Every founder, CTO, or innovation lead carries a short queue of questions that are too expensive to get wrong: which platform to bet on, how to sequence a modernization, whether the AI operating model gets built or bought. On questions like these, two failure modes quietly eat the budget:
- The team commissions weeks of analysis on a question only a decision can settle — modeling as a sophisticated way of not deciding.
- Or it forces a Friday decision on a question that genuinely needed a model, and the mistake surfaces two quarters later as a rebuild.
Models and decisions aren't substitutes. They have different jobs, and knowing which one your question wants is most of the work — decision-makers need modeled scenarios where scenarios help, and a clean decision where they don't. The triage below is a practical way to make that distinction. It is a decision method, not a claim that every Comuvia job passes an automated modeling gate.
Start with the action the answer could change
Write three things before choosing a model: the decision, the available actions, and the result that would change the action. If no plausible answer changes what you will do, additional simulation has little decision value. If the question is who has authority or whether a use is permitted, establish that constraint directly; a simulation cannot grant permission.
For example, “How efficient is our content workflow?” is too broad. “Should we increase rendering capacity, simplify the format, or improve the review step before accepting a recurring series?” is a decision with alternatives.
Worked example: a content production bottleneck
Suppose a team wants to publish reviewed technical explainers. It believes rendering is the bottleneck and is considering more compute. The workflow also includes source preparation, script review, production, final review and release. This is a hypothetical planning example. The observations below are things to collect, not performance results from Comuvia or a client.
1. Measure the existing workflow before simulating a new one
For a representative set of completed jobs, capture when each stage became ready, when work began, when it finished, and whether it returned for repair. Separate active work from waiting. Record why jobs were held: missing sources, reviewer availability, rendering failure, factual correction or release approval.
If most elapsed time is waiting for source review, faster rendering alone may have little effect on accepted output. If scripts arrive in bursts and a render queue grows while reviewers are available, compute may deserve a closer test. An average turnaround time cannot reveal that difference by itself.
2. Begin with a small capacity calculation
A first approximation compares the sustainable capacity of the stages. The slowest required stage constrains steady accepted output. That is a diagnostic, not a complete queueing model: rework, shared staff, variable job sizes and interruptions can change the result.
| Observation | Useful next step | Why |
|---|---|---|
| A single stage has an obvious, repeated wait | Fix or pilot that stage first | A complicated model may not change the action |
| Several stages share one reviewer or machine | Represent the shared resource | Independent averages would overstate capacity |
| Demand arrives in bursts | Test a range of arrival patterns | Mean demand can hide queueing at busy times |
| Rework is common | Model repair paths and acceptance | More drafts are not the same as more accepted work |
| Little trustworthy history exists | Collect a bounded baseline or run a reversible pilot | Detailed distributions would be invented precision |
3. Model only the uncertainty that changes the choice
A small queue or discrete-event model can compare additional render capacity, a simpler video format, a different review schedule and the current workflow. Use measured ranges where available. Label assumptions where observations do not exist. Keep the same demand, acceptance rules and cost boundary across options.
Test sensitivity: does the leading option remain attractive when repair rates rise, a reviewer is unavailable, or demand is lower than expected? If the ranking reverses under a small plausible change, report the fragile assumption instead of choosing a dramatic-looking chart.
4. Validate against observations the model did not use
Check that the model reproduces the behavior it is supposed to explain on a separate period or job sample. Reproducing a fitted average is weaker evidence than explaining queues, waits and accepted output in another sample. Historical fit does not establish that a future intervention will work; a limited production pilot may still be the next appropriate action.
A useful recommendation might be “test a protected review window before buying more capacity, because the decision changes with reviewer availability.” It might instead favor compute. The analysis earns its place by discriminating between actions, not by making the final choice sound inevitable.
A brief you can give an analyst
Specify the decision owner, alternatives, time horizon, outcome measure, available data and material constraints. Ask for a baseline, an assumption register, sensitivity checks, validation evidence and a written explanation of what the model cannot conclude. Require a versioned result and a review event.
Keep conditional scenario, forecast, and observed outcome separate. A stress gauge is not automatically a probability. A simulated improvement is not a measured saving. If the evidence cannot support the proposed decision, state the missing observation and the smallest useful experiment.
What a model is good at
- Examining ranges. A scenario that's "good or bad depending on three assumptions" is a model's home turf. Vary the assumptions; watch the output band; report the sensitivity.
- Catching proposals that don't balance. If flow doesn't match stock, demand doesn't match supply, or revenue doesn't cover cost over time, a stock-flow consistent model surfaces it before the spreadsheet does.
- Comparing structural alternatives. Three architectures, two governance regimes, four sourcing mixes: a model shows what each implies, faster than a strategy off-site can.
- Quantifying the boundary. What would the world have to look like for option A to beat option B? A model turns that from a debate into a number you can monitor.
What a model is bad at
- Choosing between politically loaded options. That's a decision, not a calculation. A model will inform it; it cannot replace it.
- Working without a known mechanism. A model needs a causal structure. Without one you're curve-fitting, and the curve will mislead you with confidence.
- Beating the clock. A model built well takes days to weeks. If the decision is due in 24 hours, use a decision frame instead — see three decision frames we use most.
- Surviving without governance. A model with no documented assumption set, no version control, and no audit trail is a liability wearing the costume of rigor. Treat it like code.
The three-question test
Triage flow: three questions — causal structure on one page, inputs measurable today, a named runtime bound — routing a hard question to research-first, decide-first, scope-as-research, or model-it outcomes
The triage gate. Only a question that clears all three checks earns a model; every other path is decision-support territory, and it's cheaper to find that out here than three weeks into a build.
- Can you write down the causal structure on a page? If yes, you have a candidate model. If no, you need a research step before any model — the kind of mechanism-mapping a Decision Map Workshop exists to produce.
- Are the inputs measurable now, or do they depend on a strategic choice you haven't made yet? Measurable now → model. Waiting on a choice → decide first, model second. No model can pick your strategy for you.
- Can you name a bound on the model's runtime cost? If you can't, you're not commissioning a model — you're commissioning a research project. Those cost different money and answer different questions.
Two yeses and a named runtime budget → modeling is the next step. Anything else → you're in decision-support territory, not modeling territory.
What "model-grade simulation" looks like
Comuvia's simulation work — the Simulation Decision Pack when packaged, the underlying ForeGlass Engine system when used directly — is built around four disciplines the system won't compromise:
- Stock-flow consistency. Every model that touches money or physical flows is checked against accounting identities. If it doesn't balance, it doesn't ship.
- Versioned assumption sets. Every run carries a written assumption set, traceable to the model version that produced it. When the recommendation changes, the audit trail tells you why.
- Bounded scenarios, not open exploration. A finite set of named scenarios with a runtime budget. Not "let's see what comes out."
- Executive-ready narrative. The output is a memo with the assumption table beside it — something a buyer can audit — not a 40-tab spreadsheet nobody can.
Pipeline: a business question passes through four discipline gates — stock-flow consistency, versioned assumptions, bounded scenarios, executive-ready narrative — to become a defensible recommendation; removing any one leaves an interesting demo
The four disciplines as a pipeline. The combination is what lets a simulation carry weight in a decision; remove any one and you have an interesting demo — decoration on a decision, not support for one.
This page is itself an instance of the rule: Comuvia is the AI-managed organization we advise enterprises to build, and the same governance the system applies to its own published work — versioned, auditable, bounded in cost — is what it applies to a model before that model is allowed to inform your call. See the full stack at Company AI System.
When the answer is "neither"
Sometimes the right answer is don't model, don't workshop — ship the smaller version. That's most decisions. Models and workshops are heavy machinery; reach for them when you can't proceed without them, not when you want to feel rigorous about a call you already know how to make.
And if you're unsure which side of the line your question is on, that triage is the first thing Comuvia's decision-support and simulation work does — deliberately, because it's the cheapest step in the engagement and it protects you from paying for the wrong instrument.
Bottom line. Achieved: the three-question triage gate Comuvia's Company AI System applies to its own work before committing compute to any simulation, published in full — with the four disciplines that separate a decision-grade model from a demo. Business value: you can tell in one sitting whether your expensive open question needs a model, a decision, or neither — before you pay for the wrong one.