You're automating too much

Do less. No wayyy less than that

Author

Zubin Teherani
June 04, 2025

You’ve seen it before. You or team member proudly shows off a 17-step Zap that auto-magics literally everything—except revenue. A week later that same Zap is crying in JSON because finance added a column.

If this hits home, congratulations: you may be the proud owner of an over-engineered ops stack.

If Pixar wrote your error logs, they’d title them Finding Failure.

Below is a smarter way to decide what to automate, when to keep it manual, and how to avoid the dreaded “I’m going to jackhammer a nail” syndrome (this isn’t an expression, but stay with me).

Why It Matters

  • AI isn’t cheap. Data-center spend will hit a mind-numbing $6.7 trillion by 2030, mostly to feed AI workloads. RIP planet.

  • Budgets are already screaming. This year, 33% of orgs burn $12 million+ on public cloud alone—up 4 points from 2024.

  • The tooling arms race won’t slow down. OpenAI’s Sora went from “futuristic demo” to “free inside Bing” in five months.

Spending real money on half-baked automations that won’t scale could be just as bad of an investment as my pricey espresso machine.

Rorschach but make it dairy

What Founders Usually Get Wrong

1. Automating Vanity, Not Value

It’s tempting to script the cool thing instead of the cash thing. A/B test the ROI: if automation doesn’t directly drive or unblock revenue, file under “later.”

2. Scaling Before the Party Starts

Your Google Sheet still has 42 rows. Do you really need a serverless LLM pipeline? Let demand break the process first; then upscale.

3. Cementing a Flimsy Workflow

Once you codify logic into AI prompts or code, change gets expensive. Remember how that 8-line regex became a 400-line horror flick?

4. Ignoring Error-Prone Hand-offs

Some tasks (looking at you, month-end close) are magnets for human error. These deserve automation love early—just maybe not an LLM when SUM() works.

A ~Framework~ Because Why Not

Below is a stepwise framework—Manual → Template → No-Code → Code → AI. Iterate forward only when the previous tier creaks.

Tier

When to Use

Tool Examples

Exit Criteria

Manual

New, fuzzy process

Google Docs, quick Slack DM

You repeat it weekly

Template

Repetitive but stable

Sheets formulas, canned email

Errors or boredom appear

No-Code

Moderate volume, clear logic

Zapier, Make, Airtable

Performance or flexibility limits

Code

High volume, custom logic

REST micro-service

Still hitting edge cases

AI/LLM

Complex decisioning, unstructured input

OpenAI, open-source models

ROI > $ (and headache)

Prioritization Checklist

  1. Revenue First: Does it move cash in or out? Rank high.

  2. Error Magnet: Are humans routinely face-palming here? Rank high.

  3. Iteration Rate: Rapidly changing processes stay manual or template-based.

  4. Data Shape: Structured? Use code. Unstructured? Maybe a LLM.

  5. Cost Curve: Monitor API pricing and GPU scarcity—compute costs swing like crypto.

If a task takes you less time to do manually than to think about automating, step away from the AI prompt.

TL;DR

  • Automate revenue-critical, error-prone steps first.

  • Let your workflow cry uncle before upgrading its tool belt.

  • Move progressively from Manual → Template → No-Code → Code → AI.

  • Keep an eye on compute costs

  • Complexity ≠ sophistication. A well-timed SUM() can beat a shaky GPT-4 prompt.

Internal Reading

If you enjoyed this, subscribe. If you didn’t, please just subscribe and unsubscribe later so it doesn’t hurt my feelings.

You can also check out: