05What to do instead
Operator’s playbook
Strategy between the extremes
The failure patterns are mirror images, and so are their corrections. Automation is one dimension, not the whole story — so this page holds tools for both ends: five tests that separate good automation from bad, sequences for cooling an overheated program and starting a frozen one, and a substitution check for spotting when AI is eating your market, not just your workflow.
Rev 05 · Q2 ’26
Good automation, bad automation
The same technology compounds in one workflow and corrodes in the next. Five tests separate them — none require understanding the models.
- 01
Cost of a wrong answer
Low, visible, cheap to correct — drafts, triage, internal search.
High, silent, borne by the customer — claims decisions, medical advice, pricing.
- 02
Reversibility
Can be switched off tomorrow without rebuilding the organization.
Requires letting people go first and hiring them back later.
- 03
Escape hatch
A human is one click away, and the handoff keeps context.
The bot is the only door, and it locks from the inside.
- 04
Feedback loop
Every correction improves the system; quality is measured weekly.
Nobody measures error rates after the launch announcement.
- 05
Volume vs. judgment
High volume, low ambiguity: the machine does the repetition.
Low volume, high stakes: automated because it was annoying, not because it was suited.
Running hot: how to cool down without losing the plot
The correction for overheating is not abstinence — it is sequencing. Keep the conviction; make the evidence catch up.
- 01
Cap the pilot portfolio
Three concurrent pilots with named owners beat forty with none. The cap is not a limit on ambition — it is a forcing function for ownership.
- 02
Write kill criteria before kickoff
Decide what result ends each pilot while everyone is still calm. If nothing in the portfolio has ever been killed by criteria, the portfolio has no governance — only momentum.
- 03
Sequence headcount after evidence
Automate the workflow, measure it against the human baseline for two quarters, then re-plan roles around the residual work. Cuts made before the baseline exists can never be evaluated — only regretted or defended.
- 04
Separate the story from the strategy
The external narrative may run ahead of the roadmap; the operating plan must not. Keep two documents and be honest about which one the board is reading.
Frozen: how to start without betting the company
The correction for the freeze is not a transformation program — it is a small, owned, measured start.
- 01
Buy a map before a machine
Inventory your workflows by error cost and volume before touching any tool. Most firms discover that two or three boring processes carry most of the opportunity — and none of them are the ones in vendor demos.
- 02
Start where mistakes are cheap
Internal drafts, summarization, search, triage. The point of the first workflow is not savings — it is building the muscle of measuring machine output against a human baseline.
- 03
Give one person the mandate
A named owner, a small budget, a quarterly review. Capability without an owner is a subscription, not a strategy.
- 04
Build evaluation muscle first
Measure before and after, on the workflow’s own terms. The firms that do this deploy new use cases in a quarter; the firms that don’t begin every cycle with a vendor beauty contest.
Don’t get disrupted: seeing it before it’s priced in
The frozen failure has a second shape the automation tests don’t catch: not refusing to adopt AI, but missing that AI is eating your market. These questions point at your own product, not a workflow. Two or more yeses is exposure, not a maybe.
The substitution check
- 01
Is the core thing you charge for now available free from a general-purpose AI?
- 02
Are your customers one prompt away from doing it themselves?
- 03
Did your value rest on a scarcity — of answers, access, or expertise — that the model removes?
- 04
Is your own data training the system most likely to replace you?
- 01
Run the substitution check quarterly
Point the questions inward — not “should we automate this,” but “can a free model now do what we sell?” The answer moves faster than an annual strategy cycle, which is exactly how the freeze catches incumbents.
- 02
Cannibalize before you’re cannibalized
If a free AI version of your product is coming, build it yourself while you still own the customer relationship. Chegg’s response wasn’t wrong — it was late, and late is repriced before it ships.
- 03
Move up the stack
Retreat from the layer the model commoditizes toward what it can’t give: trust, accountability, integration, the last mile. The answer is free now; being responsible for it is not.
The overview is the product.
Balane GmbH runs this monitor and works with leadership teams on exactly these questions: which workflows to automate, which to protect, and how to govern the difference. One conversation, no funnel.
Request a strategy briefing