THE COURAGE STUDIO.
Free interactive tool

AI DEPLOYMENT DECISION

A prospective tool for deciding whether to deploy an AI tool — and how to do it responsibly — step by step, with a live decision that builds as you go. Please provide your email to access the tool.

Already have access? Open the tool →

Your decision runs entirely in your browser — it never leaves your device.

AI Deployment Decision tool. Work through seven sections; a live decision, including a severity and recourse profile and the resolved decision zone, builds in the panel on the right.

THE COURAGE STUDIO.
AI Deployment Decision Matrix
Critical Analysis Tool by The Courage Studio

AI DEPLOYMENT DECISION

A prospective tool for deciding whether to deploy an AI tool, and how to do it responsibly. For a “GO” decision, a yes has to be earned at each stage. Your decision builds on the right as you go.

A note on how this tool works: This interactive tool was developed to run entirely in your browser. It makes no API calls, uses no generative AI, and nothing you enter is sent or collected. Your progress is saved only on your device, so you can pick up where you left off. It works best on a desktop browser, not mobile.

I’m deciding as an

Before you begin, our contextual anchor

Read first

This framework enforces the burden of proof for responsible AI deployment with the deployer. Refusal (categorical, conditional, or use-case-specific) is a complete and legitimate output. Assuming a “neutral” deployment stance reliably underestimates risk, because it ignores the inequities the technology will enter, encode, and amplify.

Throughout, we insist on making visible the systems of oppression: racism, racial capitalism, settler colonialism, patriarchy, transphobia, homophobia, ableism, and their intersections, as the omnipresent conditions in which all AI tools are developed and deployed.

Governing principle: Whenever possible, the people who would be most impacted by the potential harm should help define what counts as adequate recourse and acceptable risk. A deployment deemed “safe” only by those who benefit from it cannot be judged adequately.

Part 1Grounding

Each stance is legitimate and carries commitments. Assess your current approach based on actual practice, then choose your posture intentionally. Be honest, even where it is uncomfortable. Your stance can change over time, but deliberately, not by drift.

Categorical refusal is a complete decision. A categorical-rejection stance does not require running the rest of the framework for your own deployments. The tool remains useful for evaluating AI imposed on you by others, for auditing what others deploy, and for understanding the field.
Non-negotiables must be testable“We won’t use AI irresponsibly” is not an enforceable boundary. “We will not deploy AI tools to make hiring decisions where the applicants cannot contest it” is a much better starting place.
Sample boundaries for reference
  • No tools trained on scraped creative work without consent or compensation
  • No AI use for grant applications screening and decision making
  • No personal financial or medical data sharing to consumer chatbots
  • No tools from companies on the BDS priority and pressure target lists
  • No AI-generated content published without disclosure
  • No vendor that denies the right to audit or to exit with your data
Does the deployment you’re evaluating cross any non-negotiable you hold?A crossed redline is a refusal. The decision is already made.
Part 2Analysis
Who is in the room?Are the most-impacted involved in this assessment? If the answer is no one, name that honestly.
Intended autonomy levelHow much autonomy will be given to the AI tool? Reference levels by the Cloud Security Alliance.

Most AI deployments carry baseline costs paid by default: the labor of data labeling, the non-consensual extraction of training data, the energy and water footprint, and the concentration of power in those who own the infrastructure. Proportionality asks whether the benefit of this deployment justifies the apparent and hidden costs.

WHAT DETERMINES PROPORTIONALITY
  • Necessity: a real need, or a manufactured one?
  • Non-AI alternatives: what was considered, and why are they insufficient?
  • Expected benefits: specific, with evidence behind the claims?
  • Expected costs: tangible and intangible, defensible against the benefit?
  • Right-sized option: proportionate, or overbuilt for the work?
  • Distribution of benefit & risk: who captures it, who bears it, is that justified?

Is any of the following true? Check all that apply. Any one means the use case fails the proportionality test.

Part 3Deployment decision

Begin the pre-mortem process. Imagine the tool has caused harm—what are the types of harm caused by the tool? Name and analyze all that apply.

Score by the worst plausible case and the most vulnerable affected party, accounting for the autonomy you’d grant the tool. The more it can act without a checkpoint, the wider the worst plausible case.

Overall severitySevere if Intensity = 3 or Equity of Impact = 3. Significant if two or more dimensions score 2 or higher (and not already Severe). Limited otherwise. Intensity and Equity can each set the level on their own.

If the system gets it wrong, can the affected party know, appeal, recover, and be made meaningfully whole? Has the infrastructure been proven effective at doing so?

Has adequate recourse been demonstrated in practice?Recourse counts as available & operationalized only if it has been proven to work, not just designed on paper.
Overall recourseUnavailable if either dimension scores 3: the harm cannot be undone, or the person has no agency to contest it. Unproven if recourse has not been demonstrated in practice. Available & operationalized when both dimensions are workable and recourse has been demonstrated in practice.

A decision is only as durable as its record. Putting the decision in writing (with the tradeoffs and design choices made explicit, and who was in the room) is what makes a “go” accountable and a “no” defensible.

NEXT STEPS