aaron

Executives don’t read; they decide. Your brief wins if it removes hesitation in under 60 seconds. Here’s the automation that makes that happen every day without drift.

The problem: Summaries are easy; decision briefs are not. Most outputs miss owner+timeline, bury the implication, or waffle on confidence. At scale, inconsistency erodes trust and adoption.

Why it matters: Reduce executive read time by 70% and lift action rate. That’s fewer meetings, faster bets, and cleaner accountability. You’ll reclaim hours weekly across the leadership team.

What you’ll need (simple stack):

• An AI chat tool or lightweight API.
• Sources: 1–3 articles/reports/transcripts each brief.
• A fixed template (headline, 3 implications, single action with owner+timeline, confidence, read time, citation).
• A short ban list (e.g., leverage, synergy, robust, transformative, paradigm, best-in-class).
• One reviewer for early runs (30–90 seconds per brief).

Field-tested lessons: Two-pass synthesis is the right spine. Add three small locks to make it enterprise-grade: a freshness/dedupe gate, a persona router, and an “action calculus” line (impact × confidence ÷ effort) to prioritize at a glance.

Build the pipeline (8 steps):

1. Define decision types: Monitor, Decide, Delegate. Only “Decide” requires an action with owner+timeline.
2. Pass 1: Facts-only extract with numbers, named entities, dates. No opinions. (Use the prompt below.)
3. Freshness + dedupe: Drop items older than X days and block briefs that share ≥70% of facts with the last 14 days.
4. Pass 2: Executive synthesis: 1 headline; 3 implications; 1 action (owner+timeline+scope); confidence with 5–7 word reason; read time; citation.
5. Persona router: Rewrite wording for CEO/CFO/COO/CMO without changing the action or confidence.
6. Action calculus: Score each brief 1–5 for Impact, Confidence, Effort (inverse). Include “ICE Score = (I × C) ÷ E”.
7. QC pulse: Auto-score clarity, action specificity, verifiability, brevity, and ban-list cleanliness. Gate on thresholds.
8. Distribution: Subject line = headline; body = fields; route by persona; publish at a fixed daily time. Keep a one-click feedback (“Helpful? Yes/No”).

Copy-paste prompt — Pass 1 (Facts-only)

“You are a facts-only extractor. From the source, list the 5 most decision-relevant facts with numbers, named entities, and dates where available. No opinions, no suggestions. Each fact under 20 words. Then add one line with the source name and date. Output exactly:
Facts:
– [Fact 1]
– [Fact 2]
– [Fact 3]
– [Fact 4]
– [Fact 5]
Citation: [Source name, Date]
Source text: [PASTE SOURCE]”

Copy-paste prompt — Pass 2 (Executive brief + JSON)

“You are an executive brief writer for [ROLE e.g., CEO/CFO/COO/CMO]. Using only the facts list below, produce both (A) a human-ready brief and (B) a strict JSON for logging.
Rules: Professional, plain language. Total words ≤75 in the human brief. Ban: leverage, synergy, robust, transformative, paradigm, best-in-class. Do not explain your reasoning.
Human brief fields:
– Headline (≤12 words)
– Takeaways (3 bullets, each ≤14 words, state impact + implication)
– Action (one sentence with owner + timeline + scope)
– Confidence (High/Medium/Low + 5–7 word reason)
– Read (≤60s)
– Citation (source name, date)
– ICE Score (Impact 1–5 × Confidence 1–5 ÷ Effort 1–5)
JSON schema keys: headline, takeaways (array), action, confidence_level, confidence_reason, read_seconds, citation, impact_score, confidence_score, effort_score, ice_score.
Facts:
[PASTE FACTS OUTPUT]
[PASTE CITATION LINE]”

Auto-QC prompt (score and rewrite if needed)

“Score the brief 1–5 on: Decision Clarity, Action Specificity (owner+timeline+scope), Verifiability (facts → citation), Brevity (≤75 words), Ban List Clean. If any score <4, rewrite once to meet thresholds without changing the action owner or timeline.”

What to expect: 3–5 tuning cycles to lock tone and length. After that, editor review drops to 30–90 seconds per brief; exec scan time stays under 60 seconds. Persona rewrites add ~10 seconds.

Metrics that matter (track weekly):

• Decision rate: % briefs that trigger a documented next step in 14 days (target ≥40%).
• Time saved: (Old read time − new read time) × readership (target ≥70% reduction).
• Reviewer correction rate: % briefs needing edits (reduce to ≤5%).
• Confidence calibration: % actions later reversed (target ≤10%).
• ICE alignment: % shipped briefs with ICE ≥3.0 (target ≥80%).

Common mistakes & fixes:

• Vague action → Force owner, timeline, and scope in the prompt; reject if missing.
• Overstuffed takeaways → Cap each at ≤14 words; use a ban list.
• Confidence inflation → Require a 5–7 word reason and track reversals.
• Duplication → Add freshness window (e.g., 21 days) and fact-similarity check.
• Persona mismatch → Lock the action; only rewrite wording by role.

7-day action plan:

1. Day 1: Finalize the template, ban list, and decision types. Copy the prompts.
2. Day 2: Run 10 briefs manually (mixed sources). Tune word limits to keep ≤75 words.
3. Day 3: Add persona rewrites for CEO, CFO, COO. Validate the action still fits owners.
4. Day 4: Implement freshness/dedupe and the QC scorer. Set ship threshold (all metrics ≥4).
5. Day 5: Pilot with 3–5 execs. Ask only: “Would you act on this?” and why.
6. Day 6: Automate batch ingestion (RSS/shared folder → Pass 1 → Pass 2 → QC). Keep the reviewer gate.
7. Day 7: Review KPIs (decision rate, correction rate, time saved). Trim prompts until correction ≤5%.

Expectation: By brief #10, outputs feel on-brand. By #30, the action line earns trust. Keep the structure frozen and enforce the QC gate—consistency is the product.

Your move.

— Aaron

Quick win (under 5 minutes): Score your next proposal with this 7-question card. Give 1 point for each “yes.” If total ≥ 5 → default to hourly or hybrid; ≤ 3 → fixed price; 4 → hybrid.

• Is the scope still evolving?
• Are there 3+ stakeholders or approvers?
• Are there dependencies you don’t control (third parties, data access)?
• Is there unfamiliar tech or tools?
• Is the deadline rigid with penalties or launch pressure?
• Are “unlimited revisions” or creative/strategy components expected?
• Has this client changed scope on you before (or lacks a brief)?

The problem: Freelancers choose hourly vs project pricing by feel, not facts. That creates margin swings, scope creep, and awkward renegotiations.

Why it matters: A consistent rule reduces discounting, raises win rate, and keeps your calendar and cash predictable.

Lesson from the field: Use a risk score + historical hours. Price three ways every time (Hourly Safe, Fixed Value, Hybrid Guardrails). Let the numbers decide, not the mood.

1. Build your decision system
   1. Calculate two anchors from past projects: Median hours (P50) and 75th-percentile hours (P75).
   2. Use the 7-question risk card above. Score 0–7.
   3. Map decision: 0–3 = Fixed; 4 = Hybrid; 5–7 = Hourly or Hybrid.
2. Set prices with simple formulas
   1. Fixed = P75 × Hourly Rate × (1 + Contingency%). Contingency% by risk: 0–1 = 10%, 2–3 = 20%, 4 = 30%.
   2. Hourly (with client comfort) = Hourly Rate, estimate P50–P75 range, add a soft cap at 1.2× P50, weekly reporting, change-order process.
   3. Hybrid = Fixed for clearly defined core scope (e.g., P50 hours), includes X revisions + Y included hours; out-of-scope at Hourly Rate with approval after Z hours.
   4. Optional discovery (high risk): sell a small paid discovery (5–10 hours) to de-risk, then reprice with clarity.
3. Spreadsheet setup (copy/paste)
   1. MEDIAN of actual hours: MEDIAN(B2:B13)
   2. 75th-percentile hours: PERCENTILE(B2:B13,0.75)
   3. Variance% each row: =(Actual/Estimate)-1; flag overruns >20%.
   4. Fixed price cell: =P75 * HourlyRate * (1+Contingency%)
   5. Hourly soft cap: =P50 * 1.2
4. Use AI as your second opinion (not your boss)

   Copy-paste AI prompt:

   “You are my pricing analyst. I’ll give you: 1) a short project description, 2) my historical stats (P50 hours, P75 hours, % of past projects with >20% overrun, change-order frequency), 3) my hourly rate and target margin, 4) answers (Yes/No) to these risk questions: evolving scope, 3+ stakeholders, external dependencies, unfamiliar tech, rigid deadline, creative/revisions-heavy, client prone to scope change. Do the following:
   – Estimate hours at P50, P75, P90 and explain the drivers in plain English.
   – Produce a risk score 0–7 from the Yes answers.
   – Recommend one model (Hourly, Fixed, Hybrid) and explain why in one paragraph.
   – Price three options:
   A) Hourly Safe: hourly rate + estimated range + soft cap at 1.2× P50, weekly reporting.
   B) Fixed Value: price = P75 × rate × contingency (10–30% based on risk). State included deliverables and revisions.
   C) Hybrid Guardrails: fixed price for core scope with X included hours + overage at hourly rate after approval at Y hours.
   – Provide 3 short contract clauses: change-order process, revision limits, mid-project review. Keep output to bullet points I can paste into a proposal.”

   What to expect: clear hours band (P50/P75/P90), a reasoned risk score, and three concise price structures you can paste into a proposal. Validate against two similar past projects before sending.

5. Proposal language (paste-ready)
   • Change orders: “Out-of-scope items are quoted in writing and approved before work. Pricing = hourly rate × estimated hours for the change.”
   • Revisions: “Includes two revision rounds. Additional revisions billed hourly with prior approval.”
   • Midpoint review: “At 50% of estimated hours, we review progress and confirm scope. Adjustments require a signed change order.”

Metrics to track (weekly dashboard):

• Margin by model (Fixed vs Hourly vs Hybrid).
• Estimate accuracy: Actual Hours ÷ Estimated P50 and vs P75.
• Overruns: % of projects exceeding P75 hours.
• Change-order revenue % of total (healthy: 10–25% on complex work).
• Proposal win rate by model and by option (A/B/C).
• Average negotiation cycles per proposal (aim to reduce by explaining the rule upfront).

Common mistakes & fixes:

• Mistake: Quoting one option. Fix: Always show Hourly, Fixed, Hybrid to anchor value and let clients choose risk.
• Mistake: Flat 10% contingency for everything. Fix: Tie contingency to the risk score.
• Mistake: Unlimited revisions in fixed bids. Fix: Define rounds; move extras to hourly via change order.
• Mistake: Ignoring discovery. Fix: Sell a paid discovery when risk ≥ 5; reprice after.
• Mistake: Trusting AI blindly. Fix: Cross-check with at least two similar projects and your P75.

One-week action plan:

1. Day 1: Pull 6–12 past projects; calculate P50, P75, and % overrun >20%.
2. Day 2: Add the 7-question risk card to your proposal template; set contingency by risk.
3. Day 3: Run the AI prompt on 3 active opportunities; produce A/B/C pricing options.
4. Day 4: Insert the three proposal clauses; add a midpoint review to every contract.
5. Day 5–6: Present proposals; track client preference (Hourly vs Fixed vs Hybrid), objections, and win/lose.
6. Day 7: Review metrics; adjust contingency bands or thresholds if P75 is consistently off.

Target outcomes in 30 days: +5–10 margin points on fixed bids, 25–40% fewer overruns beyond P75, and a faster yes/no from clients thanks to clearer options.

Your move.

Quick win (under 5 minutes): Paste 20 open-ended responses into an AI chat and run this prompt to get immediate sentiment (+1/0/-1) and a single theme for each response — you’ll have structured data to analyze in minutes.

The problem: Open-ended survey answers are rich but messy. You can’t run percentages on verbatims without turning them into numbers: sentiment scores and repeatable themes.

Why this matters: Quantifying sentiment and themes converts qualitative insight into KPIs you can track over time, tie to NPS/CSAT, and prioritize action. You’ll know what to fix, measure impact, and show ROI.

Short lesson from the field: You don’t need a data scientist to get useful results. Two reliable approaches: 1) rule-based/classifier prompts for sentiment + manual taxonomy; 2) embeddings + clustering to discover themes at scale. Combine both for best accuracy.

1. What you’ll need
   • A CSV or spreadsheet of responses (text column).
   • Either access to an LLM (chat UI or API) or a simple tool that supports embeddings/clustering.
   • A small validation sample (50–200 responses) for tuning.
2. How to do it — step-by-step
   1. Clean: remove duplicates, trivial spam, and anonymize any PII.
   2. Quick sentiment pass: run the prompt below to tag each response as Positive/Neutral/Negative and give a short rationale.
   3. Theme extraction: either ask the AI to assign one primary theme from a short taxonomy, or generate embeddings and run k-means/UMAP to reveal clusters (useful when you don’t have a taxonomy).
   4. Validate: sample 100 tagged items, calculate agreement vs. human labels, and adjust prompts or cluster count.
   5. Aggregate: produce counts, sentiment-weighted theme scores, and a dashboard-ready CSV.

Copy-paste AI prompt (sentiment + theme)

Paste the response below. Return JSON array with fields: id, sentiment (Positive/Neutral/Negative), sentiment_score (1/0/-1), theme (one short label), brief_reason (one sentence).

Example instruction:

“Read the customer comment. Classify overall sentiment as Positive, Neutral, or Negative and assign a sentiment_score (1, 0, -1). Then assign one concise theme label (e.g., Pricing, Customer Service, Product Quality, Onboarding, Feature Request). Finally, give a one-sentence reason. Output as JSON only.”

What to expect — accuracy and time:

• Initial automated agreement vs human: 75–90% for sentiment, 60–85% for themes (improves with validation).
• Processing time: minutes for hundreds via chat batching; seconds via API/embeddings per 100s.

Metrics to track

• Sentiment distribution (% Positive/Neutral/Negative)
• Theme frequency and share of negative comments per theme
• Human-AI agreement rate (validation sample)
• Change over time (week/month) and correlation with NPS/CSAT

Common mistakes & fixes

• Too broad taxonomies — fix by consolidating to 6–8 actionable themes.
• Relying only on raw LLM labels — fix with a validation sample and simple rules (e.g., negative if contains “cancel” or “refund”).
• Ignoring context (sarcasm) — fix by adding the one-sentence reason requirement and reviewing low-confidence items manually.

1-week action plan

1. Day 1: Export responses and clean data (remove PII, duplicates).
2. Day 2: Run quick-win prompt on 50–100 items; review results.
3. Day 3: Create an initial taxonomy of 6–8 themes.
4. Day 4: Run full sentiment + theme pass (batch or API).
5. Day 5: Validate 100 items, measure agreement, refine prompts/rules.
6. Day 6: Produce dashboard CSV and top 5 action items by negative volume.
7. Day 7: Present findings and set the next review date (weekly or monthly).

Your move.

— Aaron

Make privacy your product spec. Write down the non-negotiables, only run tools that meet them in writing, and you’ll ship safe AI faster than districts that debate.

The problem: Vendors market “for education,” but few lock down data use, retention, or training. That’s COPPA/FERPA risk and parent backlash waiting to happen.

Why it matters: Once student data trains a model, you can’t pull it back. Contracts and configuration are what protect kids, budgets, and your reputation.

Lesson from the field: Districts that enforce three non-negotiables win — 1) no model training on student data, 2) short, automatic deletion, 3) school-managed accounts only. Add a tight pilot, and you’ll keep good AI while cutting risk to near-zero.

• Do: Require written commitment to no model training on student data.
• Do: Enforce auto-delete ≤90 days and deletion on request within 30 days.
• Do: Use school-managed SSO; disable personal or home accounts.
• Do: Prefer tools with zero-retention/logging modes or on-prem/tenant options.
• Do: Minimize inputs (no full names, IDs, faces, voices) unless contractually required.
• Do: Keep a pilot incident log and a parent-facing summary.
• Do Not: Accept vague policies like “may use data to improve services.”
• Do Not: Allow indefinite retention, voiceprints, biometrics, or location collection.
• Do Not: Let teachers upload student work that includes PII without a signed DPA.

Gold-standard settings to demand (put in the DPA and admin console)

• No training on any student data, metadata, or derivatives; include subcontractors.
• Data residency disclosed; encryption in transit and at rest (AES-256 or equivalent).
• Auto-deletion: raw inputs ≤30–90 days; logs/anonymized analytics off by default.
• Granular parental rights: access, correction, deletion; 30-day SLA.
• Breach notification: within 72 hours; incident report with scope and remediation.
• SSO only, role-based access; audit logs available to district.

1. Define your privacy floor (1 hour): Copy the gold-standard list above into a one-page “Privacy Requirements” doc. This is your baseline for every tool.
2. Shortlist safer tool types (30 minutes): Start with teacher-assist tools that avoid student accounts, local/on-device utilities (speech-to-text, translation), and district-hosted chat with zero-retention toggled on.
3. Screen vendors (1 week): Send your questionnaire plus DPA. Require explicit answers on PII, model training, retention, deletion, subcontractors, and SSO.
4. Configure and pilot (2–4 weeks): Turn off logs/analytics, use anonymized content, and run a supervised pilot with an incident log.
5. Decide and communicate (48 hours): Approve with signed DPA and safe settings, or pause and escalate. Publish a parent-facing summary per tool.

Metrics that prove you’re in control

• % tools with signed DPA and no-training clause (target: 100% before rollout).
• Average retention window across tools (target: ≤90 days; stretch: ≤30).
• Vendor response time to questionnaire (target: ≤7 days).
• Incidents per 100 students during pilot (target: 0 critical, ≤1 minor).
• Staff adherence to anonymization checklist (target: 95%+).

Worked example — “ClassCoach AI” (writing feedback)

1. Vendor answers: Collects essay text and first name; claims “aggregated data may improve AI;” default retention 1 year; SSO optional.
2. Risk: Medium-High (PII + training + long retention).
3. Mitigations required:
   • Written clause: no model training on any student data (direct or aggregated).
   • Auto-delete raw data after 60 days; deletion on request ≤30 days.
   • SSO-only access; disable personal accounts; logs/analytics off.
   • Parent rights process documented; 72-hour breach notice.
4. Pilot setup: Teacher-only mode; anonymize essays; one class for 2 weeks; incident log active.
5. Decision: Approve only if all clauses signed and settings enforced. Otherwise, pause.

Common mistakes and fast fixes

• Mistake: Trusting “FERPA-compliant” badges. Fix: Require clause-by-clause commitments and signatures.
• Mistake: Uploading full student work. Fix: Use excerpts or synthetic samples in pilots.
• Mistake: Letting teachers create personal accounts. Fix: Enforce SSO-only, block personal signups at the firewall.
• Mistake: Keeping analytics logs on. Fix: Default to zero-retention; audit monthly.

Copy-paste AI prompt (use with any general LLM)

“You are a K–12 Data Protection Officer. Evaluate the AI tool info below for COPPA/FERPA risk. Output: 1) Traffic-light risk (with rationale); 2) Exact list of data elements collected and why each is sensitive; 3) Required safe configuration settings (admin console toggles and account rules); 4) Contract language: no training on student data; retention ≤90 days; deletion on request ≤30 days; SSO-only; 72-hour breach notice; subcontractor flow-down; 5) Parent-facing summary (2–3 sentences) describing protections; 6) Pilot checklist and red flags that trigger a pause.”

What to expect: Vendors that can’t accept “no training” and short retention usually stall or go quiet. That’s a time-saver. The ones who engage are the ones you can scale with.

1-week action plan (crystal clear)

1. Day 1: Publish your one-page Privacy Requirements to staff; lock SSO-only for pilots.
2. Day 2: Send the vendor questionnaire + DPA; 7-day deadline; auto-pause non-responders.
3. Day 3: Build your pilot incident log and teacher supervision checklist.
4. Day 4: Select one low-risk tool; configure zero-retention; run a dry run with anonymized content.
5. Day 5–6: Start the 2-week pilot; brief parents with a 2–3 sentence summary.
6. Day 7: Review any incidents; escalate contract gaps; pause if clauses aren’t met.

KPIs for the week: 100% of pilot tools on SSO-only; zero-training clause signed or tool paused; retention set ≤90 days; incident log created and used within 24 hours of any issue.

Your move.

Good call — starting small and keeping humans in the loop is exactly how you avoid noise and get measurable wins fast.

The problem

Support chats hide buying signals but handoffs are inconsistent: sales gets partial context, leads cool off, and opportunities are lost. AI fixes the reading, not the relationships — but you must run it the right way.

Why this matters

Cleaner handoffs speed follow-up, increase conversion, and let SDRs focus on the right prospects instead of hunting for context. That improves rep efficiency and revenue predictability.

Lesson from the field

I’ve seen pilots win when teams keep labels small, force a 1–2 sentence template for handoffs, and require a human confirmation on every high-score lead during the first month. That balance yields quick trust and measurable impact.

Step-by-step plan (what you’ll need, how to do it)

1. Gather & anonymize: Export 3–6 months of chats. Remove PII (names, emails, phones). Store offline for labeling.
2. Label a starter set: Tag 200–300 chats with 5 core fields: intent, product, budget, timeline, handoff_quality. Use a 1-page guide so labels are consistent.
3. Configure AI: Use a plug-and-play classifier or LLM. Feed labeled examples and a validation set. Produce: structured fields + handoff_score (0–100) + 1–2 sentence summary.
4. Define routing rules: e.g., score >75 → immediate SDR alert + summary; 50–75 → sales queue for review; <50 → support follow-up. Keep a human confirmation required for >75 for pilot.
5. Pilot & measure: Run 4 weeks on one product/region. Sales reviews every routed handoff and records outcome (contacted, converted, false positive).
6. Iterate: Re-weight score components, relabel another 200 if F1 score is low, refine summary template from sales feedback.

Metrics to track (and targets to set)

• Time-to-contact (baseline → target: reduce by at least 50% during pilot)
• Handoff-to-conversion rate (track % of routed leads that enter pipeline)
• False positive rate (sales-marked bad handoffs)
• Sales confirmation rate for routed leads (human verification)

Common mistakes & fixes

• Mistake: Too many labels — Fix: collapse to 5 core tags and add later.
• Mistake: Sending raw chats to AI — Fix: enforce PII redaction and test on synthetic transcripts first.
• Mistake: No human review on high-value handoffs — Fix: require quick sales confirmation for scores >75 during pilot.

Copy-paste AI prompt (use with your LLM)

“Read this anonymized chat transcript. Extract customer_intent, product_interest, budget_estimate, timeline, blockers. Assign handoff_score 0-100 and produce a 1-2 sentence handoff_summary sales can use. Recommend next_action (call/demo/email). Output JSON only.”

7-day action plan

1. Export and anonymize 500 chats; store offline for labeling.
2. Label 200 chats with the 5 core fields and create a 1-page labeling guide.
3. Run the provided prompt on 300 holdouts, set an initial score threshold (e.g., 75), and prepare for a 4-week pilot with human confirmation on routed leads.

Your move.

Good point: focusing on simple, effective upsell and cross‑sell offers beats complex campaigns every time when you want fast, measurable revenue.

Why this matters: small, targeted offers improve average order value (AOV), lifetime value (LTV) and margin without needing a major product overhaul. The problem most businesses face is not a lack of ideas but poor targeting, messy execution and no testing plan.

Experience in one line: use data to segment, AI to generate tightly relevant offers, then test — repeat. That sequence produces measurable uplifts quickly.

1. What you’ll need
   1. Customer list with purchase history & basic attributes (product bought, date, value).
   2. Simple CRM or spreadsheet and an email/checkout tool that supports A/B tests or offer blocks.
   3. Access to a generative AI (chat) or a prompt tool.
2. Step‑by‑step (how to do it)
   1. Segment: pick 2–4 high‑value segments (recent buyers, high AOV, repeat buyers, cart abandoners).
   2. Use AI to create 3 focused offers per segment (upsell, cross‑sell, bundle). Use short, clear benefits and a single CTA.
   3. Design two price variants: perceived discount vs. value add (e.g., +20% product vs. free trial/bonus).
   4. Deploy A/B tests in email, checkout or post‑purchase flow. Run each test on a statistically useful sample (≥500 recipients if possible).
   5. Measure, learn, iterate weekly and scale winning offers.

AI prompt (copy‑paste):

“You are a marketing strategist. For customer segment: {segment description}, who recently purchased {primary product} at {price range}, generate 5 upsell or cross‑sell offers that are simple, productized, and low friction. For each offer include: 1) 10‑word offer headline, 2) 20‑word benefit statement, 3) suggested price or discount, 4) target delivery channel (email/checkout/post‑purchase), and 5) expected customer objection and one line to overcome it.”

Metrics to track

• Attach rate (%) — % of orders with the add‑on
• Conversion rate on offer
• Incremental revenue per user (IRPU)
• Average order value (AOV)
• ROI on promotion spend

Common mistakes & fixes

• Too many choices → limit to one strong offer.
• Irrelevant offer → refine segment or offer using purchase context.
• No test plan → always A/B and holdout groups.

One‑week action plan

1. Day 1: Export customer purchase data and select 2 segments.
2. Day 2: Use the AI prompt to produce 3 offers per segment.
3. Day 3: Create email/checkouts for A and B variants.
4. Day 4–6: Run test, monitor daily metrics.
5. Day 7: Analyze results, keep winner, plan scale.

Your move.

Short version: You’ve got the right framework. Turn it into an operational checklist, run a short pilot, and stop any tool that won’t sign a DPA or refuses to exclude student data from model training.

The problem

Many edtech vendors claim “education use” without contractual limits. That exposes districts to COPPA/FERPA risk and creates avoidable data leakage.

Why it matters

Liability, parent backlash, and potential data breaches are real. Plus, once student data gets used to train models it’s essentially permanent. Control the contract and the flow of data first—features come second.

My quick rule from experience

If a vendor won’t: (a) sign a DPA, (b) confirm they do not use student data to train models, and (c) set a short retention window or delete on request — treat the tool as high-risk and pause.

What you’ll need

• A one-page vendor questionnaire (see suggested items below).
• Standard DPA template with explicit clauses for training data, retention, deletion, and breach notification.
• Teacher pilot protocol and incident log template.
• Decision owner (principal/IT director) who can sign off.

Step-by-step (operational)

1. Inventory: 7 days to list all AI tools and owners.
2. Questionnaire: Send to vendors; require answers in 7 days. Key questions: Do you collect student PII? Do you use student data to train models? Can data be auto-deleted after X days? Will you sign our DPA?
3. Risk score: For each tool, assign Low/Medium/High based on answers (criteria below).
4. Pilot: For Low tools, run a 2–4 week pilot with one teacher, logged incidents, and admin review.
5. Contract: Only deploy broadly after a signed DPA and acceptable pilot findings.

Risk scoring quick criteria

• High: Collects PII + uses data for training or indefinite retention.
• Medium: Collects limited PII, short retention, unclear training policy.
• Low: No PII, or school-managed accounts only, vendor confirms no training use, auto-delete available.

Metrics to track

• % tools with signed DPA (target 100% within 90 days).
• Number of pilots completed and incidents logged.
• Time-to-vendor-response for questionnaires (target <7 days).
• Number of tools paused for non-compliance.

Common mistakes & fixes

• Mistake: Accepting vague privacy language. Fix: Require explicit answers and contract clauses.
• Mistake: Skipping pilots. Fix: Pilot with supervision and incident logging.
• Mistake: No owner. Fix: Assign a single decision owner and a 72-hour SLA for escalations.

Copy-paste AI prompts (use these now)

Vendor analysis (use this with a general LLM):

“You are a K–12 privacy reviewer. Analyze the following vendor response for COPPA and FERPA risk. Identify specific data elements that are problematic, rate risk as Low/Medium/High, list required mitigations for safe use in a public school, and provide suggested contract clauses for a Data Processing Agreement including explicit language on: no model training on student data, retention limits, deletion on request, parental controls, and breach notification.”

Contract outreach (short email prompt for vendor):

“We are evaluating [Tool]. Please confirm in writing: 1) whether you collect student PII; 2) if student data is used to train models; 3) your standard retention period and deletion process; and 4) willingness to sign our DPA that forbids model-training on student data.”

1-week action plan

1. Day 1–2: Complete tool inventory and assign owners.
2. Day 3: Send the vendor questionnaire and contract request to all vendors.
3. Day 4–7: Pause any tools with no response or high-risk answers; schedule pilot for one low-risk tool.

Do this and you reduce legal risk while enabling useful AI in the classroom.

Your move.