# Start With the Decision - Complete Video Narration ## Opening Welcome to SRAI Book 7, Applied AI Studio: National Sector Intelligence. This lesson is called Start With the Decision. Many artificial-intelligence projects begin with a dataset, an algorithm, or a vendor proposal. This lesson begins somewhere else: with the public decision that an institution is legally and operationally responsible for making. Before selecting a model, we must answer five questions. Who owns the decision? What intervention is lawful? How quickly must the institution act? What are the consequences of different errors? And when must the analytical system abstain or escalate to human review? The model is one component of this process. It is not the decision-maker. ## The decision contract We turn these questions into a decision contract. The contract names the public problem, the accountable owner, the precise decision, the decision horizon, the permitted intervention, the costs of false positives and false negatives, the abstention conditions, the escalation route, and the review cycle. This structure prevents a technically impressive output from being mistaken for an authorized institutional action. A prediction can be accurate and still be unusable if nobody has authority to act, if the data arrive too late, if the intervention is unlawful, or if the institution lacks the capacity to respond. ## A worked agricultural example Consider an agricultural sampling frame that may omit or misclassify active holdings. The accountable owner is the national statistical office or the responsible ministry of agriculture. The decision is not whether a model has discovered an emergency. The decision is which areas should receive frame verification before sample selection. The intervention is targeted listing and frame reconciliation. A false positive may produce an unnecessary field visit and additional cost. A false negative may preserve undercoverage and introduce bias into the frame. If coverage diagnostics are stale or geographic identifiers fail validation, the system must abstain and refer the matter through the defined survey-governance route. The analytical system may identify areas requiring verification. It does not declare a food-security emergency, allocate public resources, or authorize field operations by itself. ## Policy loss and model error Model error and policy loss are not the same thing. A false negative may be much more consequential than a false positive, or the reverse. The relative cost depends on the public decision and must be documented by accountable people. It is not a value that a model should invent. In the notebook, three illustrative scenarios produce different policy-loss totals. The calculation is intentionally simple. Its purpose is to make the assumed costs visible and contestable. Decision-makers can then ask whether the weighting reflects the real consequences, whether important groups face unequal harm, and whether the proposed intervention remains proportionate. ## The actionability gate High accuracy is not enough. The actionability gate checks whether the decision has a named owner, a defined intervention, a positive decision horizon, lawful authority, operational capacity, and sufficiently current data. If all six conditions are present, the candidate passes this limited gate. Passing does not mean that deployment is automatically approved. It means only that the proposal has cleared one necessary test and may proceed to the remaining statistical, legal, ethical, security, accessibility, and institutional reviews. If a critical condition is absent, the correct response is to pause. The absence should be recorded as a decision-relevant finding rather than hidden behind a performance score. ## Abstention and escalation Abstention is a valid analytical output. When data quality is inadequate, when the estimated probability lies in an uncertain band, when linkage fails, or when distribution shift makes prior validation unreliable, the system should say that the evidence is insufficient. In the notebook example, strong evidence can route a case to authorized human review. Low estimated risk can support routine monitoring. An uncertain result produces abstention. Poor data quality also produces abstention, even when the apparent probability is high. This is not indecision. It is a safeguard against false precision. ## Transfer to education Now consider an attendance and dropout early-warning service. The public problem is that persistent absence may precede dropout. The accountable owners are the district education authority and the school safeguarding lead. The decision is which schools or learners require proportionate review and support. The permitted intervention is human review followed by appropriate learner support. A false positive can create unnecessary review or stigmatization. A false negative can miss an opportunity to prevent disengagement. If attendance records are incomplete, delayed, or linked to the wrong person, the system must abstain and refer the case through the safeguarding and education-authority route. The system may organize evidence and identify a case for review. It may not impose sanctions, exclude a learner, or determine eligibility automatically. ## Applying the contract Choose a sector such as health, education, agriculture, or habitat. Name the accountable decision owner and the lawful action. Define the evidence, decision horizon, and failure conditions. Then state when the system must abstain, escalate, be monitored, or be retired. A strong decision contract allows a reviewer to see what the system knows, what remains uncertain, what it cannot decide, and who remains accountable. ## Closing Before deployment, return to the five questions. Is the owner named? Is the intervention lawful? Is the horizon operationally realistic? Are error trade-offs explicit? Are abstention and escalation defined? If one critical answer is missing, pause and refer the proposal for review. Start with the decision. Then determine whether data and modelling can support it responsibly. This lesson uses synthetic instructional examples. It does not authorize an operational decision or replace sector-specific legal, ethical, statistical, accessibility, security, and institutional approval.