AI products don't usually lose users over the thing everyone worries about. In feedback from AI-native software products, the model itself is rarely the reason people leave. Its intelligence, its accuracy, even its tendency to make things up, all matter less than the ordinary machinery wrapped around it: work that never arrives, meters that charge for failures, controls users can't reach, and cancellations that don't stick.

This report reads that pattern directly from customer feedback. It draws on feedback from AI-native software products (products whose core value is AI) across consumer, prosumer, developer, and enterprise contexts, gathered from support tickets, app reviews, community and social posts, surveys, and sales-call transcripts. Every record was classified with Enterpret's Adaptive Taxonomy and counted, so every percentage below is a counted share of a stated slice of feedback rather than an estimate. Enterpret is customer context infrastructure: it unifies the voice of a company's customers across every channel and turns it into a structured, counted Customer Context Graph that teams and agents can act on.

What follows is a short account of what users actually want, the six findings behind it, and the evidence for each. The evidence sits in three sections, on where AI products break, what users ask for, and why they stop trusting, and the report closes with what the pattern means for product teams.

The six findings

1. The model is the top complaint only when its output is the product. Where the deliverable is the model's output, such as transcribed words, generated voices, or drafted professional analysis, model quality leads the complaints. Everywhere else, the chassis around the model, meaning the interface, integrations, and billing, leads instead.
2. The most widespread failure is the AI action that produces nothing. Failed or stalled generations, jobs that never complete, artifacts that never arrive. At its peak this single failure covers up to 17% of a product's negative feedback in a month, and in credit-metered products it arrives already paid for.
3. Users ask for dials, not a smarter model. Off-switches, approval gates, and model pinning form the largest product-ask family in the data, and approval-gate requests outnumber refusals of agent autonomy five to one. A generically smarter model never becomes a request theme.
4. Consumption metering turns model failures into fairness disputes. Where credit metering exists, a failed generation that charges anyway becomes a billing dispute. At the most metered end of the data, being charged for failed generations covers 32% of negative feedback.
5. Trust loss is enacted financially, not voiced. Users almost never say the trust is gone; departure statements sit at or under 1% of records. Refunds, cancellations, and chargebacks carry the signal instead, and charged-after-cancellation alone reaches 28% of one product's negative feedback.
6. Hallucination is a minority complaint, but low complaints are not low incidence. Fabrication is materially voiced only where the AI writes in the user's own voice or carries professional liability, peaking around 14%, and a record-level audit found it in under 3% of sampled negative records. Complaints only capture what users catch. Where nobody verifies the output, a low rate says little about how often the model is actually wrong.

Where do AI products break, according to their users?

Findings 01, 02, and 06.

Mostly below the model. Negative feedback about AI products lands in six complaint layers: model output quality, product surface and UX, speed and cost, trust and safety, workflow fit, and non-AI operations such as billing, accounts, and infrastructure. These layers are a way of classifying where a complaint lands, the axis of Fig 1 below, and they are not the same as the six findings above. Model quality dominates only where the model's output is the entire deliverable. When users buy transcribed words, generated voices, or drafted professional analysis, the model carries the complaints. When the AI sits inside a broader product, the chassis absorbs them: the interface, integrations, billing, and infrastructure wrapped around the model.

The praise tells the same story from the other side. The model layer is polarized, praised and attacked along the same axis. Transcription accuracy shows up among the top praise themes and atop the complaint clusters at once, and in high-stakes professional products, output integrity leads both sides, with roughly 4% of praise specifically applauding low hallucination risk. The plumbing draws almost no praise at all. On the most heavily metered products, what praise exists is mostly relief that a billing problem finally got resolved.

The heaviest recurring reliability complaint is the invoked AI action that produces no usable output: the dictation that returns silence and loses a five-minute ramble, the video generation that stalls for hours with credits already burned, the agent run that sticks, the export that reports success and never arrives. Users state plainly why this is worse than a wrong answer. A wrong answer can be judged and retried, while silence consumes work, money, and time and offers nothing to correct. Its most corrosive variant is unexplained failure. Error-less dead ends form named complaint clusters throughout the data, running as high as 7% of negative feedback at one product and 13% at another. Developers respond by asking for observability. Consumers respond differently: paid plus failed plus unexplained reads as scam.

The failure AI products are most famous for, fabrication, is a minority pattern in what users actually report. It is material exactly where fabricated language carries the user's own name or professional liability: high-stakes professional work (14% of negative feedback, the peak anywhere in the data), dictation-style use where the AI writes as you (up to 4%, including profanity landing in workplace messages), and voice agents speaking wrong facts to callers (about 2%). Elsewhere it is trace-level, and a record-level audit of sampled negative records confirmed it, with fabrication described in under 3%. Read the concentration carefully. Those places are exactly where users are equipped to notice a fabrication and pay for missing it, and where nobody verifies the output, a low complaint rate says little about how often the model is actually wrong. What dominates instead is mundane wrongness, meaning word swaps, irrelevant results, translations that drift, and edits that miss the instruction, along with its sharper cousin, disobedience. The AI that ignores an explicit instruction is a recurring complaint cluster in creative and agentic products, covering 9% of negative feedback where it runs largest, several times its fabrication share in the same feedback.

Again and again, users detected quality shifts after silent model changes and asked, in so many words, whether the model had been swapped. The largest such cluster covers up to 11% of one product's negative feedback. The request that follows is versioning: pin the model I validated, and tell me when it changes.

What do users ask AI products for most?

Finding 03.

The most consistent product ask in the data is control over what the AI does automatically: turn off auto-formatting, transcribe only the language I selected, let me approve the agent's plan before it acts, let me correct output in flight, keep the model version I tested. Control asks form the largest product-ask family in the data, and in agentic use, requests for approval gates outnumber explicit refusals of full autonomy five to one. Both are the same demand. Nowhere in the data do users ask for more autonomy without paired supervision asks.

Quality asks arrive as named instruments. Generic model-improvement demand never forms a request theme. Users ask instead for scoped tools that turn general capability into their specific reliability: custom vocabularies, glossaries and document-layout fidelity, trusted grounding sources and citation pinpointing, richer filters. Quality asks and control asks blur together here, because a lever is what makes the output feel like theirs.

The buyer's asks and the user's asks pull apart. The loudest enterprise request clusters, such as deep workflow automation, training programs, and governance, live overwhelmingly in sales-call transcripts. Where sales calls dominate the intake, roughly 95% of asks are buyer or rollout voice, the flagship automation theme collapses by 98% once user voice is isolated, and a companion integration theme collapses by 93%. What active users ask for instead is mechanics: better source pinpointing, formatting fidelity, faster responses. Automation depth is a real purchase criterion, but the people who use these products every day ask to supervise and repair what already exists.

Metering adds a money-shaped family of asks: cost transparency before and after every action, drafts and previews that don't burn the meter, and hard spend caps that actually stop spend. Some users offered their own remedy, accepting restored credits in place of a refund. The ask underneath is fairness, not a better model. Stop charging blind for failures.

Why do users stop trusting AI products?

Findings 04 and 05.

Start with how rarely it is voiced. Users almost never say the trust is gone. Complaints that announce a departure are vanishingly rare, at or under 1% of records throughout the data, and in one entire complaint set, distrust is never voiced at all. What appears instead, wherever a payment boundary exists, is money in motion. Refund, cancellation, and chargeback activity runs far above any trust phrasing, and it is often the largest single share of negative feedback where support tickets double as the refund path. Three escalation behaviors travel together: scam and fraud accusations (1.5 to 3% of negative feedback where present), bank chargebacks (3 to 6%), and legal or consumer-protection threats (near 1%). That language attaches to money that moved without consent or without value received. It essentially never attaches to an honest failure.

This is also where metering does its damage. Wherever credit metering appears, the same mechanism follows: the AI fails, the meter charges anyway, and a quality bug becomes a billing dispute. At the most metered end of the data, being charged for failed generations covers 32% of negative feedback. At another metered product, a quarter of negative feedback is credit-economy complaints, including 7% that are explicit reimbursement disputes. The asks follow the meter. At one product, roughly one in five of all feedback records asks for itemized credit accounting, followed by drafts that don't burn credits.

The sharpest single trust-killer is the charge that outlives the cancel. The user took the trust-ending action, cancelled, and the product overrode it. At the extreme, this one theme covers 28% of a product's negative feedback, alongside users demanding written confirmation that their cancellation actually happened. At another product the same family covers 12 to 20% of negative feedback. Close behind sit refund paths that dead-end in automation, which is where users volunteer compromises, such as restored credits over cash, just to be acknowledged.

The pattern has two instructive edges. Where the stakes are professional liability, trust really does live at the output. Fabricated sources and citation integrity carry the trust story, complaints about leaving barely exist (under 1% of records), and the behavior that matters, professionals double-checking output before they act on it, shows up in nearly 4% of all feedback. All of that verification talk sits in sales-call conversations, and none of it is framed as a complaint, so it stays invisible to every complaint dashboard. In free products the exit looks different again: trust leaves as competitor switch-talk, and refund requests barely register. Where money can't carry the signal, it moves to verification behavior or straight to the exit.

Repetition tends to come before the exit. Where departure talk can be traced to a trigger, it pairs with repeated or inconsistent failure and with billing violations, and users often state a condition first: "if this isn't fixed soon, I'm out." Those if-then messages are the most actionable churn-risk surface the study found. On this evidence, trust rarely dies at the first failure. It dies when the failure repeats, gets billed, or gets stonewalled.

What this means for AI product teams

Six things follow for product teams, and none of them is "make the model smarter."

The first is to treat "nothing came out" as a first-class reliability metric: track the completion rate of invoked AI actions, and the share of failures that at least return an error message. Silence is the most widespread failure in the data, and when it is unexplained it converts directly into scam perception in consumer channels. The second concerns the meter. Any meter that charges for failures will manufacture fairness disputes, so refund the failed generation automatically, itemize usage, price an action before it runs, and offer drafts that don't burn credits. On metered products the refund queue is effectively the QA process, because every billing dispute in it started life as a quality bug.

The third is that every automatic behavior needs a dial: an off-switch, a scope, a preview, an approval step. In this data users ask to steer before they ask to delegate more, and they already name the instruments they want, from custom vocabularies and glossaries to model pins and approval gates. The fourth is to version models out loud. Announce changes, offer pinning, and publish a changelog, because users noticed silent model swaps throughout the data and asked for exactly this. It is a communication fix available long before the next capability fix. The fifth is to treat the cancel button as a trust feature: make it stick, and put a human at the end of the refund path. Charging after a cancel is the single most reliable generator of scam accusations, chargebacks, and legal threats in the whole study.

The sixth underwrites the other five. Measure trust in money and behavior rather than vocabulary, watching refund and cancel rates, chargeback language, competitor mentions, and verification talk in call transcripts. All of it depends on being able to see the signal in the first place, which means feedback unified across support, reviews, community, and calls, and classified so that "wrong output," "charged for failure," and "can't turn it off" stay distinct and countable.

Methodology

The study analyzed customer feedback from AI-native software products, meaning products whose core value is AI, across consumer, prosumer, developer, and enterprise contexts, and across support tickets, app reviews, community and social posts, surveys, and sales-call transcripts. Feedback was classified with Enterpret's Adaptive Taxonomy. Complaint, request, and praise themes were counted from the classified records, and every percentage in this report is one of those counts expressed as a share of the slice named alongside it, whether that is negative feedback, all feedback, or praise. The findings were then stress-tested. Layer rankings were recomputed inside and outside the support channel, and enterprise ask themes were re-counted with sales-call transcripts removed, to separate buyer voice from user voice. Results are reported at the level of patterns.

A few limits are worth stating plainly. Feedback captures what users choose to say, so these counts describe voiced feedback and make no causal claims. Low complaint volume is not the same as low incidence: where users can't verify an output, as in Finding 06, quiet in the data says little about how often the model is actually wrong. Users who never write in, including those who leave silently, are under-represented by design, so voiced departure counts should be read as a floor rather than a full measure of churn. The report also leaves out corpus size, company counts, and the exact analysis window on purpose, which means the percentages are within-slice shares rather than population base rates, and they are not meant to rank one product against another.

FAQ

What do users actually want from AI products?

Reliability of the basic transaction: output that actually arrives, no charges for failed attempts, controls over automatic behavior (off-switches, model pinning, approval gates), named quality instruments like vocabularies and glossaries, and remedy paths that actually work, such as cancels that stick and refunds a human can approve. Category-level wants like "accuracy" and "trust" are real but too generic to act on. The feedback record shows users asking for these specific, buildable things.

What is the most common complaint about AI products?

The AI action that produces nothing at all: failed generations, stalled jobs, exports that claim success and never arrive. It is the most widespread failure in the data, and its worst variant is failing without an error message, which in consumer products converts directly into scam perception when the failed attempt was also billed.

How often do users complain about AI hallucinations?

Far less than its reputation suggests. Fabrication is materially voiced only where the AI's language carries the user's own voice or professional liability, peaking around 14% of negative feedback there and running at 4% or less everywhere else, and a record-level audit of sampled negative feedback found it described in under 3% of records. Mundane wrongness and instruction-ignoring dominate instead, though low complaint volume is not the same as low incidence where users can't verify the output.

Why do users stop trusting AI products?

Repeated failure plus a money violation. Trust loss is enacted rather than announced: users rarely voice it, while refund, cancellation, and chargeback volume is the dominant trust signal wherever a payment boundary exists. The single sharpest trigger is being charged after cancelling. In high-stakes professional use the pattern shifts, because trust lives in output verifiability and users quietly double-check everything, a behavior visible only in call transcripts and never in complaint metrics.

What features do users request most from AI products?

Control instruments (turn off auto-behaviors, approve before the agent acts, pin the model version), cost transparency on metered products (itemized credit accounting is among the largest request themes in the data), integrations into existing systems of record in enterprise use, and flexible pricing. Requests for more AI autonomy come almost entirely from buyers evaluating products, while daily users ask for supervision and repair tools.

See this in your own feedback

Every pattern in this report came from feedback unified, classified, and counted with Enterpret. To track failures like "nothing came out," charge-after-cancel, or silent model swaps in your own product's feedback, book a demo.

Companion piece on why users churn: link once live.

Produced by Enterpret Research, from customer feedback across AI-native software products. Enterpret turns customer feedback into counted, cited, reproducible customer context.

‍