Select Automation as a Reversible Option, Not an All-In Bet
By Equipo Quantum Developers
Summarize:
Operating thesis
Selecting “the project with the highest ROI” often means selecting the spreadsheet with the most ambitious benefit. That comparison rewards assumptions rather than evidence. A more defensible decision funds the next block of learning: the winning project can test its critical assumption with bounded loss and an executable exit.
The UK Treasury Green Book 2026 calls for comparing options, retaining business as usual, analyzing uncertainty, and considering flexibility when new decisions can follow new information. For automation, flexibility turns a project into an option: invest a little to gain the right, not the obligation, to expand later.
From initiative to a sequence of options
An irreversible bet commits integration, licenses, process change, and autonomy before outcomes are observed. A reversible sequence separates:
- discover whether object, data, and outcome exist;
- test decisions on historical cases;
- observe live inputs without acting;
- assist a person in a narrow population;
- enable one reversible action;
- expand only with outcomes and control.
Each stage has cost, expected evidence, maximum exposure, and a later decision. The approved budget covers the stage, not an implicit promise to reach the end. Canceling after invalidating an assumption is a correct result, not a team failure.
Start with a real alternative
The worksheet compares at least:
- business as usual;
- policy, form, or data improvement without AI;
- deterministic automation;
- an assistive agent;
- an agent with bounded autonomy.
The agent does not compete with zero. It may lose to a cheaper rule or process change. Show which benefit depends on another investment. If data remediation creates value on its own, separate it from return attributed to the agent.
Use a pre-mortem to test the narrative
Before scoring, the committee adopts a hypothesis: “The project reached review and was stopped without producing the intended outcome.” Each function proposes causes:
- operations: exceptions consumed capacity;
- business: the outcome did not change even though the task was faster;
- data: the source was stale or did not represent the population;
- control: the action needed approval that the design treated as friction;
- technology: the integration could not reconcile uncertain states;
- finance: released capacity never became savings or production;
- users: the alternate process remained more trustworthy.
For every cause, record an early signal and a cheap test. The pre-mortem does not assign invented probabilities. It forces the team to seek evidence capable of refuting its preferred story.
Artifact: the reversible selection worksheet
| Field | Content |
|---|---|
| business decision | outcome, population, and owner |
| business as usual | cost, risk, and performance without change |
| alternatives | process, rule, automation, and agent |
| critical assumption | claim that destroys the case if false |
| next test | smallest experiment observing that assumption |
| maximum exposure | money, cases, data, and consequence at risk |
| reversal | command, owner, time, and destination for cases |
| success evidence | outcome, coverage, comparison, and horizon |
| kill criterion | observable signal that stops or narrows scope |
| option cost | full stage cost, including review |
| next right | what may be authorized if the test passes |
| dependency | critical source, integration, person, or policy |
Sign the sheet before starting. If criteria change after results appear, record a new version and approval.
Kill criteria that can be executed
“If it does not work” is not a criterion. It needs a signal, boundary, and owner. Examples:
- close when the outcome cannot be observed in the window;
- return to shadow after a material action lacks complete evidence;
- narrow population when harm clusters in one segment;
- stop when a critical source fails freshness or provenance;
- do not expand when cost per outcome leaves the approved range;
- retire when rollback cannot execute or leaves inconsistent objects.
NIST includes proceed-or-stop decisions, residual-risk management, and mechanisms for deactivating systems that do not meet purpose in the AI RMF Core. Kill criteria translate that capability into a specific investment.
Costs and ranges, not an aspirational number
The GAO Cost Estimating and Assessment Guide recommends documenting the technical baseline, assumptions, data, sensitivity, risk, and updates using actual cost. The worksheet includes build, integration, observability, human review, operations, incidents, security, and retirement.
For benefits, use low, central, and high ranges with coherent conditions. Identify the variable that changes the decision: adoption, exception rate, capacity realization, or review cost. The next option should test it. A central estimate does not win when it depends on an assumption nobody can observe.
Illustrative example: claim classification
Two proposals compete. One automates classification for every claim and promises substantial capacity. The other observes one category, proposes routing, and preserves human decision. The pre-mortem reveals that taxonomy changes and reopening appears later.
The second option buys evidence about category stability, disagreement, and reopening with bounded exposure. If it works, it enables a reversible action for that segment. If it fails, the enterprise still has a better reason catalog. This example is illustrative: it offers no universal percentage, schedule, or promised ROI.
The committee rule
Do not use a weighted sum as the final decision. Apply vetoes first: outcome unobservable, irreversible action without control, source without permission, owner absent, or reversal impossible. Among eligible options, compare:
information value + expected operating value − stage cost − residual exposure.
Every term need not receive a falsely precise monetary value. The committee can use ranges and explanation. What matters is giving learning explicit value and preventing the high scenario from hiding lock-in risk.
In Quantum Automation Center, catalog, states, timelines, artifacts, logs, analytics, permissions, and approvals can link the worksheet to executions and gates. Selection remains a portfolio decision; the platform preserves its evidence.
The strongest counterargument
Breaking investment into small options can favor incremental change. Some integrations, data platforms, or process redesigns need critical mass before creating value. Demanding a return from every stage may underfund the common foundation.
That criticism is correct. An option should not pretend to be independent. In shared infrastructure, a stage may produce enabling capacity—a data contract, identity layer, or observability—with clearly identified future adoption. It still needs a usefulness test and an exit from a wrong design.
When not to use this approach
Do not use the logic as an excuse to fragment mandatory infrastructure, regulatory compliance, or change with an indivisible foundation. Do not describe a political or organizational commitment as reversible when undoing it causes material harm.
Use the approach where uncertainty is material and scope can be sequenced. The best selection is not the story with the largest return. It is the option that buys decisive evidence without turning an early hypothesis into permanent debt.
Sources
Article topics
