The 70% question tells you which line is exposed. It does not tell you which two acts inside that line to rebuild first.
That second choice is where most edge programs die. Not because the stack failed. Because the firm selected a workflow that could survive a steering committee.
A lighthouse pilot is not a twin. A twin has to beat the incumbent on cycle time, quality, cost, and risk — on the same inputs — in full view of the people who own the old path. If the first two workflows cannot produce that evidence, the edge remains a lab.
What a workflow is, in this method
A workflow is a repeatable sequence that turns an input into a committed output.
It is not a department. It is not a role. It is not “customer service” or “finance.” Those are containers. The unit is the chain: request arrives → information is gathered → a decision is taken → an action is executed → a record is left.
If you cannot name the input, the output, the decision gate, and the system of record, you do not yet have a workflow. You have a slogan.
The Evolut stack is built for chains:
- Sense
- Interpret
- Decide
- Orchestrate
- Learn
The first two workflows must be long enough for that loop to matter and short enough that a parallel run can be stood up in weeks, not quarters.
The political attractors
Organizations do not pick workflows. They pick safety.
Four attractors show up every time:
The visible demo. A chatbot on the website. A copilot in the inbox. Something that photographs well and does not touch margin. It generates slides. It does not generate a deprecate-the-loser decision.
The orphan process. A workflow nobody important owns. Easy permission. No immune response. Also no consequence if it wins. The core will congratulate you and leave the real chains untouched.
The sacred cow. The process the CFO or the works council will defend as identity. You will spend the mandate arguing about existence instead of running a parallel.
The science project. A chain that requires a new data platform, a model trained from scratch, or a regulatory opinion before the first transaction. That is a research program. It is not the first twin.
If the proposed workflow sits in any of those four, it is the wrong first object — even if it is “strategic.”
Five tests for the first two
Score each candidate. Two yeses on tests 1–3 and a pass on 4–5 is enough. Perfection is how politics returns.
1. Throughput, not prestige
Does this chain run often enough that a four-week parallel produces a sample, not an anecdote?
A contract that closes twelve times a year cannot teach the stack. A claims file, an onboarding pack, a pricing exception, a supplier onboarding, a credit memo, a listing refresh — these can. Frequency is an instrument. Prestige is not.
Rule of thumb: if you cannot imagine fifty comparable instances in the first quarter of the parallel, pick another chain.
2. Coordination-heavy, judgment-light at the core
The 70% map already said this at the line level. Apply it to the act.
The rebuildable core of the workflow should be search, assembly, routing, checking, drafting, notifying. That is where agents collapse cost.
The human gate should be real and narrow: approve the exception, accept the residual risk, refuse the out-of-policy act. If every step is a judgment call, you are not looking at a first workflow. You are looking at the part of the firm that must stay inside the legal person.
A useful split: eighty percent of instances should be chainable; twenty percent should hit a named human. If the ratio is inverted, you selected a craft, not a workflow.
3. Same inputs, same customers, measurable output
The twin must run on the live diet of the old path. Not a cleaned dataset. Not a sandbox of friendly cases.
You need:
- a stable input (a file, a ticket, a form, a market event)
- a committed output (a decision, a filing, a booking, a payment instruction, a customer-facing artifact)
- four numbers that already exist or can be stood up in days: cycle time, quality/error rate, unit cost, risk events
If the current process has no system of record, that is information. Either instrument it first, or do not choose it first. A twin without a baseline is a story.
4. Contained blast radius
The first two workflows should be allowed to fail in public without taking the franchise with them.
Ask: if the twin is wrong for two weeks, what breaks? A delayed quote is acceptable. An unrecoverable payment, a privacy breach, or a licensed opinion is not — not at the start.
Contained does not mean unimportant. It means reversible. Rollback is part of the product. The first objects must be objects you can roll back.
5. A named owner who does not need a coalition
Someone with the right to change the path must want the evidence.
If the workflow’s owner is a committee, you will get a pilot. If the owner is a director who reports the numbers to the CEO, you can run a twin. The Evolut design assumes the twin reports to the chief executive. The first two workflows have to be close enough to that line that a weekly comparison does not require six signatures.
Two is a constraint, not a vibe
Why two, not one, not five.
One workflow can be dismissed as a special case. Five workflows are a transformation program, which is to say: the core has captured the edge.
Two is enough to show that the method generalizes. It is few enough that you can build the stack, the guardrails, and the human checkpoints once and reuse them. It is also few enough that you can still look at both parallels every week.
Pick two that share a data plane or a customer event if you can. A quote-to-cash pair. An intake-to-case pair. An exception-to-resolution pair. Shared context makes the Learn layer compound earlier. Do not force a pair that only looks related on an org chart.
A worked distinction
Wrong first workflow: “Use AI in HR.”
No input, no output, no gate, no baseline. A department.
Still wrong: “Build an employee assistant.”
A tool inside roles. Coordination architecture unchanged.
Better, still not first: “Redesign performance reviews.”
Low frequency. High identity. High politics. Judgment-heavy.
A first workflow: “Inbound partner onboarding from complete file to approved record.”
High throughput. Coordination-heavy. Clear input (a pack). Clear output (an approved counterparty in the system of record). A human gate on exceptions and sanctions hits. Cycle time and error rate already argued about in meetings. Blast radius contained if you keep write-back behind a review queue for the first weeks.
The second workflow should sit next to it — the first commercial act that depends on the approved record — so the stack begins to own a chain rather than a station.
What you are not choosing
You are not choosing the most broken process. Broken processes hide missing data and missing ownership. The twin will inherit the fog.
You are not choosing the process with the best executive sponsor if that process cannot pass the five tests. Sponsorship without throughput is theatre.
You are not choosing a workflow because a vendor has a demo for it. Vendor gravity is another political attractor.
You are choosing two sequences that can produce a comparison the CEO can read without translation.
A simple test
Write the two workflows on one page.
For each, fill in:
- input
- output
- decision gate and who sits above it
- current cycle time, quality, cost, risk — even if the numbers are ugly
- why it is coordination-heavy
- how you roll it back
- who reports the parallel to the CEO
If you cannot fill the page without a workshop, you do not have two workflows. You have a program.
Suppose the page exists and still looks like a chatbot, an orphan, a sacred cow, or a science project — start over. The edge only becomes the core when the first objects are allowed to win on evidence.
The line is the map. The two workflows are the first terrain. Politics will try to choose a park. Choose a road that already carries traffic.