheap, fast execution.

That's the whole game in one sentence. Point AI at the what (sensing, deciding, learning) and it compounds your judgment. Point it only at the how, and you get speed with a slow leak in your capability.

What the evidence actually shows

When you look at the strongest research, the interesting finding isn't speed. It's that AI changes the quality and breadth of the thinking, not just the clock.

Procter & Gamble ran a real experiment on this with Harvard and a few other researchers. They took 776 of their own professionals and randomly assigned them to work on actual product-development problems, either with AI or without, alone or in teams. They published it as "The Cybernetic Teammate," and the results are worth noting.

Yes, people moved faster. Individuals finished about 16 percent quicker. But that's the least interesting part. An individual working with AI matched the performance of a two-person team working without it. The AI didn't just speed up one person. It gave them the second perspective a teammate would have brought.

It got better. Without AI, the R&D people pitched technical solutions and the commercial people pitched commercial ones. Everybody stayed in their lane. With AI, both groups produced balanced solutions that crossed the boundary. The technology broke down the functional silos that usually slow an organization's learning to a crawl. And teams using AI were three times more likely to land in the top ten percent of solutions. The genuinely good ideas, not just more ideas.

Read that again. The win wasn't volume. It was better thinking, faster, across more perspectives. That's a learning loop getting tighter, not just a task getting quicker.

You see the same pattern in how fast teams can now run experiments. In pharma and physical-product work, AI has cut prototype cycles by up to 70 percent. In software, a process that used to take weeks gets done in an afternoon. When experiments get that cheap and that fast, you can run ten where you used to run one. That means you learn ten things where you used to learn one. The constraint stops being "can we afford to test this?" and becomes "are we actually paying attention to what the tests tell us?"

The part everyone skips: learning capture

Which brings me to the step almost nobody pays attention to.

Sensing, deciding, experimenting. Those are the visible, exciting parts of the loop. The unglamorous one is capture: actually pulling the insight out of what just happened, so the organization is smarter next time instead of just busier. Honestly, this is the step I have to force myself to do, and most teams I've watched never build the habit at all. It's nobody's favorite work.

This is where most teams leak value. They run 47 experiments and learn almost nothing, because the lessons live in seven different people's heads and three abandoned Slack threads. The activity looks like learning. It isn't.

AI is unusually good at exactly this, and it's underused for it. Point a model at the last twenty experiments your team ran and ask what patterns connect the winners. Feed it the customer-support transcripts from the last quarter and ask what's changed in what people complain about. Hand it the postmortems nobody ever reads again and ask what mistake keeps showing up under different names. You're not asking AI to do the work faster. You're asking it to do something humans rarely make time for: look across a hundred attempts and tell you what they add up to.

One more move pays off here: making your own organization's knowledge easy to get at. Consider what MIT's Center for Information Systems Research documented at one industrial distributor. Onboarding a new hospital used to mean manually mapping its catalog against millions of products from tens of thousands of manufacturers. Slow, expert-dependent, error-prone. By building an AI layer over their own data, they automated roughly 80 percent of that work across six million products and twenty-five thousand suppliers. The knowledge was always in the building. What changed is that it became searchable in seconds instead of locked in a specialist's head.

That's not a small convenience. New hires using AI built on internal documentation are reaching full productivity 30 to 40 percent faster, picking up in weeks the context that used to take months of asking around. When the cost of getting at what your organization already knows drops to almost nothing, every person's learning loop speeds up at once.

The capabilities, and which ones move most

If you walk the loop deliberately, here's where AI earns its place.

Sensing. You can read more signals, from more sources, faster than any team could by hand. Customer feedback, market chatter, competitive moves, support tickets at scale. The risk is mistaking volume for insight, so the real skill becomes asking sharper questions, not collecting more answers.

Deciding. This is where the latency drain hides. The analysis that used to take an analyst two weeks, pulling the data, modeling the scenarios, and writing it up, can now take an afternoon. Decisions that used to sit in a queue waiting for someone to "run the numbers" get unstuck. You can also stress-test a strategy by asking AI to argue the other side, model three different futures, and tell you where each one breaks.

One telecom worked through this deliberately instead of scattering pilots around. They built a portfolio of more than fifty AI agents, each with a named human owner and a simple rule for what it could decide on its own versus what it had to escalate. One agent cut the time to build a customer segment by 60 percent. Others now resolve more than 60 percent of in-app service questions without a human. The lesson isn't the agent count. It's that they treated AI as a set of decisions to govern, not a pile of use cases to collect. Frame the decision, let the agent act, capture what it learned, and do it again.

Experimenting. This is the one we walked through earlier. The 70 percent cut in prototype cycles, the week of work done in an afternoon. It's worth restating why it matters for the loop, not just the calendar. When an experiment costs almost nothing to run, the bottleneck moves. It stops being "which idea can we afford to test?" and becomes "are we disciplined enough to actually read what each test told us?" Cheap experiments only pay off if someone closes the loop on the result. Run a hundred of them and ignore what they teach you, and you haven't built a faster learning loop. You've just automated the production of evidence nobody reads.

Capturing. This is the biggest underused lever of them all. Sensing, deciding, and experimenting all throw off raw material. Signals, transcripts, results, postmortems. Capture is the step that turns that pile into something the organization actually knows. Humans skip it because it's invisible work. There's no demo, no launch, no applause for writing down what last quarter's three failures had in common. So the lesson gets relearned the hard way a couple of quarters later, by someone who never heard about the first time. AI takes away the excuse. It will read across two hundred support tickets, forty experiment write-ups, or a year of retro notes and surface the pattern in minutes. That's exactly the synthesis nobody ever had time to do by hand. Wire it in and you stop paying the quiet tax of relearning what you already knew.

Look at which capabilities speed up the most. They're the ones that used to depend on scarce, expensive human attention. Deep analysis, cross-domain synthesis, reading across a hundred documents to find the through-line. AI is cheapest exactly where human attention is most expensive. That's the opportunity, and most teams walk right past it.

The meta-move: using AI to get better at using AI

There's one more loop, and it sits on top of all the others.

Chris Argyris, a Harvard professor who spent his career studying how organizations learn, and how they fool themselves into thinking they have, drew a distinction decades ago between single-loop and double-loop learning. Single-loop is fixing the error. The thermostat notices the room is cold and turns on the heat. Double-loop is questioning the setting itself. Should the target be 68 in the first place? Single-loop makes you better at the thing you're doing. Double-loop makes you better at deciding what to do.

The teams getting real mileage out of AI are running the double loop on their own AI use. They look at which prompts actually produced something useful and which didn't. They notice the model is great at first drafts and terrible at final judgment, and they redesign the workflow around that. They keep finding new places to point it. They're not just using AI. They're learning how to use AI, and getting measurably better at it month over month.

That's the compounding engine. A team that improves how it learns will, over a few cycles, pull away from a team that just runs the same loop faster. In the same way compound interest pulls away from simple interest. It's slow, then it's sudden. And the payoff isn't abstract. Bain & Company built an index that ranks companies on how well they change, and the spread is stark. The most adaptable companies deliver roughly twice the profit growth and shareholder return of the least adaptable. The gap between learning fast and learning slow shows up on the income statement.

What to do this quarter

Don't launch an AI transformation. Pick one loop and go deep on it.

Start by measuring the right thing. Most teams measure output. Tickets closed, features shipped, words written. Start measuring learning velocity instead. How long does it take you to go from a question to a validated answer? From a signal to a decision? If you can't see that number, you can't shrink it.

Then find your slowest step. For most organizations, it's either decision latency, good ideas waiting weeks for analysis, or learning capture, where the lessons evaporate. Point AI at that one step, not at everything. Remember the McKinsey finding. The organizations getting real bottom-line impact aren't the ones with the most AI. They're the ones who redesigned a workflow around it. Bolt AI onto a broken loop, and all you get is a faster broken loop.

And run the double loop on yourself. Once a month, ask what you've actually learned about using these tools well. Where they helped, where they fooled you, what you'd do differently. Watch for the leak, too. If AI is producing more output but your people are thinking less, you're trading a capability for convenience. That monthly habit, more than any single use case, is what separates the teams who compound from the teams who plateau.

The uncomfortable truth about that 78 percent number is that everyone now has the same access. The model isn't the moat. How fast your organization turns a signal into a decision into an experiment into a lesson, and then does it again a little smarter, that's the moat. Your competitors aren't just working faster than you. The ones who are going to take your customers are learning faster than you.

The good news is that you can see your own loop. You can measure it. And you can start tightening it this week. The only question is whether you'll do it before the market does it for you.