PackPulse Insights · PackBack Network
Turning millions of cup scans into the three things an operator should do this week.
A redesign of PackBack's operator dashboard around a short weekly list of AI findings. Each one says what changed at a location, why we think so, and what to do about it.
- Role
- Product Lead, design owner
- Company
- PackBack Network
- Timeline
- Jan to Jun 2026
- Team
- 1 product designer, 4 engineers, 1 data engineer, our ops lead
- Platforms
- Web app, Email digest, PDF reports
1Four numbers, picked together with operators
2Every insight shows the data it is based on
3Each insight ends with a confidence level and one next step
23% → 68%
Operators active in a given week
+6.8 pts
Return rate at locations that acted on insights
41%
Insights acted on within 48 hours
3 h → 9 min
Time to prepare a quarterly compliance report
The problem
PackPulse showed everything and explained nothing. Operators logged in once a month to export a spreadsheet, and problems like a broken SmartBin were found by cleaners before anyone looked at the data.
What I did
I ran discovery with 14 operators, cut the dashboard down to four numbers, and designed Insights: AI findings that each come with evidence, a confidence level and one suggested action. I also designed the compliance report builder.
What changed
Weekly active operators went from 23% to 68%. Four in ten insights led to an action within two days, and locations that acted on them raised their return rate by 6.8 points on average.
01 · Context
Every cup is a data point, and we had about 9 million of them.
Every PackBack cup carries a QR code that works as its digital ID. It gets scanned when a SmartDispenser hands it out, again when someone drops it into a SmartBin, and again at the washing facility. Across 130+ locations that adds up to tens of thousands of events a day, and PackPulse is where our customers see them.
- Borrow↓
A SmartDispenser hands out a cup. Scan 1.
Guest
- Use↓
On site or to go. The deposit is held on their card.
Guest
- Return↓
A SmartBin checks the cup and refunds the deposit. Scan 2.
Guest
- Collect↓
Crates are logged in the Ops app on the van. Scan 3.
Field crew
- Wash↓
Industrial wash and a quality check. Scan 4.
Wash partner
- Restock↺ back to 01
Clean crates go back into dispensers.
Field crew, store staff
The people using PackPulse are not analysts. They are the operations manager of a franchise group with twenty restaurants, the facilities lead of a university campus, the retail manager at a train station. They care about three things: are the cups coming back, is the system saving or costing money, and can they show a regulator or head office that they offer reuse.
When I joined in November 2025, PackPulse was a set of charts the engineering team had built for the first pilots. It worked well for three customers. At 130 locations it had turned into a wall of graphs.
02 · Problem
Plenty of data, and almost nobody looking at it.
Usage data made the problem easy to see. In January 2026, only 23% of operator accounts logged in during a given week, and most sessions were a single visit to the export page at the end of the month.
Meanwhile, real problems went unnoticed. In one case a SmartBin at a station was offline for four days. Returns at that location fell by a third, guests left cups on the counter, and the station manager heard about it from a cleaner. The data that would have caught it on day one was in PackPulse the whole time.
PackPulse in January
- ✕14 charts on the home page, all the same size
- ✕Totals only, no comparison with last week or a target
- ✕Machine faults in a separate log nobody opened
- ✕Compliance numbers rebuilt by hand in Excel every quarter
What operators asked for
- ✓Tell me what changed since last week
- ✓Tell me why, and show me how you know
- ✓Tell me what I can do about it
- ✓Give me a report I can send without editing
The question we worked on
How might we make PackPulse tell an operator what changed and what to do, in the two minutes they give it on a Monday morning?
03 · Research
What operators actually do on a Monday.
I interviewed 14 operators across restaurant chains, campuses, stations and offices, sat next to four of them during their weekly check-in, and went through six months of support tickets with our ops lead. Three things came up again and again.
They want to be told what's wrong
Every operator described their ideal Monday as being told what is wrong. Charts helped only after they knew which question to ask.
“I don't have time to find the problem. Show me the problem.”
Return rate is the one number everyone watches
We showed more than 30 metrics. Thirteen of fourteen operators named return rate first. Cost per cup and compliance came second.
“If returns drop, everything else drops with them.”
A finding without evidence is a guess
When we showed early AI-written summaries, the first reaction was always some version of 'how do you know?'. People trusted what they could check.
“Who says? Show me the bin.”
04 · Process
Six months, and one shared definition of what an insight is.
The hardest part wasn't the interface. It was agreeing with our data engineer and ops lead what counts as an insight. We ended up with a three-part test: something changed compared with what is normal for that location, we can show why with data, and there is something the operator can do. If any part is missing, we don't show it.
In week five I prototyped three directions in code, on real data from eight locations, and tested each with five operators.
Chat with your data
Easy to demo. In tests, operators didn't know what to ask, and free-form answers were hard to trust or compare from week to week.
More charts, better filters
Our two power users liked it. Everyone else still had to find the problem on their own.
A ranked list of insights with evidence
A short list of findings, each with its data and one suggested action. Fits in two minutes and works for 24 locations as well as for one.
05 · Design decisions
What we built, and why it looks the way it does.
Four numbers on the home page, picked with operators
We cut the overview from 14 charts to four numbers: return rate, cups in circulation, single-use packaging avoided, and what unreturned cups cost. Each shows the change since the last period, and return rate is drawn against the customer's own target.
Everything else moved one click deeper. In six months, nobody asked for it to come back.
Every insight shows the evidence it's based on
An insight card answers four questions in a fixed order: what changed, why we think so, how sure we are, and what you could do. The 'why' is never just a sentence. It links to the machine log, the return events or the footfall numbers it came from.
We also list what we checked and ruled out. Operators told us this was the part that made them believe the rest.
Confidence decides how loud an insight is
Insights combine plain statistics with a language model. The statistics decide what gets flagged. The language model only writes the explanation, and it can only reference data we pass to it. That split made it much easier to test and to explain to customers.
High-confidence findings can send an email. Medium ones wait in the list. Low ones stay hidden unless someone opens 'all signals'. On average an account sees 3.2 insights a week, which is about what people said they'd read.
Email and top of the list
Returns fell 18% while SmartBin RC-02 was offline for 41 hours.
In the list, no email
Returns down 9%. No machine fault, but footfall data is missing.
Hidden by default
A small dip that fits normal weekly variation.
Operator feedback
Removes the insight and sends it to the data team's weekly review.
Locations sorted by what needs attention
Operators with many sites used to scan an alphabetical list. Now locations are ordered by open insights and return rate trend, and a simple map shows where the problems cluster. A franchise manager with 20 restaurants sees the two that need them first.
Compliance reports in one click
Since 2024, Dutch rules require reusable options for on-site consumption, and more customers have to show numbers to municipalities and head office. Operators were rebuilding those numbers in Excel every quarter, each in their own way.
The report builder produces a PDF with a fixed, documented method, so reports from two locations can be compared and an auditor can see how every number was calculated.
06 · Edge cases
Designing for the days when the data is wrong.
Insights are only as good as the scans behind them. A machine without internet still accepts cups, but its scans arrive hours later. If we hadn't handled that, PackPulse would have reported fake drops every day.
| Situation | What the operator sees | Why |
|---|---|---|
| Machine offline, scans arrive late | Numbers marked as incomplete, drawn dotted until the machine reconnects | Prevents a fake drop in returns |
| New location, under three weeks of data | No insights yet, a 'learning this location' note | Without a baseline there is no fair comparison |
| Public holiday or event | Insight held back if footfall explains the change | Nobody needs an alert about Easter Monday |
| Delay at the washing partner | Stock warning on dispensers, cause shown as supply | Points the operator to the right team |
| Operator disagrees | 'Not right' removes it and asks one optional question | Goes into the weekly review with the data team |
07 · Outcome
From a monthly export to a weekly habit.
We ran a six-week beta with eight operators and released Insights to every customer in June 2026. Weekly active operators climbed from 23% to 68%, and most sessions now start on the insight list instead of the export page.
“It's the first thing I open on Monday now. Most weeks there's nothing urgent, and that's the point.”
“I sent our Q2 numbers to the municipality without opening Excel once.”
08 · Reflection
What I'd do differently.
I would bring the field crew in earlier. Half of all insights end as a task for them, and we only connected Insights to the Ops app two months after launch. Until then, operators copied insights into WhatsApp messages.
I'm also more careful now about how much the language model writes. Our first explanations were three sentences long. The version that tested best was one sentence and a chart.
What's next
- ·Insights inside the Ops app, so a technician sees the context before arriving at a machine
- ·Cup stock forecasts per dispenser, using the same evidence format
- ·An Arabic version of PackPulse for our first customers in the Gulf
Tools on this project
Figma, FigJam, React and TypeScript for prototypes, Claude API, dbt, Metabase, Dovetail, Amplitude
Names of the people quoted have been changed to protect their privacy.
SmartBin and dispenser screens →