writing
The 70/30 Engineering Audit: What to Automate and What Must Stay Human
Most teams adopt AI backwards. They buy a tool, hand out licenses, and tell everyone to find a use for it. Six months later they've got faster email drafts, a little less boilerplate in the codebase, and no real change to how the business runs.
The problem is where they start.
Start with the work. My whole approach runs on the 70/30 Method: in almost any business, roughly 70% of the work is repeatable, mechanical, and safe to hand to a machine. The other 30% needs context, relationships, accountability, and judgment.
I'm not trying to remove people. I build an Agentic OS that handles the repeatable 70% reliably, so the people I work with spend their hours on the 30% that decides whether the business actually moves.
The unit that gets automated is a workflow, never a whole job. So take one workflow, break it into steps, and run each step through five questions.
The five questions
Is it repeatable? Does the step follow a stable pattern? If the inputs and decisions recur predictably, it's part of the 70%. If every instance changes the problem in a real way, it needs a human.
Are the inputs ready? Can a system actually get the data it needs? Structured, permitted, and reachable through an API or a database means automation is on the table. If the context lives in a senior engineer's head, in scattered Slack threads, or in documents nobody's allowed to touch, the step stays human until you fix the data layer.
Is it observable? Can you tell quickly whether it worked? Safe automation needs clear success criteria and a fast feedback loop. If quality is subjective or the outcome only shows up weeks later, a human reviews it.
Is it reversible? What happens when it breaks, because it will. A cheap rollback like a draft PR description or an internal CRM update is fine to automate. Something expensive, public, or permanent, like a production migration, a signed contract, or an email to your whole list, needs a person holding the button.
What's the consequence? Low stakes inside clear boundaries: give it to an agent. Strategy, pricing, architecture, safety, and anyone's trust: that stays with a person.
The four control modes
Automation isn't an on/off switch. Based on the answers, put each step in one of four modes.
| Mode | When it fits | What the human does |
|---|---|---|
| Automate | Stable, low-risk, observable, reversible | Set the policy, review the exceptions |
| Assist | AI can draft or recommend, but context matters | Review, edit, decide |
| Approval-gated | The system acts only after you say go | Own the consequential call |
| Human-owned | Ambiguous, high-stakes, trust-dependent | Do the work, stay accountable |
The audit in practice
Run this across a few real workflows and the split gets obvious fast. Here's how the 70/30 line tends to fall.
Engineering delivery
The 70%: scaffolding, test drafts, CI triage, release notes, routine migrations. An agent can read the ticket, generate the boilerplate, write the first tests, and summarize what changed.
The 30%: architecture, security trade-offs, acceptance criteria, the weird exceptions, and the decision to merge. An agent can draft the PR. A senior engineer owns the choice to put it into production.
Sales operations
The 70%: enrichment, CRM updates, lead routing, reminders, follow-up drafts. A system never forgets to log a call or pull a company's latest funding news.
The 30%: discovery, reading the nuance in a lead, pricing, negotiation, and owning the relationship. Software can tell you who to call and what they do. A person builds the trust that closes the deal.
Customer support
The 70%: classification, retrieval, summaries, draft replies, status checks. Agents are good at reading a ticket, finding the right doc, and preparing a response.
The 30%: escalations, refunds, a customer who's clearly struggling, anything with reputational risk, and policy exceptions. When someone's angry or the situation falls outside the playbook, you want a human with empathy and authority.
Content and marketing
The 70%: transcription, organizing sources, formatting, generating variants, prepping distribution.
The 30%: positioning, the original insight, the claims, the voice, and final approval. An agent can turn a recorded interview into a draft. A person makes sure the argument holds and the tone is right.
Protect the human 30%
The common mistake is trying to squeeze the human out entirely. Automate the 30% and you'll ship a disaster eventually. It's only a question of when.
Keep human ownership over:
- Direction. Setting the goal and defining what a good outcome looks like.
- Edge cases. The rare, high-stakes situations no script saw coming.
- Relationships. The conversations that build trust and defuse conflict.
- Accountability. Someone whose name is on the result.
- Taste. Knowing what's actually good, not just what's technically correct.
- Saying no. Deciding what not to do is often the most valuable call in the whole business.
How to fail the audit
A few ways I've watched this go wrong:
- Automating a broken process. If a workflow is a mess or nobody really understands it, automation just makes it fail faster. Fix the process first.
- Confusing a draft with a finished thing. AI writes great first drafts and terrible final ones. Anything that reaches a customer gets an approval gate.
- Removing the owner. Every automated workflow needs a named human responsible for its output and its upkeep. No owner, no automation.
- Building plumbing you could've bought. Don't burn engineering time rebuilding what n8n or a similar tool already does well. Spend it on the part that's unique to your business: the context layer.
- Measuring hours saved. Time saved means nothing if quality drops or the team just refills the hours with busywork. Track cycle time, error rates, throughput, and real business outcomes.
Run a 30-day audit
Don't try to automate the whole company at once. Pick one workflow that's high-volume, highly repeatable, and low-risk.
Week 1: map the steps and run the five questions on each one.
Week 2: build the 70% automation in shadow mode, running alongside the human process without executing anything for real.
Week 3: compare the agent's output against the human's, side by side.
Week 4: if it holds up, roll it out to a small group with clear approval gates.
Prove the model on one workflow and the rest of the business starts asking for it on their own.
Not sure where to start? Take the 70/30 readiness assessment. It takes about 90 seconds and gives you a read on which parts of your work are ready to hand off.
If you already know which workflow is dragging your team down, let's talk. We'll map the process together and find the safest, highest-value place to begin. No pitch, just a useful conversation about what's possible.