Iterate-Then-Automate Scaling Ladder
Let ops manually iterate a process until it breaks at scale, then hand exactly that broken step to technology
- Difficulty
- Moderate
- Time to result
- ~months to results
- Steps
- 5
- Confidence
- 90%
A rule for deciding when a human-run operational process should be replaced by software. You deliberately run a process manually while it is still being figured out, watch for the specific point at which manual improvements stop being viable, and only then invest engineering to automate that exact bottleneck. This frees the operations people to move up to the next unsolved problem.
Origin
Brian Tolkin's articulation of Uber's driver-onboarding evolution; he ties it to the widely-used maxim (popularized by Paul Graham / Airbnb) 'do things that don't scale, then scale the things you're doing.'
Core principles
- 01Operations teams iterate and gather qualitative insight faster than engineering in the early days
- 02Technology is uniquely good at scaling, not at figuring things out
- 03Automate the step only once manual system improvements are 'no longer viable'
- 04Automation should free human capacity for the next hill, not just cut cost
How to run it
- 1
Run the process manually while it's still being figured out
In the iteration phase, do the work by hand — e.g. onboard each Uber Black driver in a 90-minute to two-hour in-person session with deep expectation-setting.
Pro tip Manual delivery is where you learn the nuances the eventual system must encode.
- 2
Scale the manual process incrementally and watch for the break point
Batch it up: 3-6 drivers in a small class, then 20-30, then swap live delivery for a video. Each step reveals a new constraint (e.g. validating driver's-license credentials is trivial at one person, strained at 20, broken at a thousand a week).
Watch out Each scale step can surface a NEW bottleneck (credential validation) even as it solves the old one — don't assume one fix holds.
- 3
Diagnose the transition from iteration stage to scale stage
Name the moment when operational system improvements stop being viable. That is the signal the problem has moved from 'iterate' to 'scale'.
Pro tip Ask: are we still learning here, or just repeating? Repetition at volume is the automation trigger.
- 4
Apply technology to the specific broken step
Replace the manual step with software built for scale — e.g. integrate OCR/auto-recognition of driver's licenses feeding a system that validates automatically.
- 5
Redeploy freed human capacity to the next problem
The people who ran onboarding now move up to analytics, the next process needing optimization, or new product bets. The virtuous cycle continues.
Pro tip Reassure ops teams their job doesn't disappear — automating one hill frees them to climb the next.
In the wild
Uber Black drivers were each onboarded in 90-120 minute in-person sessions, then small classes of 3-6, then 20-30, then an onboarding video. Credential validation held fine at small numbers but broke at ~1,000 drivers/week, so Uber integrated OCR/auto-recognition of driver's licenses.
→ Scaled the system and freed dozens-to-hundreds of onboarding staff worldwide to work on higher-leverage problems.
Common mistakes
Automating before the manual process has taught you the nuances
Jumping to software during the iteration stage means encoding rules you don't yet understand; the manual phase is where the real-world edge cases surface.
Treating operations as pure inefficiency to be eliminated
Tolkin partly disagrees with the 'operations is a sign of inefficiency' take — the point isn't to erase ops but to free that capacity for the next challenge; there's always another hill to climb.
Is it for you?
Best for
Product and ops leaders at operationally-heavy marketplaces or logistics businesses deciding what to build vs. keep manual in the early scaling phase
Not ideal for
Pure-software products with no real-world operational component, or processes still deep in the discovery phase where manual iteration is still teaching you things
From the transcript
“we've gone from the iteration stage to the scale stage and and technology is uniquely good at scaling”
“how do we integrate with some type of OCR technology or Auto recognition of driver's licenses that feeds to A system that knows what a…”
“one you've scaled your system and two you've just created a ton of time for what at the time was probably dozens if not hundreds…”
“do things that don't scale and then scale the things that you're doing that's the phrase I always come back to”
From the episode
Lessons from scaling Uber and Opendoor
Brian Tolkin (Head of Product at Opendoor, ex-Uber)