If you own a technology or strategy call this quarter — rebuild vs. refactor, build the AI capability in-house vs. buy it, which platform carries the next three years — the expensive failure mode is rarely a shortage of opinions. It's that the opinions never get structured: six months later nobody can reconstruct why the call was made, what it assumed, or what would have changed it. That is how costly technology mistakes get defended long past the point they should have been reversed.
Three frames fix most of that. They are the core of Comuvia's Decision Map Workshop, and they are the same frames this Company AI System applies to Comuvia's own technology and strategy bets. Each is small enough to run without a facilitator — this article gives you all three, deliberately. Together they produce a decision record and expose the evidence still missing. They are a working method, not a measured guarantee of decision quality.
The three decision frames in sequence: a live decision passes through the reversibility test, the option matrix, and the horizon band, producing a defensible call — or exits early if it is cheap and reversible
The three frames in sequence. The reversibility test sizes the decision — and lets most decisions exit early. Only the ones that survive it earn the matrix and the horizon check.
1. The reversibility test
Two questions place every decision: how big are the stakes, and what does it cost to reverse?
Reversibility test as a two-by-two: small stakes and cheap to reverse means just decide; small stakes and expensive means decide but log the assumption; big stakes and cheap means pilot it; big stakes and expensive means bring out the heavy machinery
The reversibility quadrants. Only one of the four earns the heavy machinery — and most teams misfile decisions in both directions.
The heavy machinery is the Decision Map Workshop, the architecture review, the board memo. Most teams over-apply it — they treat reversible decisions as one-way doors and burn senior attention proving what a two-week pilot would have proven cheaper. Some under-apply it — they treat genuine one-way doors (a platform migration, a data-model rewrite, an AI vendor lock-in) as if they were reversible. The point of the frame is to notice which kind of door you're standing at.
If a decision is cheap and reversible, the right move is to make it quickly and learn. Time spent debating it is more expensive than the wrong answer.
2. The option matrix
Rows are the candidate options. Columns are the criteria the decision will actually be judged on six months later. Cells are short, comparable statements — not adjectives, not stars out of five.
The trap is filling the matrix with criteria that sound important but don't discriminate between the options. "Aligned with strategy" usually gets a checkmark in every row; it does no work. Replace it with the question whose answer differs across rows: which option still works if the vendor gets acquired? which one can the current team run without three new hires?
A useful option matrix has 3–6 rows and 5–8 columns. Beyond that you've stopped comparing and started cataloging.
3. The horizon band
Every option implies a time horizon at which its bet pays off. Some bets pay in 6 months. Some pay in 5 years. Some pay in 50.
Group the options by their horizon, then ask the uncomfortable question: for how many of these horizons does the team actually have the patience and the runway?
Horizon band: candidate options plotted by when their bet pays off along a time axis, with a shaded band marking the team's actual runway — one well-chosen option sits outside the band, flagged as right option, wrong horizon
The horizon band. The most common strategic mistake isn't a bad option — it's a correctly-chosen option sitting just outside the band.
Most strategic mistakes the system sees aren't bad options — they're correctly-chosen options on a horizon the team can't actually wait for. A 5-year bet inside an 18-month runway is the wrong bet, no matter how compelling the analysis.
The horizon band is also where lightweight decision-support meets simulation. Comuvia's decision-support and simulation work exists because some decisions can't be made well without the modeled long view. Most decisions can. Notice which kind you have.
Worked example: local models, hosted APIs, or a hybrid?
Consider an established team with an internal document assistant. Some documents are confidential, usage varies, and the team is deciding how to host inference. This is an illustrative decision exercise, not a client result or a claim that one hosting model is universally cheaper.
Write the decision before comparing products. “Choose how this assistant will serve its first production workload, subject to our permitted data routes, quality threshold, support coverage and review date.” That is answerable. “Which AI platform is best?” is not yet a useful brief.
First: separate the reversible experiment from the commitment
Testing a permitted hosted model on synthetic documents is a different decision from sending confidential production data to it. Running a local benchmark is different from buying hardware or promising a support service.
List the commitments separately: data transfer, model license, hardware, application interfaces, operational responsibility and any minimum contract. For each one, write the exit steps and who can perform them. A portable API interface can reduce application rework; it cannot reverse a data disclosure. A feature flag can help roll out a software change, but it cannot cancel a contract or undo a completed external action. Feature-toggle guidance
Second: compare evidence, with hard gates before preferences
| Criterion | Hosted API | Local inference | Hybrid |
|---|---|---|---|
| Permitted data handling | Check service terms and the approved route | Check access, storage and operator controls | State which data may take each route |
| Quality on representative tasks | Run the same evaluation set | Run the same evaluation set | Test routing mistakes as well as model answers |
| Peak demand and failure behavior | Measure latency, limits and outage handling | Measure queueing, capacity and recovery | Test what happens when either route is unavailable |
| Operational ownership | Name application and vendor-management owners | Name infrastructure and model-serving owners | Name the owner of routing and incident diagnosis |
| Complete operating cost | Include usage, retries and review | Include hardware use, support and idle capacity | Include both routes and routing overhead |
| Exit path | Test export and provider substitution | Test model/runtime replacement | Test that either route can be withdrawn safely |
An entry such as “enterprise-ready” does not settle a row. Prefer a test result, a contract reference, or unknown. If a data-handling requirement is mandatory, a failure excludes the option; a high score elsewhere cannot average it away. Do not assign numerical weights until the decision owner has agreed what they mean. Record disagreements instead of hiding them in a composite score.
Third: test the horizon against the commitment
A hosted route may be the simplest permitted experiment while a local route is being evaluated. A local deployment may fit a stable, controlled workload but require support the team has not yet staffed. A hybrid may be useful, or may introduce more failure modes than the workload warrants. These are hypotheses to test against the actual organization.
Name the review event: a volume change, a material model-quality change, a new contract restriction, or the end of the experiment. Specify what evidence would reverse the recommendation. A review date without a reversal condition tends to become another meeting.
The one-page decision record
Use this structure after the meeting:
- Decision and owner: the bounded choice, the accountable person and the date.
- Options considered: include a small pilot and the option to defer.
- Hard constraints: data, rights, security, support and required quality.
- Evidence: test set, versions, observed results and unresolved unknowns.
- Chosen path: rationale and the trade-off the owner accepts.
- Exit and review: reversal steps, review event and evidence that changes the call.
The output is a choice someone can execute, question and revise. For AI-specific risk questions, the NIST AI Risk Management Framework provides additional context; this worksheet is not a compliance assessment.
Running the three yourself
The sequence matters:
- Reversibility test first. Is this even a workshop-grade decision? If not, decide today and move on — the frame has already paid for itself.
- Option matrix next. Generate, prune, score.
- Horizon band last. Sanity-check the leading option against the team's actual ability to wait for its payoff.
If you run this and end up with a strong leader from the matrix but you can't shake the feeling that something is off — check the horizon. The option is usually right; the patience isn't there.
These three frames are the human-legible front of a larger decision-support capability. When the reversibility test says heavy machinery — one-way door, big stakes, an answer that turns on numbers nobody can defend — Comuvia runs the same sequence as a facilitated engagement: a named option set, an explicit assumption register, and, where the answer turns on numbers, modeled scenarios. The Company AI System that published this article drives those engagements — the frames you just read are the ones it uses on Comuvia's own bets.