A synthesis of every AI-related insight we have. 21 Dovetail insights across 4 studies, 5 themes.
🧝 Elf Workshop · Mar 2026 · 9 interviewsSmart Selection survey · Sep 2025Resolution Boost studies · May–Jun 2025V1 usability · Aug–Nov 2025Lyssna concept test · Jun 2026
Why this matters now
As we shift toward an AI-first strategy, most of our products already include AI in some form. To orient ourselves well, we need to understand the relationship our users actually have with it : how much they want, where they draw the line, and what agency means to them. The more we factor that in, the closer we get to building experiences that feel personal and genuine to our users. This document is a foundation for that.
What the research says
The relationship users have with AI is nuanced. They're willing to delegate, but it depends on the task, the context, and the moment. Execution work (layout, sorting, photo selection) moves more easily than choices that feel personal. In gift contexts, even that shifts: effort carries meaning, and AI involvement can quietly reduce it. Different users also sit in very different places: some want to stay in control throughout, some want to steer from a distance, some just want to hand it off and come back to a draft. None of those profiles is wrong. What's consistent across all of them is that trust is the prerequisite, and it's built, or broken, in the small moments.
📌 A note before reading
These insights come from research conducted in the context of the Photobook product. While we are no longer pursuing that product, the findings remain highly relevant : they reveal how our users think about AI-assisted creation, what they trust, what they want to control, and where the real friction is. That understanding applies across any product we build that involves AI and personal content. Use these findings as a foundation for prioritizing directions, not as a photobook-specific playbook.
🧝 Methodology Elf Workshop + 3 other studies
🧝 Primary study · Mar 2026 · 9 interviews
The Elf Workshop
Photo book creation is a high-effort, emotionally invested process : and our users are doing significant manual pre-work before they even open the app. We needed to understand exactly where AI can reduce that perceived effort without undermining the sense of ownership that makes a photo book feel like theirs. This study set out to:
Define the Picta role and AI's expected job description as well as the user's position as "editor-in-chief"
Identify where helpful automation tips into unhelpful loss of control and ownership of the creative process
Surface the trust conditions users need before they will stop doing everything themselves
The core insight behind the method: if you ask people directly "which tasks do you want AI to do?", you get guarded, tech-anxious answers. The Elf Workshop replaces the term "AI" with a team of elves : talented and well-meaning, but they've just met you, so they know nothing about you or why you're making a photo book. This framing lets people respond to the actual question of delegation, without the defensiveness that the word "AI" triggers.
Step 1 : Reconstruct the user's reality
Before introducing the elves, ground the exercise in the user's actual behaviour. The interviewer asks the participant to walk through the last time they made a photo book, step by step. Every action is documented as an individual card on a shared FigJam board : including the invisible "pre-work" done before even opening the app (creating albums on their phone, starring favourites). These user-generated cards become the foundation for the sorting exercise.
Step 2 : Introduce the elves
Once the workflow is mapped, the scenario begins. The interviewer introduces the elves like this:
"Imagine you've just been assigned a team of well-meaning elves to help you build a photo book. They are in the room with you, waiting for instructions. These elves are incredibly talented but you have just met them : so they don't know you or why you're making a photo book."
Step 3 : Sort into the delegation spectrum
The participant drags each task card into one of three zones:
Mine : I do this myself, the elves don't touch it
Supervised : elves do the work, I steer and approve along the way · This is where the interesting insights live
Theirs / AI : total delegation, I just care about the result
📊 Other studies referenced
Smart Selection survey · Sep 2025
Post-launch survey on feature awareness, perceived usefulness, and satisfaction.
Resolution Boost studies · May–Jun 2025
Print products (Produits Imprimés). Usability and survey work on AI photo enhancement perception.
Lyssna concept test · Jun 2026
5-participant unmoderated test on the Scout "Your Life Illustrated Poster" concept.
Key findings at a glance
🔍 The appetite is real but conditional
What people will and won't delegate, and why
1Better AI serves more technical tasks, but personal ones stay owned
2People split tasks in two: what they keep, and what they hand over
3Not everyone delegates the same way : three user types emerged
4Delegation is also contextual: the same user can delegate differently
💡 They want AI that suggests, not AI that decides
How AI should behave in the product
7The final review: no one was willing to skip it
8What stops people from delegating isn't fear of errors : it's fear of not being able to fix them
9Users prefer AI that highlights good options over AI that removes bad ones
10The problem isn't time : it's the number of decisions that pile up
🤝 Three things build the appetite for AI
What actually builds trust and unlocks more delegation
12First impressions matter : the first AI result makes or breaks things
13AI that learns your taste earns more delegation over time
14Familiar AI from other apps builds comfort fast
15Privacy has to be resolved first : it's a blocker, not a footnote
1. Better AI serves more technical tasks, but personal ones stay owned
▾
TLDR When probed with the scenario: "what if AI (the elves) kept getting better with experience?" : the effect was uneven.
We identified three different types of tasks, and three different levels of delegation:
Identity tasks
Which photos matter, the narrative, captions. Non-delegable regardless of AI quality: rooted in personal meaning, not trust.
Style & aesthetic tasks
Colours, layout, overall feel. Users want to approve these every time : even when they trust AI to do a good job technically, style is personal enough that they still want the final say.
Technical execution tasks
Page layout, cropping, duplicates, photo adjustment. The most delegable: ready to hand over once AI demonstrates competence.
Once AI learns your preferences, execution tasks move further. Personal ones never will.
Strong evidence
Three clear levels of how much people will delegate
(1) Identity tasks: never delegated, no matter how good AI gets. (2) Style tasks: the user always wants to approve. (3) Technical tasks: happy to delegate once AI proves itself.
The identity limit is not about trust in AI
It is a personal boundary. Better AI will not change it. Improving AI moves technical tasks forward, but not personal ones.
🤨 So What?
HMW distinguish between tasks where trust can be earned through demonstrated competence and tasks where the boundary is permanent?
User Highlights / The Evidence
"If you're telling me they can learn and maybe they can learn my preferences too... or they get better at maybe suggestions." [Moved 5 cards to Theirs after this probe]
"I still want to do those. It's kind of hard to give up full control from that." [On captions: unchanged despite moving 5 other cards]
"I still want to have a sort of a loop to be like, okay, do I like it or not?" [On style: also unchanged]
2. People split tasks in two: what they keep, and what they hand over
▾
TLDR People kept tasks that felt personal : choosing which photos matter, writing captions, setting the narrative. They handed over tasks that felt technical : cropping, removing duplicates, placing text on a page.
But this divide doesn't just run between tasks. It runs inside them. Several participants split a single task in two: they'd write the caption themselves, but let AI handle where it sits on the page.
The product needs to support this level of nuance : not just "delegate this task yes/no", but "delegate this part of this task."
Strong evidence
The divide runs inside a task, not between tasks
Users often split one task in two: they keep the meaningful part and let AI handle the mechanical part. For example, they write the caption themselves but are happy to let AI place it on the page. A simple on/off AI toggle per feature won't match this.
The same task can mean very different things to different people
Sorting photos is "just putting them in order" for one user, and "the most creative step" for another. You cannot tell whether a task feels meaningful just from its name.
🤨 So What?
HMW support sub-task-level delegation so that meaning and execution within the same feature area can be handled differently?
💡 Implication / Recommendation
Design delegation at the sub-task level, not the feature level. A user who wants to write their own caption but let AI place it on the page needs both options in the same flow. Controls like "AI handles layout, I write the words" are more useful than a single on/off toggle per feature.
User Highlights / The Evidence
"They can do the mechanics as long as the emotional impact is not hindered."
"The subjective like flow and artistic pieces. Like, I want those to be left up to myself. But the objective stuff, like removing the blurry photos, doing the caption layout, like, I would be curious to see how AI could help me with that."
3. Not everyone delegates the same way : three user types emerged
▾
TLDR Each participant had a consistent delegation style that was a better predictor of their choices than the type of task itself. Three profiles emerged:
🛡️
Narrative Guardian
Prefers full control
High ownership across the board. Delegates only clear technical tasks. Sees the book as a personal expression that only they can control.
🤝
Supervised Collaborator
Prefers to stay involved
Prefers a partnership model. Happy for AI to propose, but always wants to review and confirm before anything is final.
📋
Brief-and-Review Commissioner
Prefers to hand it over
Gives instructions upfront and steps back. Trusts AI to handle most of it, then reviews the result at the end.
Strong evidence
A one-size AI default will frustrate some users and underwhelm others
A user's delegation style (Guardian, Collaborator, or Commissioner) predicts their choices better than the type of task does.
Trust actions, not words: what people say about AI doesn't predict their behavior
One AI-positive participant kept most tasks for herself. A self-described "control person" ended up delegating a lot. Asking "how do you feel about AI?" gives a misleading picture.
🤨 So What?
HMW make the experience adaptive to different types of delegation styles?
💡 Implication / Recommendation
Don't assume a single default AI level works for everyone. The open question is how to detect which type of user you're dealing with : we don't have a reliable signal for that yet. One direction: rather than asking about preferences upfront, prompt users with concrete choices early in the flow ("want to pick your own photos or let us suggest a selection?") and let their actual behaviour inform how much AI involvement to surface next. Personalise progressively rather than configuring once.
User Highlights / The Evidence
"My elves do not know my family, so I wouldn't trust them to pick what photos they like or not. I would want to have ownership of that decision."
"I would just prefer to get it all out on the table now and just go, let the elves do their thing and sit back and I guess wait, forget about the book until they come back to me." [Commissioner]
"I guess I like the AI as the suggestor of, you know, make recommendations, be a partner in designing this. But I'm still the one who is making the initial and end choices." [Collaborator]
4. Delegation is also contextual: the same user can delegate differently depending on the circumstance
▾
TLDR Several participants explicitly stated that their delegation preferences would change for a different type of project. Some participants noted that certain tasks would shift zones for a more narrative-driven project. Others said they would delegate more under time pressure or without a strong concept.
This means delegation is not a fixed user trait: it is a situational state modulated by project type, emotional stakes, and time pressure.
Strong evidence
The same person will delegate differently on different projects
Project type, emotional importance, and time pressure all change how much someone is willing to hand over.
Delegation preferences are project-level, not account-level
A user making a casual year-in-review may delegate almost everything. The same user making a memorial book may keep most of it.
🤨 So What?
HMW account for the fact that the same user needs different levels of AI involvement across different projects?
💡 Implication / Recommendation
A memorial book and a holiday recap are not the same emotional undertaking for the same person : and the same logic applies across products. The amount of AI involvement that feels right will shift depending on what the project is for. Don't lock users into a fixed mode : make it easy to adjust how much AI is doing at any point in the process, including mid-way through.
User Highlights / The Evidence
"I could see where it might fall into a supervised category if I was doing a different project."
"It would be nice if I have short time and I can get AI's help."
5. Layout is the task everyone wants AI to handle : and the most painful one to do manually
▾
TLDR Page layout achieved the strongest consensus in the study: no participant wanted to keep it entirely. It was also the most frequently cited pain point and the largest time-cost task.
In short: layout is the step that feels most like work : and the one people most want AI to handle first, so they can just review and tweak.
Strong evidence
Layout was the strongest consensus for delegation across all participants
No participant wanted to keep it entirely. Even the most controlling users were happy for AI to help here.
Layout is also the biggest pain point
It is the most-mentioned frustration and the step that takes the most time. High pain + high willingness to delegate = the clearest place to build.
🤨 So What?
HMW reduce the effort of layout without removing the user's sense of authorship over the final spread-by-spread result?
💡 Implication / Recommendation
Layout is the highest-value automation target. But don't just auto-generate and present a finished book : generate a draft they react to, spread by spread. The goal is to make layout go from the hardest part to the fastest part, without making the user feel like a passive observer.
User Highlights / The Evidence
"That's the part of the process that feels most like work, because I spend a lot of time, like, choosing a template, trying to put photos in there, and then realizing... Now let me go back and start again."
"Having to do it manually is kind of a pain. But it was really just like almost like graphic designing each page."
"I don't want layout to feel repetitive, but I also want it to sync with other pages. That also takes some time. That takes actually a lot of time."
6. Duplicate removal and photo adjustment are the clearest automation candidates
▾
TLDR Removing duplicates/bad photos and photo adjustment/cropping are the tasks most readily delegated to full automation.
Participants perceived these as technical, objective tasks with "right answers" that do not require personal judgment.
The key condition: users need to be able to get things back. Even the most delegable tasks require an easy undo.
Strong evidence
Closest to full consensus on automation in the study
Duplicate removal and photo adjustment are the tasks participants most readily delegated without hesitation. They perceived them as objective, technical tasks with "right answers" that don't require personal judgment : the kind of work where AI getting it slightly wrong feels recoverable.
Even here, sentimental value can override technical quality
P09's "I know this is a scanned photo from 1985, but that's what we have" is the edge case that matters. Technical quality and sentimental value can be in direct conflict. A blurry, low-res photo isn't objectively worth keeping : but it might be irreplaceable. Any automated removal has to account for this tension, which means making it easy to review and restore what AI removed.
🤨 So What?
HMW handle the tension between objective technical quality and subjective sentimental value when automating photo screening?
User Highlights / The Evidence
"I think it's obvious, like, hey, a blurry picture. Don't want to include that."
"It seems like it's a little more technical. I'm already choosing the theme. I'm already choosing the photo. This is just getting it in the right place and position on the page."
"I'm pretty sure that they would be much better than me and it would be done more efficiently."
7. The final review: no one was willing to skip it
▾
TLDR Every single participant kept the final review for themselves : including those who handed over almost everything else.
Even when the "AI gets better" scenario moved many tasks over, the final check never moved.
This is what makes delegation possible in the first place. People were comfortable handing things over earlier in the process because they knew they'd have one last look before ordering.
Strong evidence
The final review is what makes delegation feel safe
Participants delegated more freely during the process because they knew they'd have a final check at the end.
No one was willing to skip the final check, even after delegating everything else
The "AI gets better" probe moved many cards. The final check never moved. It is the most protected step in the whole process.
🤨 So What?
HMW make the final review step feel efficient and comprehensive so that users who delegated upstream feel confident they caught everything?
💡 Implication / Recommendation
The final review is not overhead : it's the product. A fast, clear review flow is what makes all the delegation upstream feel safe. If the review is clunky or hard to navigate, users will either redo everything manually or order something they're not happy with.
User Highlights / The Evidence
"I am still responsible for the final product... checking the whole book before ordering sounds like a very easy job to me. You know, where it's like I would like to have that control."
"Checking the whole book would obviously be a me task... things could have been miscommunicated... I want to go through and kind of compare. Does it match the image I had in my head?"
8. What stops people from delegating isn't fear of errors : it's fear of not being able to fix them
▾
TLDR For control-oriented participants, the decisive factor in whether to delegate was not whether AI would produce good results, but whether they could easily undo AI output if it was wrong.
One participant put it clearly: the flip point is not the error but the cost of correcting it. If fixing a mistake means starting over, they won't delegate. If it's a quick edit, they will.
The easier it is to undo, the more willing they are to try.
Strong evidence
For cautious users, it's not about AI quality : it's about whether they can undo it
If fixing a mistake means starting over, they won't delegate. If it's a quick edit, they will.
The real question is: how easy is it to correct AI output?
How easy the correction is determines whether people try AI features at all.
🤨 So What?
HMW make it easy enough to review and correct AI output that users feel safe delegating in the first place?
💡 Implication / Recommendation
Reduce the perceived cost of trying. The barrier to delegation isn't "I don't trust AI" : it's "I'm not sure I can fix it if it goes wrong." Easy undo, clear edit paths, and visible correction options lower this barrier more than improving AI quality alone. This applies throughout the process, not just at the final review step.
A critical tension: review must be present but must not become its own burden. If reviewing AI output feels like more work than doing it manually, users will revert to doing it themselves, collapsing the delegation model.
User Highlights / The Evidence
"Automation is good when you know that you won't have to restart."
"If I know that I can easily review and adjust the results. Like if it's easy and flexible, if I know that I can edit anything without starting over, I'll be very happy to delegate."
"If I felt like they completely got it wrong, I would just take it into my own hands. But if they had some good ideas, then I might continue to collaborate."
9. Users prefer AI that highlights good options over AI that removes bad ones
▾
TLDR A consistent preference emerged across participants: AI that surfaces good options is more welcome than AI that removes bad ones.
This was observed most clearly in photo selection and duplicate removal : participants wanted AI to highlight rather than discard.
The reason is agency: highlighting lets the user say no. Removal forces them to go back and recover what was lost.
Strong evidence
Highlighting keeps users in control; removing takes control away
When AI suggests rather than acts, users feel they made the choice. When AI acts first and users have to undo it, the experience feels like a loss of control.
"Suggest and confirm" feels different from "act and undo", even if the end result is the same
Who initiates the action changes how users feel about it.
🤨 So What?
HMW frame AI contributions as additive suggestions the user accepts rather than autonomous decisions the user must reverse?
💡 Implication / Recommendation
AI will remove and discard, and that's part of the value. What matters is making those changes visible and reversible. If a user can see what was removed and get it back easily, the action feels safe. If it happens silently with no way back, it feels like a loss of control even when the output is good. Surface what changed, and make undoing it a real option : Framing also matters: "here are the best photos" lands better than "we removed 23 duplicates" users react more strongly to what was taken than to what was kept. Also see: loss aversion
User Highlights / The Evidence
"I wouldn't necessarily want them to choose to discard as much as I'd want them to highlight ones that are good."
"Highlight for me... these are photos that are blurry. Do you want us to kind of clean it up?... Give me feedback, I guess, based on the photos."
"I still want to be able to have the option to see which ones they deleted or removed."
10. The problem isn't time : it's the number of decisions that pile up
▾
TLDR Several participants had photo book projects sitting unfinished for months : not because they didn't have time, but because they didn't know where to start. Hundreds of photos, hundreds of small decisions, all at once.
What they want from AI isn't just speed. They want to stop feeling stuck.
The real value proposition is not "saves time" : it's "gets you unstuck."
Strong evidence
The problem isn't time, it's the number of decisions
One participant's project sat unfinished for months, not from lack of time, but from not knowing where to start with hundreds of photos. AI that reduces decisions matters more than AI that saves minutes.
A good AI draft changes the task from "create" to "react"
Instead of making hundreds of choices from scratch, the user only needs to say yes or no to a draft. That is a much lighter kind of work.
🤨 So What?
HMW shift the user's experience from generating every decision from scratch to reacting to a draft that already embodies most of the low-stakes decisions?
💡 Implication / Recommendation
Start with a draft, not a blank page. The biggest relief AI can offer isn't speed : it's reducing the number of decisions from zero. A good first draft turns the job from "create everything" to "react and refine," which is far less draining.
User Highlights / The Evidence
"They are taking the mental load off of me."
"I would say it's more due to the volume of the photos because I have like a lot of pictures in my camera roll... it's kind of like trying to decide like which needs to be included and what doesn't."
12. First impressions matter : the first AI result makes or breaks things
▾
TLDR If the first AI output is good enough, people keep going. If it misses badly, they take over : and often don't come back.
There's no middle ground of "let's rework it together." It's more like a pass/fail test.
The first output is disproportionately important. A bad first impression can permanently lower how much someone is willing to delegate.
Strong evidence
The first AI output decides everything
If the first result is good enough, users continue and build trust. If it misses badly, they take over and often don't come back. There is no "let's try again three more times."
A bad first impression can permanently lower how much someone will delegate
Even if AI would have done better on the next attempt, users don't always give it a second chance.
🤨 So What?
How do we make sure a bad first AI result doesn't permanently put users off?
HMW capture enough user intent at the start of a project to produce a draft worth reacting to, without the briefing step itself becoming a burden?
💡 Implication / Recommendation
Treat the first AI output as a critical UX moment, not just a technical output. Quality-gate early AI results. If the first thing a new user sees is a bad layout or a wrong photo selection, trust drops permanently. Consider showing AI capabilities in a low-stakes, low-commitment context first.
User Highlights / The Evidence
"If I felt like they completely got it wrong, I would just take it into my own hands. But if they had some good ideas, then I might continue to collaborate."
"The more they did that, the more I'd like, okay, you guys clearly understand the style I'm looking for here."
"If I am using the tool again and again and if I am observing them and they are understanding me, then I can delegate."
13. AI that learns your taste earns more delegation over time
▾
TLDR Several participants said they'd be willing to hand over more : but only once AI had learned their taste. "Once it knows what I like, I wouldn't need to check as much."
This came up unprompted across multiple interviews. It's the condition that moves tasks from Supervised to fully delegated.
Learning taste over time is the clearest path to deeper delegation.
Strong evidence
AI that learns your taste over time unlocks more delegation
Once users feel the system understands them, they need less approval at each step. Tasks move from supervised to fully delegated.
Users want to see the learning happen, not just be told it did
Trust builds when users can observe the system getting closer to their style. A learning process that happens invisibly doesn't build confidence.
🤨 So What?
HMW make AI learning tangible to the user so they can see (and judge) whether the system is converging on their preferences?
💡 Implication / Recommendation
Personalization is an opportunity to deepen collaboration over time. The more the product learns about a user's style and taste, the more tasks they're willing to hand over. This is the clearest path from a supervised relationship to a truly collaborative one : not better AI in general, but AI that gets better at you specifically.
User Highlights / The Evidence
"I also want to make sure that the tool learns from my preferences over time and it gets closer to my style instead of being very generic."
"If they can learn and maybe they can learn my preferences too... or they get better at maybe suggestions."
"More things can be done by the elves completely once they understand what I like and my personal style, whatever my personal writing style."
14. Familiar AI precedents from other apps build comfort with delegation
▾
TLDR Participants who already used AI on their phones : Apple Photos highlights, iPhone auto-brightness, Google Photos : were noticeably more comfortable delegating similar tasks in a photo book context.
They didn't need convincing. The trust was already there from a different app.
"This is basically what my phone already does" is the fastest path to comfort : faster than any feature explanation.
Strong evidence
Existing AI experiences serve as trust anchors
If someone already trusts iPhone auto-brightness or Google Photos, they're ready to trust the same kind of help in a photo book. The trust is already built.
"This is like what my phone already does" removes all hesitation
Framing features around familiar AI experiences is faster than any explanation.
🤨 So What?
HMW tap into users' existing mental models from smartphone AI to reduce the perceived novelty (and therefore the risk) of AI in photo book creation?
💡 Implication / Recommendation
Reference familiar AI touchpoints when introducing features. "Like Google Photos, but for your book" removes novelty anxiety instantly. Anchoring new AI features to things users already trust (phone deduplication, auto-enhance) makes delegation feel familiar rather than risky.
User Highlights / The Evidence
"There are apps or features in my phone for getting rid of duplicate pictures. So that's kind of not really new to me. So that makes sense that someone else could just do it or it could be delegated."
"Even with my iPhone, right? When you, like, hit that, like, magical, like, light bulb button or whatever it is that changes your photo. Like, it's always better than what I do."
15. Privacy has to be resolved first : it's a blocker, not a footnote
▾
TLDR About half of participants brought up privacy before anything else. The concern wasn't about AI quality : it was about what happens to their photos. Especially photos of children.
For these participants, privacy is a blocker: if it's not resolved first, no AI feature matters.
The questions they want answered: Are my photos used to train AI? Are they shared? When are they deleted?
Strong evidence
Privacy is a yes/no question, not a sliding scale
Either users are satisfied on privacy before they start, and normal trust-building applies. Or they're not, and no amount of good AI output will matter.
Vague privacy messages won't work: people have specific concerns
They want to know: will my photos be used for training? Will they be shared? When will they be deleted? Answer these directly.
🤨 So What?
HMW address the privacy gate early enough that it does not become a barrier to experiencing the product's core value?
💡 Implication / Recommendation
Resolve privacy before asking for delegation. If a user hasn't decided how they feel about their photos being used, they won't hand over a single task. Address it clearly, early, and specifically : not with generic "we take privacy seriously" copy, but with concrete answers to "what do you do with my photos?"
User Highlights / The Evidence
"I would be comfortable to do that as long as I know that my materials is not used for training beyond what I allowed to be used... it depends on how the data will be handled."
"If I'm uploading images of my kids, are those images going to be used or shared in some way?"
"I would definitely want to know that my pictures are not sent to a third party and that they might be deleted at a specific point after I ordered the book."
16. Captions are consistently kept : because the words have to sound like them
▾
TLDR Captions were the most protected task in the whole study. Nobody put them in Theirs/AI. Most kept them entirely in Mine.
It's not that they think AI can't write. It's that the words have to sound like them. The captions carry private context and personal voice that AI simply doesn't have access to.
That said, there's an opening: some participants were happy to get AI help with short, functional captions : just not the personal ones.
Strong evidence
Captions are the "creative identity": the minimum users will never give up
The reason is not that AI can't write well. It's that the words should sound like the person who made the book. That's a personal choice, not a quality issue.
There is a gap though: users do struggle with short, functional captions
One participant wanted help writing witty one-liners but didn't want AI writing her full personal captions. The line runs between "personal voice" and "functional label."
🤨 So What?
HMW respect the creative identity of personal voice while addressing the real friction users feel around low-stakes, functional caption writing?
💡 Implication / Recommendation
The bar users set is clear : AI can enhance, not replace. For captions, that means AI handles the context : date, location, who's in the photo : while the user fills in the meaning. Framing it as assistance ("what do you want to say about this moment?") reinforces that the voice stays theirs.
User Highlights / The Evidence
"I want my own voice to be shared. It's my memories, you know, so that would be my own thing."
"They probably don't know like my heart and my mind."
"I don't want it to be kind of vague or sound like it's AI language, I guess."
17. Even when AI gets style right, people still want to approve it every time
▾
TLDR Style was almost always placed in Supervised : and it stayed there even when participants moved other tasks to Theirs/AI.
One participant who handed over almost everything still kept style: "I still want a loop to be like, do I like it or not?"
Better AI doesn't change this. Style approval isn't about trust in the output : it's about wanting to be the one who decides.
Strong evidence
Style needs ongoing approval, not just a one-time setup
Even when users said AI would get better at style over time, they still wanted to approve each result. Competence alone doesn't remove the need for aesthetic sign-off.
Style approval is permanent, not a phase that goes away
Unlike technical tasks, aesthetic choices seem to need human approval every time, not just until AI "earns" the user's trust.
🤨 So What?
HMW support aesthetic decision-making in a way that respects the perpetual need for approval without making it feel burdensome?
User Highlights / The Evidence
"I still want to have a sort of a loop to be like, okay, do I like it or not?"
"That's just the sort of thing I like to choose for myself. I find that enjoyable."
18. For gifts, AI output that looks too good can actually hurt the gesture
▾
TLDR When a photo book is a gift, the recipient knows the creator. If the book looks too professional, it signals that AI made it : which can undermine the personal value of the gift.
This only came up for close relationships. For archive books or gifts to people who don't know the creator's style, nobody mentioned it.
In gift contexts, "better" doesn't always mean "more impressive." Sometimes it means "looks like me."
Limited, directional only
For gifts, AI output that looks too polished can actually hurt the gift
When the recipient knows the creator well, a book that looks too professional signals that AI made it. That reduces the personal value of the gift.
This concern only comes up in close personal relationships
It didn't come up for archive books or gifts to people who don't know the creator's style.
🤨 So What?
HMW ensure AI assistance enhances the creator's own expression rather than replacing it with a visibly higher-polish output?
💡 Implication / Recommendation
In gift contexts, "better" doesn't always mean "more polished." Consider offering a deliberate "personal touch" mode where AI assistance is lighter : rougher layouts, less perfect crops : so the result reads as handmade. The value of a gift book is often its imperfection.
User Highlights / The Evidence
"If I'm creating it for my friend and she knows me, she knows how I think and if I bring her something which is absolutely too good maybe for what I would have created, she would know it."
"As long as the tool enhances my thinking instead of replacing it, I am comfortable."
"I don't want it to be too generic, especially in the sense of a photo album."
20. Current tools are all-or-nothing. Nobody wants either extreme.
▾
TLDR Participants who'd tried auto-generated photo books (Shutterfly, Once Upon) described the same frustration: either you do everything yourself, or you hand the whole thing over to AI and get something generic.
Neither felt right. What they actually wanted was to specify the content : which photos go where, what the narrative is : and let AI handle the rest.
That middle ground is the actual opportunity : user intent at the start, AI execution in the middle (to varying degrees), human review at the end.
Strong evidence
Current tools force a choice: do everything yourself, or let AI do it all
The middle ground, where users set the direction and AI handles the layout, doesn't exist in any current tool. That is the gap.
Auto-generated books feel generic because the tool never asked what you wanted
The problem is not that AI produces bad output. It's that it never received any user intent to begin with.
🤨 So What?
HMW close the gap between "do it all myself" and "auto-generate everything" that current tools force users into?
User Highlights / The Evidence
"If I had literally been able to sort of like, say, I want page one to be these photos, I want page two to be these photos... that would be great. But, like, it's either all or none."
"Generating generic photo album, maybe like something. Because it's not personalized at all. It's basing the structures based on when the photos were taken."
21. Smart Selection ran without anyone noticing : including people who had used it
▾
TLDR Neither participant was aware that Smart Selection had run. One had no idea any automation had happened at all : despite having used it. The other had a vague memory but couldn't say what it had done.
Both went through photo selection manually anyway.
A feature that runs invisibly doesn't build trust : it just gets ignored, even when it works.
Limited, directional only
The fact that no one noticed is itself the finding
Smart Selection is one of our key features for making photo selection easier. But users went through it manually anyway because they didn't know the feature had run.
AI ran, did its job, and got zero credit for it
Same pattern as Resolution Boost: the feature worked, but no trust was built because users never saw it happen.
Photo selection is high-effort and genuinely entertaining : until it isn't
Photo selection is consistently described as one of the most time-consuming tasks in photo book creation. But unlike layout or duplicate removal, users don't want it fully automated : they find entertainment in it. The process triggers nostalgia, emotional connection, and a sense of creative ownership. The structural problem is timing : photo selection typically begins before the user has committed to a service or invested in an outcome. The cost of quitting is near zero. The moment effort overtakes enjoyment : and it inevitably does at scale : there is nothing holding the user in. This is the Photo Selection Cliff : a danger zone where effort exceeds enjoyment but sunk cost hasn't yet accumulated enough to push the user through. The design goal is to extend the entertainment window past the effort tipping point, or to build enough investment early that the cost of quitting catches up before the user drops off.
🤨 So What?
HMW keep photo selection entertaining as volume increases : and how do we build enough commitment before the cliff hits?
💡 Implication / Recommendation
The entertainment window is a design lever. Surfaces that make selection feel like discovery (grouping by trip, surfacing a forgotten moment) extend engagement past the tipping point. Early commitment mechanics : saving favorites, starting a collection : build sunk cost before the cliff hits. The goal isn't to make selection faster : it's to make it worth staying for.
User Highlights / The Evidence
"Not at all." [when asked about awareness of automated layout features]
"It said, like, add up to 100 pictures or something. So I added as much as I could."
"I would want to decide what photos get included. I enjoy that process. Like, it's fun... it's more about like the initial nostalgia of looking through things."
P09
Quantitative
Quantitative
22. Users want AI to fix their photos but can't name the problem: "just make it better"
RQ: How do users perceive AI photo enhancement features like Resolution Boost and Smart Crop?
▾
TLDR Users have a felt sense that AI could improve their photos : but most can't say exactly how. The most common survey response to "what AI improvement do you want?" is just "make it better" (23.9%).
Resolution Boost has a strong emotional case: it lets people print photos they'd otherwise have to leave out. But it has a visibility problem. Users don't notice the warning, don't see what changed, and don't connect their satisfaction to the feature that caused it.
One signal worth watching: customers with gifting intent convert at 25% vs 16.2% for the next highest group. Gifting creates a committed deadline : "I still need a gift by X date" : that makes stopping less of an option.
Limited, directional only
The desired improvement is undefined for most users: no mental model for what AI can do
In open survey responses, the single largest category of desired AI photo improvement was "general improvement, just make it better" at 23.9%. Specific needs came up but less often: exposure and lighting (16.8%), sharpness and focus (7.2%), composition (5.7%). Users have a felt sense that something is improvable but can't name the problem. This is a blank page problem for AI feature design: the user wants the output without being able to specify the input.
Emotional use cases are the strongest value proposition: "save the photos that matter most"
The highest perceived value for Resolution Boost was the ability to print emotionally important photos that would otherwise be too low-quality to use: old scanned family photos, early smartphone photos, blurry-but-irreplaceable moments. The emotional attachment that drives users to print low-quality photos at any cost is what makes Resolution Boost a strong acquisition and retention lever. The technical message ("higher resolution") is the weakest framing. The emotional message ("save the photos that matter most") is the strongest.
Users don't notice or attribute the improvement: the feature ran invisibly
In Resolution Boost testing, users were satisfied with the final print quality but didn't connect that satisfaction back to the feature. The warning that flagged low-quality photos wasn't memorable enough to register. The before/after review screen caused confusion. Even users who interacted with the feature didn't recall doing so after the session. Same invisibility pattern as Smart Selection, AI works, attribution fails.
Larger baskets associate with higher perceived need of AI enhancement
Quantitative data showed respondents with larger photo selections reported higher perceived need for AI quality improvement. Quantitatively this makes sense, the manual curation cost of 200 photos is much higher than 30, and AI feels more necessary. This suggests AI enhancement positioning should be calibrated to basket size: for users with large collections (holiday trips, multi-year archives) the value message is strongest. For small, curated selections, it's less relevant.
Central tendency skews low: AI should be positioned as assistive, not broadly transformative
Survey distribution for AI photo improvement skewed toward occasional use rather than broad enhancement. The dominant mode was: identify and fix the photos that need it (blurry, underexposed, low-resolution) rather than improving every photo across the board. "AI should fix the ones that need it" was the median expectation. This means positioning Resolution Boost as a safety net for borderline photos (not a blanket quality upgrade) better matches actual user expectations. Overpromising broad improvement creates disappointment.
Face-related improvements are a top-tier outcome for those wanting AI-driven enhancement
When asked to rank desired AI improvement outcomes, face-related results (sharp, clearly recognizable faces) came out on top. This connects directly to the emotional use case: photos that matter most are photos of people, and the failure mode that most frustrates users is a blurry or unrecognizable face. "Make sure every face is clear" is a more resonant message than generic quality improvement, and maps exactly to the photos users would otherwise have to leave out of a printed book.
🤨 So What?
AI photo improvement has real demand and emotional resonance, but the product has to make the value visible and attributable. The emotional use case ("save the photos that matter most") is a much stronger message than any technical description of what the feature does. And the same invisibility problem observed in Smart Selection shows up here: AI acts, user doesn't notice, no trust is built.
*HMW make the value of AI photo improvement visible enough that users notice and remember it?*
User Highlights / The Evidence
Open survey responses on desired AI photo improvement: "General improvement (just make it better)" (23.9%. Exposure and lighting) 16.8%. Sharpness and focus, 7.2%.
23. For some products, buyers need to know a real person made it
RQ: Does the AI authenticity concern transfer to non-photo-book products? (Scout: Your Life Illustrated Poster)
▾
TLDR In a concept test on the Scout "Your Life Illustrated Poster" (Lyssna, Jun 2026, n=5), the gift framing landed without prompting : anniversaries, retirements, grandparents, Mother's Day all came up on their own.
One participant named AI vs. human artist as the single most important factor in her purchase decision.
The perceived value of the gift is the effort behind it. AI involvement challenges that.
Limited, directional only
AI vs. human artist is a purchase decision question in illustration-adjacent products
When asked what information would most affect her decision, one participant immediately named AI involvement, not price, not shipping, not customisation options. Her reasoning: this is a product that involves illustration, a category she associates with human craft. Knowing real artists are behind it would increase her trust. She even connected the brand's print history ("printing since 2015") to human involvement as a trust signal. This mirrors the Elf Workshop pattern where authenticity concerns were more salient in gift and craft contexts.
"The effort is THE gift" pattern replicates across products
The same participant also said she wouldn't buy this for herself, it would feel "too conceited, like a look how good my life has been gift." This is the same pattern identified in the PB Explo Éditeur research: in personal gift contexts, the object says something about the creator's investment of effort, and an AI-made gift reduces the perceived value of that investment. The pattern appears across different products. It may be a general property of AI in gifting contexts, not specific to photo books.
The "how is it made" question is the main comprehension gap
Two other participants asked about cost and the illustration process, not from distrust, but curiosity. The concept page answers "what" clearly. It doesn't answer "how." In craft-adjacent categories, "how" is a trust question as much as a curiosity question. Answering it transparently may close both gaps at once.
🤨 So What?
Early signal, not confirmed — but it adds to a pattern. In products that feel craft-adjacent (illustration, personalised gifts, photo books), some buyers weigh AI involvement as a quality signal. Being transparent about what AI does and what human craft brings may need to become standard across new Picta product pages, not just this one.
HMW communicate the role of AI and human craft in our products so it builds confidence rather than raising doubt?
💡 Implication / Recommendation
Don't hide the creative workforce behind the product : lean into it. Users in craft-adjacent categories respond to knowing that real human expertise is in the loop. That's a value proposition, not a disclaimer. The AI + human craft combination is something worth naming explicitly on product pages, not something to downplay out of fear it raises questions.
User Highlights / The Evidence
"How much AI is used in this product and are there real artists behind it. If they've been doing it since 2015 I feel like there would be less AI involved and I would have more faith if there were actual artists involved."
There's no single answer to how users want to work with AI : and that's the point.
Results are more nuanced than a one-way street. People have a delegation style, but it shifts depending on their emotional state, the project, and the context. That means we shouldn't think of delegation as a fixed user setting. Instead, we should prioritize tasks that are closest to repetitive execution and always think about tasks at the sub-task level: which parts carry personal meaning, and which are just mechanics. The mechanics are what we can safely hand over by default. The meaningful parts should never be assumed.
Some tasks move easily : layout, sorting, photo selection. Others : the ones tied to meaning, voice, or emotional weight : stay in the user's hands regardless of how capable AI gets. And the same person who delegates freely on a quick holiday recap will want full control on a memorial book. Context, time pressure, and what the book is for all shift the line.
There's a useful shortcut buried in the contradiction, though. Layout was the most consistently delegated task across the board : not because everyone felt the same way about it, but because the cost of doing it manually was high enough for almost everyone. That's the practical entry point : start with the tasks where the effort-to-meaning ratio is most evident.
The catch is that "willing to try" isn't the same as "fine with anything." Even on layout, the motivation differs:
One person delegates because they find it technically hard
Another because they're time-pressed
Another because they just want a starting point to react to
The task is the same. What they need from the handover is different. That's where the nuance lives : not in which tasks to automate, but in how those automations are presented, how much visibility users have into what ran, and how easy it is to course-correct.
01
Design for different levels of collaboration
Not all users want the same level of AI involvement. Guardians, Collaborators, and Commissioners all exist in our user base. The product should accommodate all three : not force a single default delegation level on everyone.
02
Give flexibility
Delegation preferences change by project, mood, and time pressure. Someone making a memorial book delegates very differently from the same person making a holiday recap. Don't lock users into a fixed AI setting : let them adjust per project, or even mid-process.
03
Make AI affordances visible : framing and engagement
When Smart Selection ran, users didn't notice : and went through photo selection manually anyway. A feature that works invisibly doesn't build trust, it just gets ignored. What stops most people from delegating more isn't fear of bad output : it's fear of not being able to fix it. Surfacing what ran, what changed, and what's still in their hands is what makes delegation feel safe enough to try.
How AI actions are framed matters as much as what they do. Keeping users invested long enough to feel the value is itself a design challenge. Quantitative science offers a useful lens here : progress mechanics, sunk cost design, and commitment devices are worth exploring as we design AI-assisted flows.
04
Give agency
Design AI as additive, not subtractive. Highlight rather than remove. Suggest rather than decide. The direction of the action matters: users accept AI that adds options and resist AI that takes them away : even when the end result is the same.
And give users a way to express intent before AI acts. Auto-generated books feel generic not because the AI produced bad output : but because it never received any user intent to begin with. A brief moment of intent-capture ("what's this book about?", "who's it for?") changes the entire relationship between user and output. It shifts the experience from "AI did something to my photos" to "AI helped me make something."
05
Build trust with the small things
Trust is not a nice-to-have. It's the prerequisite for any delegation to happen at all. This means: first impressions matter disproportionately, recovery options must be easy, privacy must be addressed upfront, and AI learning must be visible : not a black box. Every feature that builds trust unlocks more delegation. Every feature that breaks it sets things back.