Thoughts
·
9 min read
Design system drift: when the world moves away from your system
Tools count products that leave the system. Nobody counts a system that has left its context. A watch list, six signals and one hour a quarter make both visible.

In short
A design system can be followed perfectly and still be out of date. This study separates the two kinds of drift, names the tools that count the first, and shows a small early warning for the second.
01 Two kinds
Conformance drift is products leaving the system. Context drift is the world leaving the system.
02 Early warning
A weekly list of dated changes turns a sudden change into lead time.
03 One review
Six numbers, limits set in advance, one hour a quarter.
A design system can be followed by every team and still be out of date. I call that context drift: the world moves on and the system does not. Tools count the other kind of drift well. This study shows how to see both, with an early warning that a small company can run by hand.
Two kinds of drift
Conformance drift is products leaving the system: detached components, hard-coded colours, forks. It is well described. UXPin defines design drift as the gap between the interface a system intends and the interface that ships. Eleken writes it as a formula: repeated exceptions, times time, times the number of teams. Magic Patterns separates drift from decay, the slow ageing of documentation and tokens.
Context drift is the reverse. Every product follows the system, and the system has stopped fitting the world around it: the platforms it runs on, what customers have learned to expect elsewhere, the law, and the company’s own strategy. The platforms are not the ones drifting. They are the context, and the system is what stays behind. I have not found context drift described as something a team can measure, so most of this study is about that half.
Exhibit 1
Two separate questions. Do products follow the system? Does the system still fit its context?
What it means. Green dashboards prove obedience, not fit. Exceptions and forks only see the bottom row. The top right needs a signal from outside the system, such as how long it took to answer a dated change.
Concept diagram, no measured data.
What tools count today
For conformance drift you no longer need to count by hand. Figma reports how many instances of each component exist and in how many files and teams, and its analytics API reports how often a component was detached. zeroheight scans code repositories for component usage, colour usage and package versions. Buoy looks for duplicate components in code and stops drift in the pull request.
For context drift I know of no tool. The four signals for it are counted by hand, from sources most companies already have.
Exhibit 2
Six signals, where each number comes from, and whether a tool reports it today.
What it means. The first two numbers can be automated this week. The other four need an owner, not software, which is why they are usually missing.
Sources and method
Sources: Figma Library Analytics, zeroheight, Buoy. I found no tool for the four context signals. That is the result of a short search, not proof that none exists.
The warning comes early
Context drift looks sudden only from inside. Take dark mode. Windows 10 had a dark app mode in August 2016. macOS followed in September 2018. Samsung announced a system-wide night mode in November 2018. Apple announced dark mode for the iPhone in June 2019, and Android 10 and iOS 13 shipped it within three weeks of each other in September 2019. A team that kept a list had 37 months between the first of these dates and the day iPhone users could turn their whole phone dark.
A law works the same way. The European Accessibility Act was published in June 2019 and has applied since 28 June 2025. That is six years of notice, with a date.
Exhibit 3
Each bar is the lead time between a dated signal and the day iOS 13 shipped system-wide dark mode, 19 September 2019.
What it means. A change that looks sudden usually leaves a trail. A watch list turns the trail into lead time, and the second platform is the moment to start.
Sources and method
Sources: release notes and announcements from Microsoft, Apple, Samsung and Google. Reference date is the iOS 13 release, 19 September 2019. The Apple announcement is dated to the month. The selection of signals is my own.
So the early warning is a list. Once a week someone, or an agent, reads the release notes of the platforms your products run on, the changes in the browsers, new accessibility rules and new laws, and writes down every change that has a date. It takes minutes and it can be automated. The list decides nothing. It makes sure the decision is yours.
Six signals
Count six signals each quarter. No study gives a universal number for when a deviation starts to matter, so the limits are a team decision. Set them before the first count, and write down what happens when one is crossed. Other fields already work this way. Google’s SRE workbook publishes an example error budget policy: a limit, an agreed consequence when it is crossed, and a named person who settles disputes. In quality control, the Western Electric rules count eight points in a row on the same side of the centre line as a signal. A run can matter as much as a single spike. The numbers below are my starting points, not measured values. Change them, but change them before you look at the data. If you have no feel for your numbers yet, count for two quarters first and set the limits from your own normal.
Exception rate (conformance). Deviating elements divided by all new elements in a release. From detach counts or lint rules for hard-coded values. In a small team, a tally in the design review. Act above one in five.
Local forks (conformance). Components in code that do not come from the library. From a code search or a scanner. Act when they rise two quarters in a row.
Decision age (context). Years since the last recorded decision behind each core rule. From the decision log. If there is no log, that is the first finding. Act above three.
Outside questions (context). Sales or support asking why two products differ. From one tag in the ticket system. Act on the same question twice.
Outside comparison (context). Your work beside three products your customers use, judged for thirty minutes by people outside the team.
Response time (context). Months since a dated change on the watch list that the system has not answered, counted against the trigger you set for it.
A crossed limit does not prove drift. It puts the item on the review agenda with an owner. A limit crossed two quarters running goes one level up, to whoever decides on budget.
The signals count deviations. They cannot tell decay from the next version of the system. So ask four questions of every deviation: is there an owner, a written reason, a checkable assumption and a date to adopt it or roll it back? Four times yes is evolution. Any no is drift, and for a safety-relevant rule, such as a fault colour, one no is enough.
Adopt, watch, or decide not to
Not every change on the list deserves an answer. Apple introduced 3D Touch in 2015 and removed it from the iPhone 11 line in 2019. A system rebuilt around it had to be rebuilt again. Three questions sort the list:
Reach. Does the change reach our users, as a default or only as an option, on every device or only on new ones?
Spread. Is it on more than one platform? The Kano model describes what follows: features that delight at first turn into expectations as time goes by. Once a pattern is everywhere, customers stop seeing it as a feature and start seeing its absence as a fault. That last step is my reading, not a measured result.
Cost of ignoring. Law, accessibility, and whether the brand stays legible.
Each item then gets one of three decisions: adopt, watch, or not for us. The reason goes next to it. Ignoring a change is a decision. Ignoring it without a record is drift.
Late is not the same as drifting
None of this says the fastest answer is the best one. In a study of 500 brands, Golder and Tellis found that almost half of the market pioneers failed. Lieberman and Montgomery list what later entrants gain: they can free-ride on the pioneer’s investment and wait until the uncertainty has cleared. That research is about entering markets, not about adopting a platform feature, so I read it as a caution, not as proof.
So choose a stance for each change, not for the whole company: lead, follow fast, follow late, or stay out. A law leaves only the question of when. A convention on every platform costs more to ignore each year. A new pattern is where leading can pay, and where it can strand you, as 3D Touch did.
What separates a late follower from a drifting system is a trigger written down in advance: we adopt this when the second platform ships it, or when the third customer asks. Response time is then measured against your own trigger, not against the fastest system in the chart.
Exhibit 4
Two separate questions again. Did you move early or late? Was it decided in advance?
What it means. Speed is not the measure. The record is. A late answer with a trigger is a strategy. The same date without one is drift.
Concept diagram, no measured data.
How long systems take to answer
Response time can be measured in public. System-wide dark mode shipped on iOS on 19 September 2019. Five design systems answered between 5 months before and 73 months after. Two were ready in time, one of them Google’s own Material. The other three needed 15, 39 and 73 months. Five systems are not a sample, and the milestone is my own selection, so read it as a hint: which of your rules are variables, and which did you pour in concrete? A late date can also be a choice. The chart shows when each system answered, not who was right.
Exhibit 5
The line is Apple: dark mode for every iPhone app, iOS 13, September 2019. Each bar shows how long a design system took to follow.
What it means. Everyone saw dark mode coming. Two systems were ready on the day; the slowest needed six years. When a change in your context has a date, response time measures drift directly. The two latest systems rebuilt their colour foundations first.
Sources and method
Sources: release notes and announcements from IBM Carbon, Google Design, GitHub, Atlassian and Salesforce. Milestone is the first public release, betas and opt-in previews included (own selection). The Salesforce theme left beta in October 2026. Reference date is the iOS 13 release. Material is Google’s own platform system, so its date is not a response to iOS. Five systems, not a sample. Data to November 2026.
What this looks like in a small company
Large platforms settle such questions with experiments. At Meta, by the account of one of its research scientists, several A/B tests are reviewed every day. That takes millions of users. A company with sixty people, two product lines and one designer has other instruments, and they are enough: the design review, the ticket system, the sales team and a list.
The page below is invented to show the format. It is not a client.
Exhibit 6
The review page of an invented company with sixty people and two product lines. Four numbers cross their limit.
Design system review, Q4 2026Illustrative
Exception rate
14%
Up from 9% last quarter. The limit is 20%.
WatchLocal forks
7
5, then 6, now 7. Up two quarters in a row.
DecideDecision age
4 of 11
core rules are older than three years without a review.
DecideOutside questions
3 times
“Why does the app look different from the display?”
DecideOutside comparison
2 of 3
comparison products were judged clearer than ours.
WatchResponse time
17 months
since the Accessibility Act applied. Focus states not yet reviewed.
DecideWhat it means. No experiment platform is needed. A tally, a tag, a list and an hour are enough to turn a feeling into four decisions.
Invented values to show the format. Not a client and not measured data.
This is the build, measure, learn loop from Eric Ries’s The Lean Startup of 2011, applied to the system itself. The weight has moved. One recent review of the loop argues that building is no longer the bottleneck. For a design system that means the scarce part is measuring and deciding.
The review
One owner, one hour a quarter, six numbers on one page, with the teams in the room. Each rule over its limit is kept, revised or retired. Each item on the watch list is adopted, watched or declined. The reason goes next to it. Why that record matters is the subject of Design systems remember answers and forget the questions.
When you do not need it, and how to prove it wrong
With one product, one team and a two-year horizon, skip the review: let people deviate and fold the good deviations back in. Drift gets expensive when several teams deviate in different directions and nobody decides which one becomes the rule.
To prove me wrong, run the review for two years and record the signals. If they do not rise before a team asks for a relaunch budget, the review is theatre. Stop.
What I expect next
I expect design tools to start showing whether a system has left its context, not only its components. The list of dated changes is the part that is easiest to automate. How such changes travel from one platform to the next is the subject of the next study.
Sources (27)
- Definitions
- UXPin on design driftEleken on design driftMagic Patterns on drift and decay
- Tools
- Figma library analyticsFigma analytics APIzeroheight code scanningBuoy drift detection
- Platforms
- Windows 10, August 2016macOS MojaveSamsung One UIiOS 13 announcement and releaseAndroid 10iOS 13 release
- Law
- European Accessibility Act
- Limits
- Google SRE workbook, error budget policyNIST, Western Electric rules
- Timing
- 3D Touch removedKano modelGolder and TellisLieberman and Montgomery
- Method
- A/B tests at MetaBuild, measure, learn reviewed
- Design systems
- Carbon dark themesMaterial guidanceGitHubAtlassianSalesforce

