Page Title

Privacy and Security

P

We embed privacy and security from day one: consent and minimization, encryption in transit/at rest, least‑privilege access with MFA, and a clear incident path. We prefer de‑identified data where feasible, avoid uploading sensitive info to general models, and use DPAs/BAAs when regulated data is involved. Before scale, teams document user notices, opt‑outs, and human override. Evidence: policy/consent snippet, access matrix, incident SOP, vendor terms.

Relevance & Urgency

R

We fund AI when it’s the best tool for a real bottleneck—throughput, wait times, accuracy, or reach—not because it’s fashionable. A needs assessment compares AI to simpler options and shows the work wouldn’t happen (or wouldn’t work) otherwise. We ask for a 60–90‑day pilot plan tied to specific benefits and equity outcomes. Evidence: gap statement, alternatives considered, early proof plan, ROI + equity metrics.

Attribution and Metrics

A

Belief isn’t proof. We require a causal plan (A/B, shadow, or DiD), baselines, and owners/cadence. Track three lenses together—efficiency (time/capacity), safety (errors/incident severity), and equity (subgroup outcomes)—so averages don’t hide harm. Show confidence intervals and pre‑declared thresholds for go/adjust/stop. Evidence: 1‑page eval plan, baseline vs pilot chart with CI, subgroup table, rollback triggers.

Cost Realism

C

AI isn’t “free once built.” We expect line‑item TCO—build, run, govern—plus usage‑based forecasts (volume × tokens × price) and controls to prevent runaway spend. Include scale math (10× volume), caching/cheaper‑model strategies, and a named cost owner. If vendor pricing shifts or usage spikes, the plan should still hold. Evidence: TCO sheet, spend controls, scale scenario, fallback model choices.
Report‑only: Environmental footprint (ops)—estimated CO₂e range + reduction plan.

Timeline Clarity

T

Pilots drift without calendar discipline. We ask for a 30/60/90 plan with go/adjust/stop checkpoints, pre‑registered success criteria (efficiency, safety, equity), and scoped protections (e.g., assistive use before automation). Timebox the unglamorous work—data access, policy reviews, staff training—so you can ship safely and decide with evidence. Evidence: dated milestones, success thresholds, mid‑point review, decision memo template.

 Implementability (Feasibility)

I

Great ideas fail on people, data, and workflow fit. We assess skills and capacity, data quality/rights, integration into real processes, and accessibility (devices, languages, assistive tech). If there are gaps, the plan names partners and sequencing. Usability checks with real users are a must. Evidence: capability map, data readiness notes, integration diagram, training plan, accessibility checklist, usability findings.

Change Resilience

C

Models drift; vendors change. We expect version pinning, a change log, shadow/A‑B for new features, watch‑metrics with alert thresholds, and safe‑degrade modes (tighten human‑in‑the‑loop or fall back to human‑only) if metrics slip. Document how you’ll adapt to policy/vendor shifts and keep users informed. Evidence: drift dashboard, deployment plan, fallback matrix, change log.

 Accountability & Oversight

A

Humans own outcomes. We ask for named owners (product/ops, metrics, incident), a kill switch, rollback triggers, and a simple incident process (detect → notify → fix → learn). Public‑facing transparency about AI use builds trust; internal governance materials speed approvals. Evidence: owner list, policy snippet, incident SOP, transparency text, oversight cadence.

Lived Expertise & Trust

L

Trust is earned with people, not just metrics. We look for co‑design with frontline staff and communities, plain‑language notices, feedback/appeals routes, and proof of adoption by subgroup (not just topline). The goal is tools people choose to use—and that improve outcomes for those furthest from service. Evidence: co‑design notes, consent/notice copy, feedback pipeline, usage/retention by subgroup.