The Paradox of Automation
It often feels like one step forward, two steps back when you build automations. But finding the bottleneck means you found the real work. Read: "Everything Starts Out Looking Like a Toy" #270
Hi, I’m Greg 👋! I write weekly product essays, including system “handshakes”, the expectations for workflow, and the jobs to be done for data. What is Data Operations? was the first post in the series.
This week’s toy: an ultrasonic knife for home cooks. This seems like science fiction - it uses ultrasonic waves to vibrate the knife edge so quickly that it creates the equivalent of a much sharper blade. Edition 270 of this newsletter is here - it’s September 29, 2025.
Thanks for reading! Let me know if there’s a topic you’d like me to cover.
The Big Idea
A short long-form essay about data things
⚙️ The Paradox of Automation
Automation often seems easy and hides an underlying truth. Every time you automate a step, you don’t erase all the work.
Spoiler: You advance the workflow to reach the next bottleneck.
In practice, you improve one thing to find a new issue:
Automating client summaries leads to waiting for client approval
Automating ticket triage leads to understanding what’s holding up resolution time Automating intake lets you focus on fulfillment
But finding the next bottleneck is not automation failure, even when it means you’ve found a new problem. It’s a success! Process bottlenecks are signals to let you know where things have broken down or are blocked.
Often you think the automation can solve this problem simply by connecting the dots. But sprinkling automation on your problem doesn’t guarantee a solution. The side effect here amplifies whatever is happening in the current automation. If you’re stacking up things that are difficult to do, automation helps you do that faster. (Not always good!)
So how do you get better at automating your manual processes?
First principles thinking suggests you need to solve the underlying process and then automate. If you focus on your bottlenecks and fix the process there you’ll have a good sequence of automate, block, solve, automate.
The real question isn’t “Where can we add automation?”, it’s “What has to be true before a machine can safely do this work?”
You can’t assume the existence of an API
The best case for automation involves API calls, clean data, and predictable responses. Your API might beven tell you the actions available, the signature of the responses, and even fully document these calls using online documentation.
That’s not what you get in professional services or in a manual workflow.
The API doesn’t exist. Instead you have an operating procedure and you don’t have a good assurance that it will be done the same way every time, because it depends on people.
Here’s a quick summary. Take something as ordinary as filing for a business license. There is no /applyForLicense endpoint where you can specify a state where you want to create a specific business type in a specific state.
Instead, you get:
Many state web sites, each with their own login quirks.
Multi-factor authentication tied to the business owner’s phone or email.
Dynamic forms that show or hide fields depending on your entity type.
A receipt that only appears as an HTML page or arrives later by email.
From the outside, the request sounds simple: “Get me licensed in Illinois and Colorado.” From the inside, it’s two completely different obstacle courses with no standard interface and no official way to automate them.
A tale of different states: Illinois vs. Colorado
Here’s how the process works in two different states:
Illinois
You search for the business, then re-enter agent and officer info already on file.
You submit separate forms for incorporations, amendments, and annual reports with no shared record.
You pay fees through dated, inconsistent checkout pages.
You download PDFs as proof instead of receiving structured data.
Colorado
You search for the business, but filings still require retyping some stored details.
You complete every filing as a form submission with no programmatic endpoint.
You pay per transaction without a reusable payment profile.
You get a stamped PDF back instead of a structured confirmation.
To a client, both are “get me a license.” To your back office, they are totally different processes. Humans can brute-force that with judgment and experience. Machines can’t improvise like that — not without structure.
The ladder to machine-readiness
To move work out of human hands, you need to climb a ladder from the things that can be done by a human to the next bottleneck that you find.
This sequence diagram gives you an idea how this works:
Step 1: Document
Write the checklist for each state. These should include the Trigger, required inputs, exact steps, and definition of “done.” You might need to set up a series of states where you can stop or need to continue to avoid losing work.
Step 2: Make the process visible
Turn every application into a case record or a series of steps and outcomes. The simplest of this looks like one row per state and company, with columns where you can note whether the next step in the process is completed. If you have related data, you can put a link to that related data in the table.
Step 3: Assist first
The best way to think about “process-assisted automation” is to imagine information arriving just in time for the operator. A version of this looks like an iframe where you place the actual page or resource being used and keep your instructions in context.
If the application doesn’t allow iframing, you need to create another control flow (a iist of steps and pages that drive another browser window).
Step 4: Automate next
Getting from step 3 to step 4 is a big automation gate. Assisting the operator doesn’t require controlling the browser or creating an API experience. On the other hand, true automation needs to either remote-pilot a website with the user’s authentication or to send information to an API and receive a response.
when you know what’s supposed to happen, you have authentication and control, it’s possible to start filling information and let the operator confirm important decisions.
Step 5: Delegate last
If you achieve automation and an operator agrees that it works most of the time, you’re probably ready to delegate steps to a machine process.
But you can’t delegate steps when the automating process isn’t able to complete the work.
What has to be true for pure automation
Each rung of the ladder moves you closer to these conditions for machine-readiness:
Access: ability to proxy on behalf of the user (via API or browser)
Bounded: clear start and finish (submit license, receive receipt)
Structured data: inputs collected upfront; no chasing clients mid-process
Explicit rules: logic written down (“If nonprofit, attach Schedule C”)
Tolerable risk: mistakes are low-cost or caught by guardrails
Fallback path: machines know when to stop and hand back to people
Until these are true, the step stays with humans. Once they’re true, the machine can safely take it on.
The paradox in practice
Automating licensing doesn’t eliminate work.
It makes the next constraint visible:
If machines fill the forms, the bottleneck shifts to client approvals.
If machines manage approvals, the bottleneck shifts to payments.
If machines handle payments, the bottleneck shifts to regulatory exceptions.
Each bottleneck you expose is progress — It’s a new target for redesign.
Getting beyond the bottleneck
Automation won’t save messy workflows. What it demands and rewards is clarity. If you automate a bad process, you’ll get more bad process.
Business licensing is just one example. Insurance filings, lease registrations, government certifications — all follow the same pattern. They start as opaque, human-only SOPs. Step by step, you make them visible, assistive, then automatable.
The paradox of automation is that you never really finish. Each solved step surfaces the next bottleneck. But that’s the point. That’s how you scale professional services beyond headcount.
What’s the takeaway? Automation reveals bottlenecks, not eliminates them. Fix the process first, then automate one step at a time.
Links for Reading and Sharing
These are links that caught my 👀
1/ Skills through games - I love this article on board games mapping the specific user goals taught by each game. It goes far beyond “learn about real estate by playing Monopoly or strategy by playing Risk.”
2./ Star Trek future - An argument that we are closer than ever to a new economic system of abundance. Is it techno-fiction or a forecast to the future? Check back in a few years.
3/ Choose your own adventure - Before ChatGPT, there were books that allowed you to experience an interactive story, one page at a time.
What to do next
Hit reply if you’ve got links to share, data stories, or want to say hello.
The next big thing always starts out being dismissed as a “toy.” - Chris Dixon