Dealer AI Pilot Desk

A working model of the AI Strategy Intern job at a national auto dealer network: an intake board that ranks AI use cases, stage gates that decide when an idea moves up, a tracker for a dealer pilot of AI lead replies, an impact readout, and the playbook, training and status report that go with them.

Built by: Abdullah Al Mneni, Business Technology Management, Toronto Metropolitan University Written for: Ford of Canada's AI Strategy Intern posting (job 73202, Oakville, Summer 2027) Live page: https://abdullahmanini.github.io/dealer-ai-pilot-desk/

Independent concept, synthetic data. Not affiliated with or endorsed by Ford Motor Company of Canada, Limited. Every dealer, lead, audit and survey answer comes from the seeded model in python/model.py. Only two public facts are used: the dealer count by province (ScrapeHero, Sep 22, 2026: 473) and the stage names Sandbox, Functional Program and Enterprise Bet, which appear in Ford's public AI job postings. The model simulates every pilot dealer with and without the tool, so the true effect is known and the readout can be checked against it.

Headline results (synthetic)

• Portfolio: 22 use cases (12 internal, 10 for dealers) scored on value, feasibility, data readiness and risk, with the checks each must clear: privacy and Quebec Law 25, CASL, AI disclosure, French review, human decision, legal and AI council, baseline.
• Adoption: 473 dealers invited, 160 volunteered, a lottery put 80 in wave 1, 65 switched the tool on, 51 sustained it. French-speaking dealers waited a median 7 weeks (French templates arrived in week 19). Only 7 of 15 dealers on CRM-C activated.
• Impact: wave 1 vs the waitlist, difference-in-differences: +15.33 pp leads answered within 15 minutes (95% CI +12.92 to +17.74; true +13.38) and +2.12 pp appointments (CI +0.85 to +3.39; true +1.35). Comparing pilot dealers with everyone else gives +17.82 and +3.00, because volunteers were already ahead before go-live. A placebo go-live at week 7 finds -0.12 pp.
• Guardrails: audited reply errors 2.29% once dealers are past their first two weeks (target 2%); French replies 12.5% before the week-22 fix, 2.6% after. Opt-outs +0.05 per 100 leads with a CI of -0.19 to +0.30, so the pilot cannot rule out a rise above the +0.10 margin.
• Decision: start wave 2 on Nov 2 with three conditions (CRM-C waits for integration, French templates signed off by a qualified translator, errors under 2% for two weeks running).

Repository

Path	What it is
python/model.py	Model parameters and the seeded simulation (same random numbers with and without the tool)
python/generate_data.py	Writes data/raw/*.csv and outputs/model_truth.json
data/use_cases.csv	The 22-idea portfolio and my 1-5 scores
python/pipeline.py	Loads CSVs into SQLite, builds staging and marts, runs the data tests
sql/staging.sql, sql/marts.sql	Dedupe, typed pilot table, dealer-period, weekly and adoption marts
sql/tests.sql	15 data tests with error or warn severity
python/analyze.py	Scoring, funnel, segments, difference-in-differences, placebo, guardrails, survey
python/build_site_data.py	Bundles results into assets/site-data.js
index.html, assets/	The one-page desk (hand-drawn SVG charts, no libraries)
docs/	Assumptions, dealer playbook, AI literacy session, intake template, leadership brief, generated results
DESIGN.md	Design tokens and type for the page

Reproduce

Python 3.9+ standard library only.

python3 python/generate_data.py
python3 python/pipeline.py
python3 python/analyze.py
python3 python/build_site_data.py
python3 -m http.server 8777

Then open http://localhost:8777.

Limitations

Simulated dealers, effects, error rates and survey answers; normal approximations on dealer-level means; the use-case scores and stage criteria are my judgement from public information. The privacy and language checks are questions for legal and privacy teams, not legal advice. Built with Claude Code for the coding. The scope, stage gates, scoring rules, planted problems and decision are mine, and every figure was checked against the query output.