Most products fail because they ignore the messy reality of how people actually use them. I help you create products that work by understanding the overall system, designing for  behaviour, emotions and people’s real context. 

Most importantly, I think about risk - always.

I’ll help you see what can go wrong to users and the business, so we can mitigate it and make informed trade-offs (rather than nasty surprises)

Here’s how I help you:

🤝 If you're a potential client:
→ UX Audits: analysing your product and identifying opportunities
→ Fractional design and research support that drives business, and user impact
→ Clarity on strategic direction and vision 


📕 If you're here for my content, I write about: 
→ How to consider the *overall* experience
→ Connecting design impact <> business impact
→ Identifying different types of risk (and mitigate it)
→ Designing for how people *actually* use things by considering context, behaviour and emotions 
→ Tips to help along your design processes

I'd love to hear about how I can help you. Drop me a message on LinkedIn, or send an email to: hello@ampersandco.studio

And if you don't have a project, I'm always down for a coffee chat!

Emily Anderson

Issue #1 of Beyond the Happy Path newsletter is live!

This one is all about Churn. It’s not the sexiest thing, I know

It's easy to assume that when people stop using our products and services it's because they *wanted* to. We focus on reasons like: they didn’t see value or it’s too expensive

But that’s not always true.

Sometimes people leave because they had no choice. They leave because they *have* to:

Maybe their ability changed and they can't use it anymore. Maybe they're struggling financially and can't afford it right now. Or maybe someone passed away.

In this issue I dive into:

➡️ Why we need to think about churn in terms of “wanted to” vs “had to” reasons (i.e voluntary, involuntary, and situational churn)

➡️ Real examples that can help to reduce churn

➡️ Ideas for how you can start considering it in your process

Incase you missed it, you can read the post in the link below (and subscribe for future issues too  👀)

https://lnkd.in/ef9CMEAE
---

It's very early days of the newsletter and I'm not going to pretend to have it all figured out. So if you've got any feedback of what you want to see more / less of, I'd love it here it!

Designer | Reducing risks to users and businesses | Founder, Ampersand | Speaker

Most teams are already using AI in the wrong ways for research—rushing to generate “insights” that collapse under pressure. I've run 1800+ hours of experiments to fix that.

I help product and design teams:

Make faster decisions by turning messy feedback into trusted insights in hours, not weeks.

Adopt AI faster (and smarter) than competitors with proven workflows that show what works and what doesn’t.

Learn the essentials for customer research—because data insights you can trust aren’t the same as what AI “gurus” in other spaces are teaching.

Stop wasting time on nonsense outputs and build bulletproof insights your team can act on without regret.

Here’s how we can work together:

🎯 Advisory Roles
I partner short-term or fractionally with growth-ready teams to adopt AI across your discovery processes in ways that work for you (not that other team).

🚀 Custom Workshops
Hands-on, team-specific trainings. You’ll leave with repeatable AI workflows, with a focus on SYSTEMS your team will actually use the next day.

📚 Courses
The most effective way for smaller teams to learn highly practical, pressure-tested workflows. These systems have already been proven by 300+ researchers, PMs, and designers—your team gets a playbook they can trust.

--

💌 If you want your team using AI to learn faster than the market, email me at caitlin@aicustomerresearch.com

--

📌 I've trained individuals and teams from these companies:
YouTube, Canva, Volvo, Bolt, Ramp, Spotify, Meta, Instacart, Asana, Lufthansa, Scribd, Vinted and more

Caitlin Sullivan

I'm building a Skill Library for Product Discovery workflows in Claude Code 📚

General skill libraries are popping up everywhere. 

But not enough Skills for PMs and researchers doing real discovery work.

In need of...
- JTBD interviews 
- feedback triage 
- transcript cleanup
- CSV analysis

I needed these for my own work. 

So I built a library.

My newsletter subscribers get it in the next edition (next week!).

Want it? 

**Tell me a workflow you wish was automated for you.** 
⇉ That's how I'll prioritize what to add next.

And you'll get the library, too. ✌️

AI for Customer Research ✦ AI insights systems, without the hype ✦ Training growth-ready teams to know your customers faster.

Synthesis isn't analysis. I set up 3 Claude Code slash commands to make sure I can trust its results...

The problem with AI synthesis is still: 
it looks done.

There’s typically clean output, well-phrased themes, and punchy recommendations.

It always *feels* like we made fast progress.
So we move on.

I used to do this, too.
In 2023, I'd review the output, make a few tweaks, call it done.

Models are so much better now.

And yet — mistakes still hide between the lines of tidy outputs.

〰️
AI analysis requires verification.

Not just "I reviewed every quote" — but asking:

"Is this really so representative of what customers said that we can easily make our biggest product decisions on it?"

I run verification at three points before anything leaves my desk.

↳ /verify-01 — Mid-analysis check

Before I go deeper:
→ Are the patterns emerging actually in the data?
→ What's being overweighted because it's memorable vs. frequent?

↳ /verify-02 — After themes are set

Before I write insights:
→ Pull quotes that contradict each theme
→ Flag themes supported by only 1-2 sources
→ Check: would I cut this if a stakeholder pushed back?

↳ /verify-03 — After insights are written

Before I let anyone ACT on these insights:
→ Does every insight trace to specific evidence?
→ Am I overstating? Would "some" be more accurate than "most"?
→ If someone asked "how confident are you?" — what's my honest answer?
→ Will I stand behind this when someone makes a big decision on it?

Yes, it means going slightly slower to go faster.

But shipping insights you'd actually bet on = spending less time worrying later. 

—

(This has the secondary benefit of often helping me write risky assumptions for testing next.)

—
🙋‍♀️ If you're using Claude Code for customer insights work:  
What does your process look like? I'd love to hear what's working for others.

Best-in-class AI customer discovery workflows come from **evals**.
I've run 10+ hours of evals weekly. Now Claude Code does it for me 🎉

Here's the thing: if you're using AI for customer insights work…

you need a way to know if your prompts and workflows are actually working.

Not just "this output looks good."
But..actually working.
Consistently.
Across models.
Across sessions.
Over time.

Otherwise, you're shipping products on insights that might actually CHANGE next time you run your latest prompts.

This is why I run evals weekly.

It's part of how I ended up running courses and advising clients on AI-assisted customer discovery - I have systems for this stuff.

But it's hugely time consuming.

Testing the same workflows over, and over, and over across all the models...

"Can’t Claude Code just do that for me?" (I had to ask.)

I've hesitated to hand this over to AI, because it felt like my human-brain-led evaluation of AI's outputs was the whole point.

The truth is, it's not.

When you have a good system — a proper eval framework, clear metrics, space for weighing outputs based on more subjective observations -

AI can do this just as well.

So I’ve crafted evals for Claude Code to run for me on my most common and time-consuming tests across models, tactics and full workflows.

Example:
- Run the same prompt across 3 models
- with the same set of transcripts.
- evaluate the results
- output all results and findings in a document I can review manually
- compare outputs per model and workflow over time

Plus, I get documentation I can actually reference later instead of losing my outputs and notes across tabs.

I won't pretend though that this works spontaneously.

I already had the eval and system to use, then I semi-automated it.

If you want Claude Code to run evals for you, these prep steps are non-negotiable:

→ A clear method for evaluating outputs that you've tested manually, many times. Nail your actual criteria.

→ Scorecards with metrics you can measure, plus space for the subjective observations that don't fit in a rubric.

→ Knowing what "good" looks like before you ask AI to replicate it.

→ Prompts you've already refined. The automation runs YOUR system at speed. It doesn't create the system (and it’s system recommendations won’t always improve on what you’re capable of).

And as with more complex studies or experiments, more complex AI workflows and evals take more oversight. This isn't always set-and-forget.

...but often it's really, really close.

📌 Is this something you want to learn how to do?

I'm thinking about running a hands-on workshop: 
How to build eval frameworks that test whether your AI workflows for customer insights are actually trustworthy — AND set up your first simple evals - API connections and all.

If that sounds useful, tell me in comments 👇

Running Claude Code agents in parallel isn't the magic. ⚡️ Good feedback loops and checkpoints are...

Here's what I keep seeing: 

More and more people using Claude Code.

And trusting Claude Code to do everything for them. 

Let it run, accept the output, move on.

👉 I don't do that.

I've watched it confidently summarize data but leave out essential parts of the picture:

- Missing contradictions between participants
- Flattening nuance into tidy bullet points that look right but aren't

So I build in verification steps. 

Agents that check other agents. 
Instructions that force it to flag uncertainty in reliable ways.

I'm running a free live demo on Wednesday:
↳ 3 agents analyzing 3 segments in parallel (with verification)
↳ Plus a final pass that tests whether output actually holds up

It's just 30 min (maybe 35). 
Link in comments if you want to join or watch it later. 👇

I created 20+ r𝗲𝗮𝗹𝗶𝘀𝘁𝗶𝗰 𝗺𝗼𝗰𝗸𝘂𝗽𝘀 of 8 products in 1 hour. Test before you build anything → Steal my prompts (FREE) 🚀

I love the idea of demand-testing all the possible products I've had in my head for years - before having to produce them.

But…

Most prompts I tested for mockups in Nano Banana last year *failed*.

𝗕𝗮𝘀𝗶𝗰 𝗽𝗿𝗼𝗺𝗽𝘁𝘀 𝗴𝗶𝘃𝗲 𝘆𝗼𝘂 𝘃𝗶𝘀𝘂𝗮𝗹𝘀 𝘁𝗵𝗮𝘁 𝗱𝗼𝗻’𝘁 𝗳𝗶𝘁 𝘄𝗵𝗮𝘁 𝘆𝗼𝘂 𝗵𝗮𝗱 𝗶𝗻 𝗺𝗶𝗻𝗱.
- Weird text that isn’t English
- Visuals that don’t fully line up
- Style that doesn’t follow your real app’s guidelines

So I spent my holidays figuring out what works.

And I did - generating highly realistic, “demand test”-ready visuals for versions of products and experiences I wish I had 😃

〰️

Here's what I made:

💛 A Whoop with a screen
💛 An Uber airport van option with toddler car seats + luggage space
💛 A limited edition Hoka Bondi inspired by Leif Podhajsky's art
💛 Leopard print TRX bands
[+ more]

What used to take designers (or me, ex-designer) days to prepare for product tests now takes 1 hour across 8 products. 🤯

You can do this, too.

But you need the *right kind of prompts*.

𝗨𝘀𝗲 𝗺𝘆 𝗽𝗿𝗼𝗺𝗽𝘁𝘀 𝗮𝘀 𝘁𝗲𝗺𝗽𝗹𝗮𝘁𝗲𝘀 that work, for FREE.

- Comment 'NANO' (or anything else) 👇
- Connect with me → I'll DM them to you

〰️

Skip these 3 steps and you'll never know what your AI analysis got wrong --

AI is getting pretty good at synthesis.

But if you haven't done your own thinking first, you'll either:

- Mistakenly trust whatever AI gives you
- or constantly second-guess the insights you're taking into your next meeting 

Before I run any AI analysis on research data, I consider 3 steps non-negotiable:

1️⃣ I review a sample of the data myself
↳If you haven't touched the raw data, you won't catch what AI misses.

2️⃣ I take my own notes - about what I observed, not just what I heard
↳What someone said is easy to summarize. 
↳What you noticed in how they said it? That's harder to capture - and it's your
check against AI flattening nuance.

3️⃣ I form my own theories before AI does
↳No hypothesis going in? Dangerous. You'll probably adopt AI's. 
↳Form your own theories first. They'll give you something to push back with.

〰️
This might be one of my simplest posts.

But the mini list is one of the things my AI Analysis course students call out as fundamental. 

It often changes how they think about their whole process.

Because *good* AI analysis of any kind starts way before you open up your LLM web UI or Claude Code.

What if you didn't need to obsess over whether AI outputs are trustworthy—because you already have a solid gut feeling for what should be in there?

That's what these 3 steps give you.

🙋‍♀️ Any other prep steps you're doing before you run AI on research?

I'm handing off work I didn't think AI could handle - because I taught it my own systems…

I've been an eager adopter of AI, and yet…
I was skeptical that it could do a lot of things as well as me. 

But Claude Code changed my mind.

Not because the tech changed - because I finally shifted from prompts to systems in a big way.

8 things I'm now trusting Claude Code to do:

DISCOVERY WORK:
1️⃣ Find insights across all existing discovery

Searching all existing research to answer "what do we already know about X?" - across years of projects.

There are limitations. Context windows matter. But I'm surprised by how often Claude pulls up truly insightful findings from massive feedback piles.

--
2️⃣ Spot the riskiest bets in my product plans

I used to teach this - how to map assumptions and find the riskiest ones. Ran a workshop on it at Hatch Conference a few years back.

But when you have good systems for things like this, Claude Code can simply... run them.

--
CONTENT WORK:
3️⃣ Fact-check my AI content

Models can now search the internet and their own documentation for the latest details.
Anything I write gets checked by Claude - like my engineer and technical writer sidekick.

--
4️⃣ Find new ideas in content I've already made

One night of bad sleep (toddler parent life) and I'm not sharp enough to find gems in my own work. Claude does it for me - possibly better than I ever did?

--
5️⃣ Answer new questions with stuff I've already written

Claude scours my materials so I don't reinvent answers every time. Same insights workflow, applied to my own content.

6️⃣ Keep my docs up to date
I hate documentation. My files have always been messy - design files, discovery files, writing, you name it. 

Now, end of every Claude session: "Review what we did and recommend which files to update."
Claude suggests the right edits 80% of the time on the first try.

--
CREATING SYSTEMS THAT WORK:
7️⃣ Building reusable systems, not one-off prompts 

The jump most people still haven't made.
Figure out what works (see #8), then make it repeatable. Every time I do something with Claude, I ask: Should this be a slash command, a multi-agent process, or a skill?

(Post coming soon on how to decide between them.)

--
8️⃣ Test what actually works (evals)

I used to manually test everything myself. It took most of my work week. Then I systematized -> still hours every week.

Now Claude runs evals that used to take 20+ hours in about 2, while I do other things. I review outputs, feedback anything not to my standards, and we update the eval files to fix those issues next time.
—

This shift isn't about better prompts.

It's about zooming out from them and teaching AI your full system - how you think through and weigh options, what your standards look like, how you define terminology, etc.

...then letting it run.


What's on your weekly to-do list that you don't think you'd hand off? 🙋‍♀️

I'm starting 2026 with 𝗙𝗥𝗘𝗘 𝗔𝗜 𝗖𝘂𝘀𝘁𝗼𝗺𝗲𝗿 𝗗𝗶𝘀𝗰𝗼𝘃𝗲𝗿𝘆 𝗹𝗲𝘀𝘀𝗼𝗻𝘀: Analysis across 6 models, realistic mockups for testing, Claude Code... 👩‍🔬

Session 1: 
3 𝗧𝗮𝗰𝘁𝗶𝗰𝘀 𝗳𝗼𝗿 𝗜𝗺𝗺𝗲𝗱𝗶𝗮𝘁𝗲𝗹𝘆 𝗕𝗲𝘁𝘁𝗲𝗿 𝗔𝗻𝗮𝗹𝘆𝘀𝗶𝘀, next week!

𝗪𝗵𝗮𝘁 𝗜'𝗹𝗹 𝗱𝗲𝗺𝗼:
Same prompt. Same data. 6 models. 
Wildly different results.

I'll show you exactly why this happens — and 3 fixes you can apply immediately so your AI analysis actually holds up.

This is just the first in a series of free Lightning Lessons I'm running to kick off the year:

→ Jan 14th: The 6-model experiment 
→ Jan 28th: 𝗣𝗿𝗼𝗱𝘂𝗰𝘁 𝗧𝗲𝘀𝘁𝗶𝗻𝗴 𝘄𝗶𝘁𝗵 𝗔𝗜: Test-ready mockups with Nano Banana 
→ Feb 4th: 𝗚𝗿𝗮𝗻𝗼𝗹𝗮 𝗕𝗲𝘆𝗼𝗻𝗱 𝗡𝗼𝘁𝗲𝘀: 4 uses for real insights 
→ Feb 25th: 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝗳𝗼𝗿 𝗨𝘀𝗲𝗿 𝗜𝗻𝘀𝗶𝗴𝗵𝘁𝘀 - what chained agents can do

...and more coming ⚙️

They're mostly ≈30 minutes each. 
Zero fluff. 
Focused on just the stuff you can use the same day (I know you're all busy!)

-
I've spent 1700+ hours testing AI for customer insights.

These lessons are the fastest wins from my latest tests that I want you to start your year with, too. ✌️

〰️
𝗪𝗵𝗶𝗰𝗵 𝗼𝗳 𝘁𝗵𝗲𝘀𝗲 𝗮𝗿𝗲 𝘆𝗼𝘂 𝗺𝗼𝘀𝘁 𝗰𝘂𝗿𝗶𝗼𝘂𝘀 𝗮𝗯𝗼𝘂𝘁? 

Drop your questions below, and I'll answer some in the sessions.

🔗 𝗟𝗶𝗻𝗸 𝗶𝗻 𝗰𝗼𝗺𝗺𝗲𝗻𝘁𝘀

Free Lightning Lessons for AI Customer Discovery in 2026

Your AI analysis is done. Deck is ready. You still don't trust it enough to hit send. 👉 I'm running a FREE session to fix that...

Trust issues are still a big deal with AI outputs.

Because you don't know if AI analysis 
- would hold up if someone challenged it. 
- would give the same answers running it again
- is solid enough to actually explain where it came from.

Despite LLM improvements last year, there are still *a lot* of challenges with getting AI analysis right.

I want to help more people fix that. 

Free session tomorrow: 
〰️ 3 Tactics for PMs - Immediately Better AI Analysis 〰️

I'm showing how to get outputs that are consistent and reliable enough to actually put in front of leadership.

↳ Live demos
↳ Models compared
↳ 40 minutes with AMA

I've taught people at Canva, Figma, YouTube and more to get AI analysis results they could stand behind.

🤝 Live attendees get a discount code for my full course.

Link in comments.

I joined Kate Syuma's course as a guest speaker last night. 🤝
Two AI topics we couldn't stop talking about:

1️⃣ Getting AI to actually synthesize data well 

Not just summarize it, but surface the insights you'd miss. 

Most people still run one prompt or 2 total AI steps before getting insights.

The real time savings + quality improvements happen when you build in your own criteria - what good looks like, plus how to verify it.

--

2️⃣ Synthetic users

I'm deep into this space right now. 
Can we trust them? When? How much?

How can we act ethically here?

--
I'd love to hear more voices on this topic.

✦ Synthetic users: real signal or false confidence? What have you experienced, what's on your mind? ✦

"How do I know when I have enough data to stop and decide?"
--> We still have to figure that out, even when AI is doing everything faster -- and we dug into it on Building With Buyers.

Plus we talk about the latest on where AI helps, where it doesn't, how to know the difference...and the list of other topics in Anna's post below. 

Thanks, Anna Furmanov. 🙏

Go to market using research, not assumptions | Host of Building With Buyers podcast

I wrapped up Season 5 of Building With Buyers, but then this convo was TOO GOOD + too relevant so then I un-wrapped up Season 5 😂 at least for a bit

Caitlin Sullivan joined me to talk about how AI is transforming customer research

Caitlin is an expert in AI for Customer Research. She trains growth-ready product & design teams to know their customers faster. Clients include YouTube, Spotify, Ramp, Meta, and Canva.

we chat about:

🌟The Shift in AI Usage for Customer Research
🌟Building Systems 
🌟Research Challenges with AI Insights
🌟what's actually really nice e.g. Completeness in Research with AI
🌟Common Mistakes in Customer Research
🌟Emotional Nuances in Research & Why they're important
🌟Understanding AI's Interpretation of Emotions
🌟The Role of the Human in AI Training
🌟Catching AI Errors 👀 
🌟Knowing When You Have Enough Data
🌟Favorite Tools

omg so good 😍 

p.s. Udi Ledergor remember when you asked me this Q almost one year ago I was still a bit unsure but now I am more sure 😇 

Ep. 270 on Building With Buyers

lemme know what you think

"Good. Fast. Cheap → Pick TWO" has been the rule for decades 
⏩ AI broke the triangle - or did it?

Right now, I'm on the side of, "the triangle is broken."

My partner and I are testing ideas for physical products.

Pre-AI: Find a manufacturer. Get the thing made. Iterate. Produce again. 

Spend €1000s and weeks before you have something worth testing.

What we did: 
Generated many realistic product photos in two hours. 
€20 Gemini subscription. 
Put them in front of real customers that same day.

Good. Fast. Cheap. 
All three.

I'm also running 3 customer discovery projects at once right now. 

Pre-AI that would've meant choosing one or hiring help. 

Now: All three, cheaply moving fast, and done properly.

So is the triangle dead? 

〰️
🙋‍♀️ I'd love to hear more perspectives - 
Where is the triangle broken? Where is AI still not able to deliver all 3?
〰️

AI for Customer Research ⚡️ Using AI to figure out customers 3x faster. Training Fortune 500 and growth-ready teams to use AI.

taplio

The Battle

Emily Anderson

Caitlin Sullivan

Followers

Posts

Engagements

Likes

Performance ScoreEmily Anderson's LinkedIn Performance

Performance ScoreCaitlin Sullivan's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Emily Anderson

Caitlin Sullivan

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Emily Anderson's LinkedIn Performance

Performance ScoreCaitlin Sullivan's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreEmily Anderson's LinkedIn Performance