Years ago we ran what turned out to be a clean experiment. A Procter & Gamble distributor, Navo Orbico, was facing high-churn, low performance and low motivation from the sales team. The Octalysis Groupo designed a completely new , fully gamified, sales team experience. The company let its sales reps keep using their existing CRM if they so wanted. It was functional and efficient, built for data entry and reporting, and it was boring. We also offered employees to use teh Octalysis rebuild: an immersive interface where their territory became an open sea, they captained a trading ship, they made real strategic choices about how to grow, they collaborated to keep a shared port city alive, and the work surprised them along the way.
Effectively all employees chose the engaging version, and they kept choosing it for years rather than for a novelty-driven few weeks, with revenue up 28.6% and voluntary adoption at 99.5% measured across the program’s multi-year run. The efficient CRM tool sat right there on the shelf, fully available, and they left it there.
That choice is the most important data point I know about AI employee engagement, which is worth defining before going further. AI employee engagement is the degree to which employees voluntarily and repeatedly use an AI tool because they want to, with their use deepening over time, rather than because policy requires it. By that definition the AI agents most companies are deploying right now are the efficient, boring CRM. They are more capable than anything we have built before, and capability is not what won the P&G reps over. Understanding why is the difference between an AI rollout your people choose and one they quietly route around.
Companies Buy the Promise, Not the Engagement
No company sets out to buy a boring tool, not in teh case of a CRM system nor in the case of an AI model. They buy a promise, and the promise is technological: the agent answers complex questions accurately, retrieves the right policy in seconds, drafts the proposal, reconciles the report, works across every system at once. That use-value is real, and it is what gets evaluated in the procurement process and demonstrated in the pilot. The business case is built entirely on what the AI can do.
What almost nobody evaluates is whether anyone will still want to use it in month six. Long-term engagement does not appear on the feature comparison, because it is not a property of the software. It is a property of the experience designed around the software, and that experience usually does not exist, because no one was responsible for it. The vendor sold capability, the buyer bought capability, and engagement was assumed to follow. It did not.
This is the gap that the entire field of behavioral design exists to close, and it is invisible until adoption stalls. A capable AI agent with no designed experience is exactly the position the old P&G CRM was in: it did its job, and people abandoned it the moment something more engaging was available. The efficiency that justified the purchase is the same efficiency that fails to create a single reason to return once the immediate task is done.
The Octalysis Framework draws a sharp line between extrinsic drives, the rewards and pressures that push from outside, and intrinsic drives, the meaning and mastery that pull from within. A purely functional tool, at best, runs on the thinnest extrinsic motivation: use it because you have to, because the task requires it, because someone is watching the dashboard. That earns compliance and nothing beyond it. It never reaches the voluntary, sustained, years-long use the P&G reps demonstrated, because compliance and enthusiasm come from completely different parts of human motivation.
What is Missing: Motivation by Design
To see what a functional AI agent lacks, it helps to name what it would need. The Octalysis Framework, the gamification model developed by Yu-kai Chou that is now studied and applied by design teams worldwide, identifies eight Core Drives of human motivation.
Consider what the P&G interface actually switched on. Epic Meaning & Calling came from the narrative: a rep was no longer entering sales data, they were a captain expanding a trading empire, and the mundane became part of something larger. Empowerment of Creativity & Feedback came from genuine strategic agency: reps decided how to grow their territory and saw the consequences of their choices, rather than following a fixed script.
The social and emotional layer ran just as deep. Social Influence & Relatedness came from the shared port city, where the whole team’s fortunes rose and fell together, so people had a reason to help each other rather than hoard. Unpredictability & Curiosity came from the surprises the system delivered, the moments reps did not see coming that made them want to look again tomorrow.
The old CRM activated none of these. It offered, at most, a weak form of Development & Accomplishment through quota tracking, and a background hum of Loss & Avoidance through the threat of falling behind. Two drives, both on the extrinsic, left-brain side of the framework, both producing the minimum necessary effort. That is precisely the motivational profile of nearly every AI agent shipping into the enterprise today: capable, efficient, and motivationally almost empty.
The reason this matters for AI specifically is that intelligence and motivation are separate problems. A more capable model answers the question better. It does not, by getting more capable, give anyone a reason to want the interaction. This is not just an Octalysis claim; it is the central finding of Self-Determination Theory, the decades-old body of research by Deci and Ryan showing that durable motivation comes from autonomy, competence, and relatedness rather than from the quality of the external output. You can double a model’s accuracy and the motivational profile of the experience does not move at all, because accuracy was never the thing people were responding to when they chose the engaging tool over the efficient one.
The Clippy Failure, in Behavioral Terms
There is a worse fate than being ignored, and Microsoft mapped it for the entire industry decades ago. In the late 1990s it introduced a cheerful animated assistant that appeared to offer help, and within a few years it had become one of the most resented features in the history of software, switched off by default and eventually retired after users made their feelings impossible to ignore. The lesson has been told as a joke for twenty years. It is actually a precise behavioral case study, and it is about to repeat at a much larger scale.
Clippy did not fail because it was unintelligent for its time. It failed because it inserted itself into the user’s work with no understanding of who they were, what they were trying to accomplish, or whether its help was wanted in that moment. Every appearance was an interruption the user had not chosen, which is a textbook trigger of the wrong kind of Loss & Avoidance: the small, repeated cost of having your attention hijacked.
It offered the user no agency, no sense of progress, and no relationship that deepened over time, treating the novelist and the accountant identically. In Octalysis terms, it activated a negative drive while activating none of the positive ones, and that is a reliable recipe for resentment.
There is a detail in Clippy’s history that sharpens the point. The original research version used a Bayesian engine that tried to model what the user was actually doing, and that engine was stripped out before launch, leaving a rule-based shell that fired the same suggestions at everyone. Even the engineers knew that adapting to the individual was the whole game. The execution shipped without it, and the resentment followed.
A modern AI agent has vastly more capability and exactly the same exposure to this failure. Drop a brilliant assistant into an employee’s workflow with no sense of their role, their goals, their level of skill, or what would make returning to it worthwhile, and you have built a far more articulate Clippy. It will give excellent answers and still feel like something happening to the person rather than something they are choosing.
But does the AI tool know who this person is and what they are trying to do, or does it treat the novice and the expert, the curious and the anxious, exactly alike? People do not sustain engagement with bots that talk at them. They sustain it with experiences that respond to them, and responding to a human being means responding to their motivation, not just their query.
So the real questions for anyone deploying an AI agent are not about the model, because the model is already good enough. The first question is how to make interacting with the agent genuinely engaging, the way the P&G interface was engaging. The second is how to keep it from sliding into the Clippy trap, where capability and intrusiveness combine into something people actively avoid. Neither question is answered by the technology. Both are answered by design.
The Four Phases of AI Employee Engagement
Behavioral design treats engagement as a journey with distinct phases, not a single state to be switched on. Every Octalysis engagement runs through four of them, and an AI agent inside a workforce passes through each one whether anyone designed for it or not. Mapping a typical rollout against the phases shows exactly where AI employee engagement leaks away.
Discovery: why an employee would start at all
Discovery happens before the employee has committed to using the agent. The question it answers is why they would bother, given that they already have a way of working that functions. A rollout email and a mandatory training session are not Discovery; they are an instruction, and an instruction borrows authority rather than creating desire. The P&G reps had a real reason to try the new interface: it looked like something they might actually want to do. Most AI agents are introduced as an obligation, which means they fail Discovery before the first interaction, and an experience that fails Discovery spends the rest of its life pushing uphill against indifference.
Onboarding: the first win that earns a second visit
Onboarding is the first real loop, and its job is to make the employee feel competent and certain they have come to the right place. This is where a felt early win matters more than any feature. A new user who gets a single genuinely useful, slightly surprising result in their first session forms a reason to come back; a new user who faces a blank prompt and infinite possibility forms the conclusion that the tool does not really fit their work. Prompting an AI agent well is a learned skill, and almost no enterprise rollout treats that first loop as something to be designed toward a win. The result is the familiar pattern: an ambitious first try, a couple of mediocre follow-ups, and then a return to the old way of working.
Scaffolding: from obligation to mastery
Scaffolding is the long middle where a habit either forms or dies, and it carries the single most important move in behavioral design. Early engagement can run on extrinsic drives, the targets and the tracking, but those curdle into boredom or resentment if they keep leading. The design has to shift weight toward the intrinsic, right-brain drives over time: the employee’s growing mastery of the tool, the creative agency it gives them, the standing it earns them with colleagues. The P&G system sustained years of use precisely because it made this shift, rewarding strategy and collaboration rather than mere activity. An AI agent that feels identical in week fifty to week one has no Scaffolding at all, and a tool that never deepens is a tool people eventually outgrow and abandon.
Endgame: veterans who shape how the team uses AI
Endgame is for the employees who have mastered the agent and done nearly everything it offers, and the question is what role remains for them. In a well-designed system, these veterans become the people who teach newcomers, who model advanced use, who carry the practice outward through the organization and become its internal champions. This is enormously valuable for AI adoption specifically, because peer demonstration beats any training program. Yet almost no AI deployment gives its power users any status, any platform, or any reason to invest in the community of practice around the tool. The most capable users are treated as a usage statistic rather than the engine of cultural adoption they could be, and the organization leaves its best path to broad engagement completely unbuilt.
What P&G Proves About Designed Motivation
The temptation is to read the P&G result as a story about gamification graphics, as if a nicer interface won the day, but that misreads what happened. The reps did not choose the engaging system because it was prettier, and adoption was not driven by efficiency gains either: the same program improved critical sales behaviors like accurate reporting and up-selling by 60%, which is the opposite of people choosing a smoother shortcut. The work was recognized with a Best Gamification Project award, but the award is not the point; the voluntary, durable adoption is. They chose it because of the motivation designed into it, and they kept choosing it because that motivation was the durable, intrinsic kind rather than a novelty that wears off.
This is the behavioral point that matters for every AI deployment. Given a free choice between an efficient tool and a motivationally designed one, people will choose designed motivation, and they will sustain that choice for years, even when the efficient option remains available at no cost. Efficiency is a reason a tool is allowed to exist. Motivation is the reason a human being returns to it voluntarily. Those are different currencies, and AI strategy consistently confuses them, investing everything in the first and nothing in the second.
The pattern holds at a scale far beyond a distributor’s sales force, again in our own client work. At CAIXA Econômica Federal, one of Brazil’s largest banks, active participation across a workforce of more than 80,000 people in 4,500 locations sat near 10% under conventional top-down management. Redesigning the daily experience around intrinsic Core Drives rather than directives lifted recurring net income from 8.6 billion to 12.7 billion reais in a single year, roughly a billion dollars at the time, and moved the bank from second to first in its market. The mechanism was identical to P&G’s: people given genuine agency, meaning, and social connection in their work engage with it differently than people handed an efficient system and told to comply.
A capable AI agent could have served both of these workforces well, handling the functional load that humans should not have to carry. But an agent dropped on top of the old, motivationally empty structure would have produced another efficient tool people used only when required. The technology would have changed. The engagement would not have, because engagement was never going to come from the technology.
Why the Model Cannot Supply the Experience
It is reasonable to ask why a sufficiently advanced AI cannot simply generate this engaging experience itself. The answer exposes the core limitation, and it is the same one that defines our work: an AI agent can see what an employee asks and what they do, but it cannot see why. It does not know whether this person is driven by mastery, by recognition from their peers, by the fear of getting something wrong, or by curiosity about what else is possible. So it speaks to everyone in the same competent, neutral register, and a tool that treats every employee identically is a tool no employee experiences as theirs.
This is the deeper problem we examine in our pillar on AI user engagement: behavioral data describes what happened and is silent on what would make someone care. Reading motivation rather than mere behavior is exactly the gap our patent-pending Octalysis AI Engagement Engine was built to close, by inferring which Core Drives are active for an individual from the signals they already generate, without subjecting anyone to a survey.
An agent that knows an employee is driven by social standing can surface their contribution to the team; an agent that knows another is driven by mastery can show them how far they have come. Same underlying model, a completely different experience, because the motivation layer is finally present.
Until that layer is built, deliberately, by people who understand human motivation, the most advanced AI in the workplace will keep arriving as a brilliant stranger that never learns who you are. Capability without that understanding is precisely what Clippy had, in primitive form, and precisely what the bare CRM had at P&G. The reps had a name for the alternative, even if they never said it out loud. They called it the tool they wanted to use.
Designing an AI Agent People Actually Choose
Designing for AI employee engagement is the reverse of how most rollouts are sequenced. The technology comes last, after the motivation, not first with the motivation assumed. A few principles from our work make the difference between an agent people choose and one they endure.
Build a real Discovery moment rather than a mandate. People should encounter the agent as something that offers them a win they want, not as a compliance requirement, because desire that is borrowed from policy evaporates the moment the policy is not watching. Engineer a genuine early success into the first interaction, so the employee leaves their first session feeling more capable than when they arrived, which is the foundation of any returning habit.
Layer the intrinsic Core Drives onto the function deliberately. Give employees real agency in how they work with the agent rather than a single prescribed path, which activates Empowerment of Creativity & Feedback. Connect their use of it to a shared team outcome, which activates Social Influence & Relatedness, and let them feel themselves mastering it over time while it occasionally surprises them with something useful they did not ask for. None of these are model capabilities. All of them are design decisions, and together they are what the P&G interface had and the CRM lacked.
Attend to the manager, because the same software produces opposite motivation depending on how it is framed. McKinsey’s Superagency in the Workplace report found that employees are roughly three times more likely than their leaders assume to already be using AI for a substantial share of their daily work, and that the binding constraint on scaling is not employee resistance but leadership moving too slowly. Framed by a manager as a way to grow, the agent activates Empowerment of Creativity & Feedback. Framed as surveillance or as a prelude to cutting jobs, the identical agent activates Loss & Avoidance and fear, and people perform the minimum and resent the rest. The motivational frame is a leadership decision, and it is at least as decisive as the design of the tool.
Finally, design the full journey rather than the first screen. Most engagement budgets are spent making the introduction smooth and nothing on Scaffolding or Endgame, which is why adoption curves spike and then sag. The reason the P&G reps were still choosing the gamified system years later is that the experience kept deepening, kept giving mastery and status and surprise, long after the novelty was gone. An AI agent designed only to impress in the demo will be abandoned on roughly the same timeline as every disengaging tool before it.
This is urgent rather than optional, because of the ground the technology is landing on. In 2025 Gallup found only 20% of employees worldwide engaged at work, the lowest level since 2020 and the first back-to-back annual decline it has recorded. When AI agents arrive into a workforce where four in five people are already running on minimal motivation, the question is not whether the tool is capable. It is whether deploying it shrinks that 20% further or finally starts to grow it.
A disengaged workforce will not be rescued by a more capable chatbot any more than the P&G reps were going to be rescued by a faster CRM. They are reached the way they have always been reached, by an experience designed around what actually moves human beings, and AI makes that experience reach further and adapt faster than ever once the motivation is built underneath it.
The model is not your problem. The experience around it is, and that is the problem we have spent two decades learning to solve. Contact The Octalysis Group and we will design the engagement journey your AI agent is missing, before your people decide it is just another tool they would rather not open.
Joris Beerda is CEO and Co-Founder of The Octalysis Group, the behavioral design consultancy behind workforce and customer engagement programs for Microsoft, P&G, Porsche, and CAIXA Econômica Federal. Explore the firm’s case studies for the full results referenced here.
Frequently Asked Questions
Why do AI agents and workplace chatbots fail to engage employees?
Because they are bought and built for capability, not engagement. An AI agent that answers questions accurately is the equivalent of an efficient but boring CRM: useful when someone is stuck, ignored the rest of the time. Engagement comes from the experience designed around the tool, including agency, social stakes, visible mastery, and meaningful surprise, which activate the intrinsic Core Drives. None of these are properties of the model, so a more capable model does not produce a more engaged workforce.
What is the Clippy problem in AI employee engagement?
Clippy was Microsoft’s late-1990s assistant that became infamous for interrupting work without understanding what users wanted. In behavioral terms it triggered a negative drive, the cost of hijacked attention, while activating none of the positive Core Drives, which is a reliable recipe for resentment. A modern AI agent is far more capable but can repeat the exact failure if it inserts itself into the workflow with no sense of who the person is or what would make the interaction worth repeating.
What did the Procter & Gamble case prove about employee engagement?
Reps were given a free choice between their existing functional CRM and an immersive, gamified interface for the same job. Effectively all of them chose the engaging version and kept choosing it for years, with revenue up 28.6% and voluntary adoption at 99.5%. It proves that designed, intrinsic motivation beats raw efficiency for sustained engagement, even when the efficient option remains available at no cost.
How do you make an AI agent genuinely engaging rather than just functional?
Layer the intrinsic Core Drives onto the function on purpose: give people real agency and strategic choice, connect their use of the agent to a shared goal, let them feel themselves mastering it, and build in genuine surprise. Design the full journey across Discovery, Onboarding, Scaffolding, and Endgame rather than just a smooth first screen, and ensure managers frame the tool as growth rather than surveillance. These are deliberate design decisions, not capabilities of the underlying model.
Does AI have a role in employee engagement at all?
Yes, a significant one, once the motivation design exists. AI is well suited to personalizing which challenge or content an employee sees, timing nudges, calibrating difficulty, and eventually inferring what motivates each person. The mistake is leading with the technology and assuming engagement will follow. The technology amplifies whatever motivational architecture is already in place, for better or worse.