Every orchard owner knows the first rule of harvest: pick what you can reach. Start with the low branches, fill the baskets, and use the early crop to learn the work. Nobody mentions the second rule, which matters more. Before you pick anything, make sure the trees will still be standing next season.
The skill is telling the two situations apart before you spend the money.
AI projects follow the same logic. There is real value in starting with fruit that hangs low. There is also a long record of companies picking the wrong fruit from the wrong tree, in an orchard that was about to be cleared. The skill is telling the two situations apart before you spend the money.
What the warning sounds like, and why it is right
Digital transformation advocates have been saying the same thing for years: automating a process that is already broken gives you a faster broken process. McKinsey is now saying it about AI. In an April 2026 piece on customer care in banking, the firm describes banks that layer AI onto legacy operating models and automate broken processes rather than redesigning them, and it puts the result bluntly: “AI simply amplifies existing inefficiencies.” Its example makes the mechanism plain: an AI copilot added to a broken workflow gives agents very little relief. The tool is fine; the workflow around it is the problem.
The firm’s broader survey says the same thing in numbers. In The State of AI in 2026, published in August, McKinsey reports on 1,719 respondents in 97 countries, surveyed from May 4 to June 8, 2026. About 37 percent say AI has contributed positively to their organization’s EBIT, essentially unchanged from the year before. About 6 percent qualify as high performers, also flat. What sets those high performers apart, in McKinsey’s telling, is that they redesign workflows around AI instead of inserting AI into the workflows they already have. Nearly three-quarters of them report doing so, against roughly a quarter of everyone else. These are McKinsey’s survey results and McKinsey’s interpretation, and we treat them as a strong signal and not a law of nature.
Take the warning seriously. If you point AI at a process that crosses four departments, three systems, and two sets of unwritten rules, the best case is that you encode the confusion in software, and the worst case is that nobody can tell whether the output is right.
The other side of the coin
Now read that warning again for what it does not say. It does not say every process needs a redesign before AI touches it. It says that complexity and dysfunction travel with the process, and that AI amplifies whatever it is given.
Some processes are small enough that the redesign and the automation are the same piece of work. They live inside one team. They run in one system, or in a spreadsheet and an inbox. One person can describe them end to end. They are slow, repetitive, and annoying, and everyone involved already knows it. For a process like that, redesign is cheap and automation is cheap, the people who would benefit are in the room, and a mistake costs a bad afternoon and not a bad quarter.
That is the fruit worth picking first. Illustrative examples, not a list we measured: first-line triage of internal IT tickets, first drafts of month-end variance commentary for the finance team, answers to routine internal policy questions, or summaries and tagging of the documents one team receives every day. In each case the process is self-contained, the work is plainly inefficient, and nobody is arguing about whether it should exist.
Check the orchard before you pick the tree
The catch is the phrase “should exist.” A self-contained, inefficient process is a good target only if you are confident it will still matter in a few years. That is the orchard question, and it is the one skipped most often because the tree itself looks so appealing.
A process that fails any of those tests is a tree in an orchard about to be cleared, no matter how low the fruit hangs.
Zoom out before you choose. Will the product or service this process supports still exist in three to five years? Is the system it runs on scheduled for replacement, merger, or retirement? Is the regulation, customer expectation, or contract behind it stable? Is the team that owns it likely to be reorganized, outsourced, or absorbed? A process that fails any of those tests is a tree in an orchard about to be cleared, no matter how low the fruit hangs.
Here is a case from our own recent writing. A manufacturer could spend its first AI dollars speeding up quote entry for a commodity line. Our 2026 AI Dividend Map puts industrial manufacturing in the Rising Bar shape, with commodity lines exposed to buyers’ procurement agents that source on spec and price alone, and we wrote about what that means in Preparing for AI Buyers Agents. The quote-entry process is self-contained and inefficient. Whether it belongs in the orchard depends on a strategic question that sits above the process, and a look at a map of your industry is a better place to start than a process diagram.
Why the harvest is worth more than the fruit
If you pick the right fruit, the savings are real but modest. The larger return is what your organization learns while picking.
A small, contained project teaches you to find the data, clean it enough to use, and decide how a human reviews the output. It teaches you to write down who owns the use case and what could go wrong, which is exactly the inventory discipline argued for in The 9% Problem. It gives you a chance to practice governance at low stakes, so that the policy, the review step, and the logging exist before the stakes rise. We describe the idea in Governance is a Freeway, Not a Roadblock: marked lanes let teams move faster, and a small project is the cheapest place to paint them.
It also gives you a measurement habit. Pick the unit of work, count what comes through the door, and hold the result against a baseline, the way we lay out in The Control Group Is Gone. A technology number that has survived one small project can earn a line on the weekly scorecard we described in Including Tech in the Weekly Meeting. By the time you attempt something hard, you have people who have shipped AI before, a review process they trust, and an honest record of what worked.
Then have the transformation conversation
Once the low fruit is picked, you have reached the point to stop and ask the larger question, before you aim at what is higher in the tree. The next class of targets is different in kind: processes that cross departments, depend on several systems, sit inside the supply chain, face customers, or would stop the business if they failed. This is where McKinsey’s warning applies with full force, because redesign is no longer optional and no single team can decide it alone.
We do not mean this as a polite way of saying wait. Set a date. A reasonable harvest is a quarter or two, with a small number of projects, and the strategic conversation goes on the calendar before the first project starts. We have written before about what does and does not count as digital transformation. The conversation we mean covers what the business will look like, which processes are central to it, which can be redesigned, and which should be retired and not automated. The harvest gives you a more realistic basis for answering them, because you now know what your organization can build, govern, and measure.
What this will not do
Low-hanging fruit does not add up to a transformation, and a company can pick it for years and still be standing in the same orchard. The method has a failure mode of its own: small wins that are never followed by the larger conversation. It also depends on the zoom-out being honest. A leader who has already decided to keep a process alive can find reasons that it will last, so ask someone outside the process to check the answers.
What you get is something narrower and valuable: early returns with low risk, a team that has done it before, and a clear view of which trees deserve the bigger investment.
What you get is something narrower and valuable: early returns with low risk, a team that has done it before, and a clear view of which trees deserve the bigger investment. If you want help sorting your candidate processes into the right orchard, or setting the date for the larger conversation, an independent assessment is a good place to start. Either way, before you approve the next AI project, ask one question of the process it targets: will this still matter in three years?
