# What a startup should not automate yet

> A practical framework for founders to separate repeatable work from strategic learning and decisions that still require human judgement.

- Author: Viktor Berthelius (BRTHLS)
- Published: 2026-09-01
- Category: automation aiops
- Tags: startups, automation
- Language: en
- Canonical: https://www.brthls.com/magazine/what-not-to-automate-startup-en
- Source: BRTHLS Magazine — https://www.brthls.com

---

## Problem

Once a startup sees AI produce, classify and respond quickly, every visible task starts to look automatable. Yet a frequent task is not necessarily a mature task. If the team is still discovering what customers value, automation can freeze a hypothesis that should remain open to learning.

## Thesis

AI's first advantage is not removing work; it is accelerating learning. Automate after observing a stable sequence, not to avoid the uncomfortable conversations required to understand it.

## Framework

Use three gates before building:

- **Repeatability:** are the input, decision rule and output stable enough?
- **Reversibility:** can an error be detected and corrected without harming trust, cash or compliance?
- **Learning value:** does doing the work manually still produce strategic information?

| Work | Initial decision | Evidence required |
| --- | --- | --- |
| Moving data between systems | Automate | Stable pattern and error control |
| Answering new objections | Assist | Reviewed library of cases |
| Choosing a segment or price | Keep human | Repeated market signals |

## Why it matters now

NIST and OECD guidance both emphasise context, measurement and appropriate oversight. For a startup, the operating translation is simple: speed is valuable only while the team can still learn and correct course.

## Anti-example

A founder automates outreach before understanding why ten customers agreed to meet. The system scales messages, but it also scales an undifferentiated proposition. Activity rises while signal quality falls.

## Protocol (3 steps)

1. **Observe five cases:** run the flow manually and record exceptions.
2. **Define the boundary:** state what AI may resolve and when it must stop.
3. **Compare learning:** measure time saved, errors and new signals—not volume alone.

## Related

- [The 14-day AI experiment for a startup](/magazine/14-day-ai-experiment-startup-en)
- [AI Fluency for educators: teach judgement, not just prompts](/magazine/ai-fluency-educators-judgement-en)

## Sources consulted

- [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework)
- [OECD AI Principles](https://oecd.ai/en/ai-principles)

## Next step

Choose one repetitive flow this week and write its stopping condition before automating it. If you cannot define that condition, you are still learning the process.

---

_Cite as: Berthelius, V. (2026). "What a startup should not automate yet". BRTHLS Magazine. https://www.brthls.com/magazine/what-not-to-automate-startup-en_