AI user engagement is the degree to which people choose an AI product, assistant or agent for the actions a business depends on, keep using it on a consistent basis, and are enthralled by the experience it gives them. A first try is only the start. Engagement shows in what follows: people reach for the AI every time the task comes up, their use becomes broader and more skilled, and they would object if you took it away.
Three examples show what that looks like:
- A bank customer asks the assistant in their banking app to set up a savings goal for a house deposit. The bank needs customers to save and plan through the app. This customer is engaged when they come back after every salary payment, because the assistant shows what last month’s choices did to the target date and gives them a real decision to make about this month.
- An accounts payable analyst hands supplier invoices to an AI agent and reviews what comes back. The company needs invoices matched correctly, with less rework. The analyst is engaged when they use the agent for every suitable batch and have started proposing rules for the cases it gets wrong, because they can see those rules reduce the corrections the following week.
- A patient logs each treatment day with an AI coach. The healthcare provider needs patients to stay on their treatment. The patient is engaged when logging no longer feels like admin, because each entry teaches them something about their condition and shows how other people in the same situation are doing.
In each case the business gets the action its metric depends on, and the person gets an experience they would miss.
The problem this article addresses is the distance between those examples and what many companies see in their own data. AI is being used: in Gallup’s Q2 2026 survey, more than half of U.S. workers said they use AI in their role. What lags is what that use delivers. Many people who use AI spend a large part of the time it saves on fixing its output, many have hidden their use, and most enterprise pilots never show up in the P&L. The technology is rarely the cause. The experience around the AI was built for functionality, and nobody designed it for the motivation of the person using it. That makes AI user engagement a design problem more than a technology problem. The question I answer here is how to design an AI product or rollout so that people take the actions the business needs, keep taking them, and enjoy the experience.
Yu-kai Chou, the creator of the Octalysis Framework, and I co-founded The Octalysis Group, and we have applied the framework in projects for more than 175 companies. This article is our hub for AI and engagement. I wrote it for people who build AI products for customers and for people who bring AI to employees, learners or patients, because the design question is the same for all of them. Several of the projects I use as evidence involved no AI. I say so where they appear, and I use them because the motivation underneath is the same.
What AI User Engagement Means
Every design project we do starts in the same place: a business metric the company wants to improve, the Desired Actions people need to take for that metric to move, and the motivation of the people who have to take them. AI changes none of that. Our hub on what gamification is and what it is not explains the reasoning in full. For AI, three parts of the definition need precision.
The action is specific. Using the assistant to draft a client proposal in the middle of normal work is a Desired Action. Logging in is not. A prompt sent says little either, because a prompt can be the start of real work or a box someone ticked. This is why seats provisioned and prompts sent mislead. They tell you who has access. They say little about whether work got done, or whether the person would choose the tool again.
Consistent means every time the activity comes up. How often that is depends on the activity. A coding assistant may be in use all day. An AI that helps someone file an insurance claim is used a few times a year, and it is doing its job if people choose it each time and those few uses go well. Daily use is one possible goal, and it only makes sense where the business need and the activity call for it.
Enthralled means the experience itself draws people in. We look to games here, because nobody has to play a game and people drop one the moment it gets boring. A game that keeps its players for years has been designed around human motivation. An AI product earns the same response when using it gives people something they care about: a result they are proud of, a strategy of their own, something to discover, other people to do it with. One good sign is a user whose tenth week looks broader and more skilled than their first.
Some AI needs no engagement design at all. A fraud detection model runs in the background and nobody has to choose it. An assistant whose purpose is to make a task disappear should be judged on the task disappearing, and minutes per session would be the wrong target. This article is about the cases where a person has to choose the AI: a workplace assistant, a consumer product, an agent inside a loyalty program or a health app.
What the Data Says About AI Usage and Engagement
Gallup’s Q2 2026 survey of more than 22,000 U.S. workers found that 52% use AI in their role, 30% use it a few times a week or more, and 15% use it daily. That is a strong base. More than half of the workforce has started, and three in ten use AI every week.
The results are less clear. MIT’s NANDA initiative reported that about 95% of enterprise generative AI pilots deliver no measurable P&L impact. The lead author attributed the result to a learning gap and flawed integration, and did not blame the quality of the models.
Workday’s September 2026 survey of nearly 6,000 workers shows what using AI can feel like day to day. A third of AI users say they lose at least half of the time AI saves them to fixing or reworking its output. And 41% say they have deliberately concealed their AI use on a work task. The experience also differs by level: 44% of senior leaders report heavy AI use, compared with 9% of individual contributors.
Read together, the figures describe an engagement gap more than an access gap. People have the tools and many use them. Fewer have an experience that improves their work and that they are glad to talk about. We have written about each part separately: why people ignore AI agents, our reading of the 95% figure in AI agent adoption failure, and the different experience of leaders and their teams in the agentic gap.
A Design Problem More Than a Technology Problem
Email, CRM and e-learning each came with the promise that they would change the world. AI is getting the same promise now.
Gamification went through its own version about ten years before AI did. In November 2012, TechCrunch reported a Gartner prediction that 80% of gamified applications would fail to meet their objectives. Gartner’s Brian Burke blamed poor design: companies fixated on points and badges and did not build meaningful motivation into the experience.
The AI version of the badge is the chatbot window. It is the visible, easy addition, and it stands in for the work of designing the experience around it.
Most product teams are good at making a product logical, so that a user gets from A to B to C as efficiently as possible. A product can be logical and intuitive and still fail to engage people over time. Functionality matters, and a design should never assume that people will want to use a product because the functionality is impressive. With AI that assumption is tempting, because the functionality is so impressive.
What AI Does to Human Motivation
Yu-kai’s recent writing explains why a capable AI can leave people less motivated. In Gamification and AI: Motivation Design in the AI Age he states his thesis in one sentence: “AI creates motivation vacuums.”
Every time AI takes over a piece of execution, it removes an accomplishment loop that a person used to run. Someone who wrote a report felt they had done something when they sent it. Someone who edits an AI draft for ten minutes and sends it has supervised the work. He points to McKinsey’s 2026 research: “the employees using AI the most are simultaneously the most engaged and the most likely to quit.” His conclusion: “cheap intelligence makes execution nearly free, which means human motivation becomes the scarce input.”
Mapped onto the Core Drives of Octalysis, his reading is that “AI amplifies the drives built on novelty and fear (CD7, CD8), holds meaning roughly steady (CD1), and starves the drives built on doing, making, and connecting (CD2, CD3, CD4, CD5).” A generated answer can surprise you, and the fear of falling behind on AI is real. What thins out is the feeling of having accomplished something yourself, of having shaped the result, of owning it, and of doing it with other people. Those have to be designed back in.
He gives product teams a test that I use as well: “does this feature give the human a real accomplishment loop, or does it just give them a result?” For leaders he recommends a second measure next to usage: “Track a second number alongside adoption: felt ownership over outcomes.”
The commercial side of the argument is in his article Thriving in an AI World: “AI collapses the price of competence. It does not collapse the price of meaning.” When every company can generate competent output, judgment, trust and care become the scarce part. In customer support, AI can answer routine questions at scale, and the companies people remember will be the ones that know when a human should step in. In his words: “Efficiency becomes abundant. Care becomes scarce.”
I have described the same problem at the level of one team in designing meaningful work when AI can do the task: a role can stay useful to the company while giving the person in it little they consider worth doing.
Why Mandates and Usage Targets Are Not Enough
When usage lags, the usual next step is a mandate or a usage target. Usage goes up, because people do what is measured. A lot of them paste in a paragraph, tick the box and go back to how they worked before.
A mandate runs on Core Drive 8: Loss & Avoidance. People use the tool to stay out of trouble. That is Black Hat motivation. It creates urgency, and people give it the minimum the rule requires. Yu-kai says the same about “adopt AI or get left behind” messaging: “It works for a quarter. It burns people out by the year.”
Employees say so themselves. McKinsey asked them what would increase their daily use of generative AI. Formal training came first at 48%, smooth integration into existing workflows second at 45%, access to tools third at 41%, and incentives and rewards fourth at 40%. At the bottom were explicit instructions from managers (30%) and OKRs or KPIs tied to usage (22%).
A manager’s support is a different thing from a manager’s instruction. Gallup’s State of the Global Workplace 2026 looked at U.S. organizations investing in AI. Employees who strongly agree that their manager actively supports their team’s use of AI are 8.7 times as likely to strongly agree that AI has transformed how work gets done in their organization.
The usual additions to a rollout share the weakness of the mandate. Octalysis separates extrinsic motivation, where people act for a result, a reward or something they get to keep, from intrinsic motivation, where the activity itself is the reward. You do not have to pay anybody to be creative. Usage targets, leaderboards of prompts sent and badges for completing the training are all extrinsic. Extrinsic design can work well where the activity is already intrinsically motivating. Duolingo can lean on its streak because people want to learn a language, and Strava can lean on measurement because many people already enjoy cycling and running. Processing invoices or updating a CRM is different. Few people do those for their own sake, so more targets and counters double down on the linear, extrinsic character the work already has. An AI tool for that kind of work needs intrinsic design as well: a strategy the person chooses, colleagues they work with, something to discover.
AI User Engagement Through the Four Experience Phases
The first time someone uses an AI tool, their motivation and skill are very different from the hundredth time. Octalysis describes this through four Experience Phases: Discovery, Onboarding, Scaffolding and Endgame. They follow the person’s experience with the product, and the calendar does not define them. Most AI products put nearly all their effort into the technology and the first session. Here is how I apply each phase.
Discovery: a specific reason to try
Discovery is why someone would want to try the product at all. The usual AI pitch is that it can do anything. That impresses people and gives them nowhere to start. The strongest Discovery is specific. A colleague shows you one task that used to take an hour and now takes a few minutes, in the place where you already work. In a company, the most common Discovery is a mandate, with the results described above.
Onboarding: a first win
Onboarding is where people learn what they can do and reach a first win. It can be quite linear and predictable. People should feel that this is easy, that they can do it and that they came to the right place.
A new AI user often meets setup admin first and an empty prompt box second. Prompting is a skill that few products teach. Good onboarding starts with a task the person already does. If the agent drafts vendor emails, the first task should be the email the user writes every day, so the first output beats the manual version immediately. That finished result is Core Drive 2: Development & Accomplishment at work. People also differ in how they like to work with AI. Some want to talk to it, some want a conversation partner, and some want an agent that executes and nothing more. Onboarding should let each of them find a way of working that fits. The general principles are in our article on crafting the best onboarding sequence.
Scaffolding: where consistent use is decided
Scaffolding is the regular use. People take mostly the same Desired Actions again and again and slowly become senior users. If the design stays linear, it gets stale and starts to feel like work.
In many workplace rollouts the tool looks the same in week ten as on day one. It does not suggest the next skill, show what colleagues are doing with it, or give the user anything to build. This is the phase where we gradually add intrinsic design: meaningful collaboration, different strategies for different people, surprises along the way, and a clearer link to a purpose people care about.
Scaffolding is also where Yu-kai’s accomplishment loop has to be protected. People should not feel they hand their work to an agent and watch. Let them find the agent’s faults and change how it behaves. Take the accounts payable analyst from the opening. They notice that most of their corrections involve partial deliveries, propose that the agent route those cases to a person, and see the following week what that did to the number of corrections. The analyst made a choice with a visible consequence. That is Core Drive 3: Empowerment of Creativity & Feedback, and it turns checking output into work the person owns.
Endgame: a role for your most skilled users
Endgame is about the veterans. The design has to give them the ability to become ambassadors inside and outside the company. For an AI product that means roles: the person whose agents and prompts others use, who teaches colleagues, and who helps shape the next use case.
Our audit of the platform of Astana Hub, the technology park for IT startups in Kazakhstan, shows what a missing Endgame looks like. New users had no guided first experience and no early wins, and a member of two years was treated the same as one who joined that day, with no way to become a mentor or a content creator. The redesign covered all four phases, and the Astana Hub case study reports a 504% increase in active users. An AI product that treats its most skilled user like a newcomer has the same gap.
Products, People and Processes: Where AI Engagement Applies
AI and engagement matter in every area we work in. The Desired Actions and the people differ, and the design question is the same. The client projects in this section involved no AI. Each one shows the motivation design an AI needs to work with.
Products: customers who choose your AI
In loyalty, AI already personalizes offers at scale. It predicts what a member is likely to buy. It has no view on why that member would care about the program. U.S. consumers belong to roughly 17 loyalty programs and actively use fewer than half, which suggests that better-targeted coupons do not settle the matter. Our article on AI loyalty program design goes through this in detail.
For LATAM Airlines, we analyzed the business metrics and the motivational profiles of LATAM Pass members and replaced standard mileage banners with a treasure hunt built around the Desired Actions from that analysis. The case study reports a 153% increase in credit card acquisitions. This is what an AI agent in a loyalty program needs to draw on: activities, progression and feedback that give different members something they want to do. An agent cannot personalize a meaningful choice that the product never provides.
The same holds in banking and insurance, where the design will rarely look like a game. The savings assistant from the opening is one example. Another is an assistant that supports different strategies for different customers. Someone who mainly buys gardening supplies could earn bonuses at certain shops, so the product responds to what that person already cares about.
People: employees, learners and patients
For employees, the AI tools that promise engagement mostly predict who is at risk of leaving. They do little about what would make people want to stay and contribute, as I argue in why AI employee engagement fails without a motivation architecture. CAIXA Econômica Federal, the Brazilian state-owned bank, is our reference for what that architecture can do. More than 80,000 employees in about 4,500 branches had to get behind one sales goal. We designed a cooperative program around a shared national goal, with collectible puzzle pieces that employees earned for individual targets and traded with colleagues to complete sets. The CAIXA case study reports that voluntary participation rose from 10% to 92%. That is the starting point of our employee engagement work. Once a program like that exists, AI can adapt it to each person.
For learners, the Microsoft Learn design addressed people attracted to competition and accomplishment alongside people who valued social interaction and collaboration. It included progress and achievement systems as well as teams, communities and learning partners. An AI tutor that only recommends the next lesson leaves most of that unused. In learning and development, the AI should know why those options exist and which one fits the learner in front of it.
For patients, a 2025 review in Frontiers in Digital Health reported that AI-based interventions improved medication adherence by 6.7% to 32.7% compared with standard care, based on seven eligible studies. A reminder treats adherence as an information problem. Sidekick, the app in our Pfizer case study, lets patients track sleep, nutrition and mood alongside their treatment, set their own goals, learn about their condition in short modules, and connect with mentors and peers. The case study reports 30% fewer patient dropouts, with 68% of users still active daily after six months. Our article on AI patient adherence applies this to AI coaching, and it is the approach behind our health and wellbeing work.
Processes: AI inside the way work gets done
Most enterprise AI arrives as a change to a process, which makes it a digital transformation question: will people use the new system for the work it was introduced to improve?
Our project for Navo Orbico, a Procter & Gamble distributor, concerned a CRM. Routine CRM work was being neglected, and cash incentives had stopped working. We turned the sales process into a journey of captains trading for a shared city-state. Reps chose their own strategy for growth. Hitting KPIs earned health points that protected the whole city against mishaps and upgraded it over time. A rep who had a weak period could recover by helping colleagues with their sales process or by taking part in training. Participation was voluntary. The P&G sales gamification case study reports 99.5% adoption and a 28.6% revenue increase against historical performance.
Agent programs raise the same question with a new complication. A leader usually chooses what the agent does and keeps the benefit. A team member often receives an agent configured by someone else and spends part of the day checking its output. The Desired Actions differ by role: delegating a suitable task, reviewing with care, disclosing AI use, sharing an approach that works. Each role needs its own answer to what the experience gives the person and what it costs them.
Content Personalization Versus Motivation Personalization
AI can improve the personalization of an experience, and it does that well. Recommendation engines and language models decide what a user sees next. In my view they still miss the motivation of the user at that moment. An AI works from a slice of what a person has done, and it sees what they did without seeing why.
Two users can leave the same flow at the same step for opposite reasons. Take a fitness app with an AI coach that notices both users have stopped logging workouts. A content-based system sends each of them the same kind of nudge, tuned to their favorite workout and the time they usually train. The first user wanted a harder goal and gets a reminder to do the same easy routine. The second wanted a training partner and gets a cheerful message about their own numbers. The nudge arrives at the right time and still misses the motivation.
A common shortcut is to sort people into generic player types. We build player types from the client’s actual users, by separating them on two characteristics that matter for that product, such as people focused on efficiency versus people who enjoy the process itself. The dominant types each get a journey that fits them inside one experience. That is a useful start, and motivation also shifts as people gain skill.
The Octalysis AI Engagement Engine
You could ask people what motivates them. A survey interrupts the experience, and the answers go out of date quickly. Intent and motivation are the data an AI lacks.
That is why we built the Octalysis AI Engagement Engine. It reads the motivation of each individual user from how they use the product, with no survey needed. It becomes more intelligent over time: the more of a person’s behavior it sees, the better its read of what drives them, including when that changes as they move from first use to mastery. With it, a product can shape the challenge it sets, the social cues it shows, the rewards it offers and the way it speaks around the actual person. The same assistant can stretch the user who wants mastery and reassure the user who wants to belong. The Engine does not replace product design. It pays off when the experience has meaningful options to adapt. Our article on the AI Engagement Engine explains what it does and why.
Yu-kai names the same opportunity. Deep personalization used to require human coaches, mentors or designers, and AI removes that constraint: “Used correctly, it’s also the tool that can execute the fix at a scale human designers alone could never reach.”
The next step is an agent, and eventually a whole journey, that adapts to the person. I describe that direction in gamification in the age of agentic AI, including the condition that people can always correct the agent when it has misread them.
Is This Just Change Management?
It is a fair question, and part of the answer is yes. Training, integration and manager support all matter, and they top the list in McKinsey’s survey for good reason. A person who does not know how to use a tool will not use it. A tool that does not fit the workflow gets skipped.
Change management usually leaves one question open. Why would a person want to use the tool once the training ends and nobody is watching? That is a motivation question, and it is the one Octalysis is built to answer. In practice the two work together. Training gets someone to the first win, and the design of the experience decides whether the person keeps taking the Desired Action afterwards. I make the full case in AI adoption is a motivation design challenge.
How We Approach AI User Engagement in a Project
Every Octalysis design project starts with a Strategy Dashboard that connects business metrics, users, Desired Actions, feedback mechanics and incentives. The behavioral design has to lead to an uptick in a business metric. More activity alone does not count.
The metric has to be sharp. “AI adoption” is too vague, so we refine it until it names something the business cares about, such as claims handled without rework or proposals sent within a day. In almost all of our projects, clients discover at this stage that they were designing for the wrong thing and need to revise their priorities. Finding that out before anything is built reduces the risk of the whole program.
Then we get specific about the Desired Actions for each role and lay them out in the order people meet them. We also look at what pulls people back to their old tools and habits, and work with that pull.
From there we walk through the four phases and look for the thin one. In many rollouts it is Scaffolding. People got a decent first session, and nothing changes after that. Our behavioral design and gamification services page describes the areas where we do this work.
How to Measure AI User Engagement
Seats and prompts measure access. To measure engagement, follow the definition: did people choose the AI, do they use it consistently, and do they value the experience?
Gallup’s split is a good place to start. It separates people who use AI in their role at all from those who use it a few times a week or more and those who use it daily. Next to that, I would track six measures:
- Voluntary return: the share of users who come back without a notification or a manager asking them to, at a frequency that fits the activity.
- Breadth and depth: whether a person’s use spreads to more kinds of tasks and gets more sophisticated over time.
- Time to first felt success: how long a new user needs to finish something they value.
- Ownership: the share of users who customize the tool, build something with it or share results with a colleague.
- Net time saved after rework: the time AI saves once the fixing is subtracted. A third of AI users in the Workday survey lose at least half of it.
- Felt ownership: Yu-kai’s second number. Ask people whether the work still feels like theirs. If that answer trends down while usage trends up, the work itself needs redesign.
The first four come from product data, and each maps to a phase, so you can see where the experience loses people. Splitting all six by level, from executives to individual contributors, shows whether leaders and their teams are having different experiences.
When Engagement Is the Wrong Goal
Engagement metrics can mislead. A product that holds attention through pressure and compulsive loops can post excellent session numbers while leaving people worse off.
Black Hat design covers Core Drive 6: Scarcity & Impatience, Core Drive 7: Unpredictability & Curiosity and Core Drive 8: Loss & Avoidance. It creates urgency, and it has a place. Yu-kai’s article on White Hat and Black Hat gamification explains the distinction. The risk lies in how heavily a design leans on it. The less in control people feel, the more they look for a chance to drop out. In a product people rely on for real work, our practice is to use the threat of a loss sparingly and always give people a straightforward way back, as the Navo Orbico design did for a rep after a weak period.
AI companions show the risk clearly. Yu-kai cites a 2026 randomized controlled trial in which people paired with a human peer ended up less lonely than people paired with a chatbot. He warns that an AI that displaces human contact “is optimizing for engagement metrics while quietly starving the Core Drive that actually sustains people.”
The better question is whether people become more capable by using the product. For a coding assistant, that means developers who understand more of their own systems. For a support tool, it means agents who resolve harder problems because the tool handles the routine ones.
Questions to Ask Before Your Next AI Product or Rollout
If you are planning an AI product or rollout, or improving one where use has stayed shallow, these are the questions we start with:
- Business metric: Which metric is the AI meant to improve, and which specific actions by which people move it?
- Discovery: Why would someone try this on their own? If the answer is that they were told to, expect lukewarm use.
- Onboarding: What does a new user finish in the first session? If the answer is that they explore, there is no first win yet.
- Scaffolding: What is different about the experience in week ten compared with week one? What can the user decide, shape or correct?
- Endgame: What can your most skilled users become? Status among peers and the chance to build things others use are the usual starting points.
- Managers, for internal rollouts: Do the people who lead teams use the tool themselves, and do they talk about it?
- Measurement: Are you tracking completed work, voluntary use and felt ownership, or only licenses and prompts?
When a company brings us an AI product or rollout, we diagnose which Desired Actions matter for the business metric, what the current experience gives each group of users for taking them and what it costs them, and which Experience Phase is thin. Then we design the choices, feedback, progression and social interaction that make people want to take those actions, and we define how to test the result. Leaving this unresolved has a cost. The licenses renew, people keep using the AI at the minimum their role requires, the rework stays hidden, and the business metric the investment was meant to improve does not move. If you recognize that situation, Contact Us and tell us which action you need people to take with your AI.
Frequently Asked Questions
What is AI user engagement?
AI user engagement is the degree to which people choose an AI product, assistant or agent for the actions a business depends on, keep using it on a consistent basis, and are enthralled by the experience it gives them. Consistent means every time the activity comes up, which can be all day for a coding assistant and a few times a year for an insurance claim. It is measured by completed Desired Actions and voluntary use. Provisioned seats and prompts sent only show who has access.
Why does AI use often stay shallow after a rollout?
Most AI tools are designed for functionality, and the experience around them is left undesigned. New users meet an empty prompt box with no first win. Regular users get the same experience in week ten as on day one. Experienced users have no role to grow into. Mandates and usage targets raise the figures for a while and produce minimum effort.
What is the difference between content personalization and motivation personalization?
Content personalization decides what a user sees next, based on what they and similar users did. Motivation personalization adapts the experience to why the person acts. Two users can stop at the same step for opposite reasons, one wanting a harder challenge and the other wanting company, and the same nudge will miss at least one of them.
What does the Octalysis AI Engagement Engine do?
The Engine reads the motivation of each individual user from how they use the product, with no survey needed. It becomes more intelligent over time: the more of a person's behavior it sees, the better its read of what drives them. The product can then adapt its challenges, social cues, rewards and tone to the person.
How do you measure AI user engagement?
Track voluntary return at a frequency that fits the activity, the breadth and depth of each person's use over time, the time a new user needs to reach a first success they value, and the share of users who customize, build or share. Add net time saved after rework, and ask people whether the work still feels like theirs.
