By Dr. Islam Gouda
A woman renews her home insurance every year without complaint. Claims get paid on time, and when her basement flooded two winters ago, a claims adjuster she’d spoken with before picked up on the second ring. By any measure her insurer would use internally, she is a model customer: satisfied, low-maintenance, retained.
This year, she doesn’t renew. She never even opens the renewal email. Her phone’s artificial intelligence (AI) assistant, the one she now uses for most of her shopping, quietly compared six insurers on her behalf, found one offering nearly identical coverage for 11% less, and switched her over during a routine monthly check-in she barely noticed. She didn’t decide to leave, and she never spoke to the new insurer. Something else made the decision, on terms neither company controls.
Ask the executives at the company she left why she went, and you’ll get the answer their dashboards give them: Nothing changed. Satisfaction held. Net Promoter Score held. Yet the revenue is gone, because nothing about the experience failed. What failed was upstream of experience entirely: Nobody, not the algorithm and not the customer, ever quite decided that this insurer, specifically, deserved the credit for the good service she’d been getting. Revenue that nobody has claimed is revenue anyone can take.
This is happening across categories that have nothing else in common: electronics, software, travel, groceries, professional services. A customer’s AI assistant increasingly does the comparing and often the buying, rewarding whichever option looks best on the dimensions it can measure, largely indifferent to the relationship a company spent years building. The company that created the value and the company credited for it are starting to be two different companies.
The Same Story, Inside Your Own Organization
Now watch the identical pattern happen somewhere you have probably felt it yourself.
A customer service representative completes a well-designed training program, a simulation-based course on handling objections. Six months later, her call quality and close rate are up sharply. When leadership reviews what changed, they credit the AI coaching tool built into the call software, the one that flags objections in real time, because that tool is what everyone in the room can see working, moment to moment. Nobody mentions the training. The learning and development (L&D) team that built it is never asked, and next year’s budget for the program quietly shrinks in favor of more AI tooling.
Nothing about the training failed. Completion rates were strong, the scores were good, and behavior genuinely changed on the job, the hardest thing training is ever asked to prove. What failed sat upstream: Nobody ever traced the improvement back to the learning experience that made it possible.
L&D has lived with a milder version of this for as long as the field has existed, which is why so much of its own literature is devoted to proving business impact. What’s new is not the difficulty; it’s where credit lands instead of going unclaimed. It used to default to nowhere in particular. Increasingly, it defaults to somewhere specific and visible: An AI tool sitting closer to the moment performance shows up, easy for an executive to credit in a budget meeting, whether or not that tool is why the employee knew what to do with what it suggested.
This article calls the pattern Attribution Instability: the growing gap between whoever, or whatever, created the value a person experiences and whoever, or whatever, gets credited for it. It shows up when a customer’s AI shopping agent reassigns loyalty to a cheaper competitor, and just as clearly when an AI performance-support tool quietly absorbs the credit that belongs to the learning experience underneath it. Both are the same structural problem, arriving in a function that has never had to solve it before, because for most of marketing’s and L&D’s history, a customer or a manager could usually tell who created the value they received. That assumption is fraying because something now sits closer to the moment the result becomes visible, and whatever sits closest tends to get the credit, correctly attributed or not.
AI Didn’t Create This Problem; AI Revealed It
Distributors and private-label retailers have quietly redirected credit away from the party that earned it for decades. Online price comparison later eroded the loyalty premium a trusted brand once commanded, precisely because comparison used to be expensive.
Agentic AI is the first technology to mediate all of it at once: what a person sees, how options get compared, and whether an action happens at all. General-purpose AI assistants now field comparisons that used to go to a search engine; retailers build shopping assistants into their own apps; enterprise buyers run vendor comparisons through AI tools optimized for a platform’s own definition of a good match, not any seller’s differentiation.
Inside organizations, AI coaching assistants and performance dashboards do the same thing to employee performance that shopping agents do to customer loyalty, sitting at the exact moment value becomes visible and quietly absorbing the credit for it. None of these tools set out to strip credit from whoever earned it. The by-product is that the question of who deserves it quietly stops getting asked at all.
3 Ideas Every Learning Leader Needs, in Plain Terms
1. Relational Attribution
Relational Attribution is the largely automatic judgment a person makes, at the moment they receive value, about who or what is responsible for it, the same way people instinctively look for someone to praise after good news.
A manager reviewing an employee’s improved performance is making this judgment every cycle, whether or not anyone names it, and it doesn’t always resolve toward the learning program strongly enough to survive a flashier AI tool appearing in the same window of time.
2. Attribution Integrity
The second, Attribution Integrity, is rarely measured because it isn’t the same thing as satisfaction or completion rates. It is whether a program gets credited consistently at every point a learner or manager encounters its results, not just on average.
A program that was praised in an end-of-course survey and then becomes invisible by budget-review time has a worse attribution problem than one only modestly liked throughout, because that whiplash damages credibility more than mediocrity ever would.
3. Impact Durability
Impact Durability is the share of a learning function’s credited value that would survive if a comparable AI performance-support tool became available tomorrow. Some impact is durable because a manager genuinely understands what the training built; some survives only until something more visible appears in the same workflow.
Two programs can show identical completion and behavior-change scores while carrying wildly different amounts of the durable kind, and until AI tools started sitting inside everyday workflows, almost nobody could tell the difference from outside.
The Attribution Matrix
Put these ideas together and a diagnostic emerges, built from two questions:
- Is the value we deliver genuinely excellent?
- Do the people who benefit from it know, specifically, who or what to credit for it?
Most leadership teams have a confident answer to the first question and no answer at all to the second.
The elements of the matrix are described below.
Durable Growth
Excellent value, clearly credited. A learning program with a named facilitator whom managers specifically ask for by name tends to live here, the internal equivalent of a specialty insurer whose agent calls after every claim, not just at renewal.
Invisible Excellence
Excellent value, unclear credit. A training program that measurably built a skill, while the praise for using it goes entirely to the AI tool that surfaced the moment, is a textbook case. This is the most dangerous quadrant, because every metric an L&D team already tracks looks fine right up until the budget conversation. The opening example of the insurance customer switching providers falls here.
Borrowed Time
Mediocre value, clearly credited. A mandatory compliance course with a well-known reputation for wasting time sits here, protected only because nothing else has replaced it yet.
Silent Erosion
Mediocre value, unclear credit. Weak performance with no relationship to fall back on. The program’s credibility is already gone; nobody has said so out loud yet.
Most leadership teams can place their own programs on this grid within minutes of honest conversation, and most are startled by where they land, often discovering strong instructional design sitting in Invisible Excellence, quietly credited to a manager’s reputation or an AI tool instead.
The matrix isn’t meant to replace a proper evaluation program; it is intended to force a conversation that the leadership team would otherwise never have. It also has real limits:
- The matrix is most useful in situations where an alternative to the AI tool exists, results can be compared, and attribution to the AI tool or the training can genuinely be contested. This is common in settings like retail or coached sales training.
- The matrix is less applicable to areas where specific, irreplaceable expertise, such as a surgeon’s or pilot’s, is needed or where long-term relationships are in play, since AI is not a direct substitute for key elements that influence performance or attribution.
Who Is Most Exposed
Vulnerability tracks how much of a program’s value lives entirely inside a dashboard number rather than a person, not how digital the program is. Fully self-serve AI upskilling platforms are the most exposed, because there is no instructor that anyone remembers and no human moment for credit to attach to. The same is true of onboarding fully replaced by chatbots, and of programs measured only by a quarterly completion number nobody remembers by the next budget cycle.
Programs with a strong, recognizable human presence, a named facilitator or owner who shows up in performance conversations, are comparatively well positioned, not because they are immune to AI but because they already have something concrete for attribution to attach to.
The risk was never really about how much AI a learning function uses; it is about whether anyone, or anything, remains identifiable as the party responsible for the value learners received.
What Changes for Learning Leaders
For the chief learning officer or head of L&D, showing L&D’s contribution means treating attribution as a line item alongside completion rates, separating program impact by durability, and treating a decline in the durable share as an early warning—long before it becomes a budget cut. It also means asking a genuinely uncomfortable question about the function’s own AI tooling investments: Does a coaching tool that quietly gets the credit for a skill it did not build end up erasing the very evidence that would justify next year’s learning budget?
For program owners and instructional designers, it means recognizing that a well-built course, if nobody remembers who built it once results show up, may build real capability without protecting the program that built it. The work shifts toward making sure something concrete—a name, a specific session—is what people’s attention lands on when they explain why performance improved.
For people analytics or the chief people officer, it means auditing performance reviews for the moment credit gets assigned—and asking whether that moment ever mentions a learning program at all. A review that credits only a tool or a manager’s coaching is structurally set up to erase L&D’s contribution, regardless of how good it actually was.
For whoever owns AI tooling decisions, and for executive leadership more broadly, it means asking a question no standard reporting package answers well: If every AI performance tool in the building were removed tomorrow, how much of the improvement being credited to those tools would disappear, and how much would still be there because someone actually learned something? Very few organizations can answer that today, and fewer still have anyone whose job is to make the answer better.
Where to Start
The first move is almost always working through the Attribution Matrix, because it costs an afternoon and changes what a leadership team argues about afterward.
The second is an audit of where a learning journey currently goes faceless, marking the points where a name or facilitator disappears into a fully automated flow.
The third is smaller still: Pick one high-value program and start asking, in manager check-ins, who or what the manager believes is responsible for an employee’s improved performance.
Most L&D teams have never asked this in so many words, and the answers are usually more useful than another quarter of completion tracking, because they measure the one thing tracking was never built to see.
The Next Competitive Advantage
For 20 years, the dominant playbook has rewarded whoever builds the smoothest, most frictionless experience, because a smoother experience was reliably a more attributable one. Autonomous AI tools are the first force powerful enough to break that link: It is now possible to deliver genuinely excellent training that a manager credits to nobody in particular, and to lose a program’s budget without a single complaint ever being filed against it.
That breaks the old playbook, but not the truth underneath it. This is not a call to resist AI; the insurer in the opening example did nothing wrong by building a good app, and the coaching tool did nothing wrong by flagging objections in real time. The problem was never that anyone automated too much. It was that nobody was accountable, by name, for making sure credit had somewhere specific to land once the automation was in place.
The learning functions that do well over the next decade will not be the ones that slow their adoption of AI. They will be the ones that pair every new layer of automated support with a deliberate answer to a much older question: When this works well, who, exactly, is going to get the credit? That question used to answer itself. It no longer does, and pretending otherwise is the single costliest assumption a learning function can still be making.
Image credit: AlonzoDesign

