A concept for Ford of Canada's AI Strategy team · Abdullah Al Mneni · October 2026

Prove it at 80 dealers. Then take it to 473.

This page works through the AI Strategy Intern job on a model. It has an intake board that ranks AI ideas, stage gates that decide when an idea moves up, a tracker for a dealer pilot of AI lead replies, and the readout I would bring to the AI Strategy Manager. The dealer network and every number on the page are synthetic.

AI use cases scored: 22
Dealers modelled: 473
Leads answered in 15 min: +15.3
Wave 2: Go

Independent student project. Not affiliated with or endorsed by Ford Motor Company of Canada. No Ford, dealer or customer data was used.

How ideas move

Every idea takes the same road

Ford's public AI job postings describe the central AI team's pipeline as three stages: Sandbox, Functional Program and Enterprise Bet. I used those names so a Canadian intake board would feed the global one without a translation step, and I added an intake stage in front. Each gate asks one question.

Stage names come from Ford's AI Strategy job postings on careers.ford.com. The gate questions and criteria are my proposal.

Use-case portfolio

22 ideas, ranked

My starting backlog. Each idea is scored from 1 to 5 on value, feasibility, data readiness and risk. The scores are my judgement from public information, not Ford's. Move the weights and the ranking changes. Open a row to see the idea and its next step.

Weights

#	Use case	Stage	V F D R	Priority	Checks before it leaves Sandbox

Intake form

Score a new idea

The form an employee or a dealer would fill in. It returns a priority score, where the idea should enter, and the checks it has to clear. The example below is filled in already; change anything.

Dealer pilot · AI first reply to web leads

Who actually switched it on

All 473 dealers were invited in week 10. 160 volunteered, and the vendor could onboard only half of them at once, so a lottery picked wave 1. Wave 1 went live in week 13 (Jul 20). The waitlist starts in week 28 (Nov 2). That lottery is what makes the impact readout fair.

Adoption funnel

Active: the tool answered leads in each of weeks 23 to 26. Sustained: it answered at least half of all leads in each of those weeks.

Where wave 1 stalled

Wave-1 dealers	Dealers	Activated	Median weeks to switch on

Two fixable causes. French reply templates arrived in week 19, six weeks after go-live, so French-speaking dealers waited. Dealers on CRM-C had no integration at launch, and only 7 of 15 switched the tool on.

Dealers by province

Province counts follow a third-party dealer count (ScrapeHero, Sep 22, 2026: 473). Quebec has 89, about one dealer in five, which is why French readiness is a gate and not a nice-to-have.

Impact readout

Did replies get faster?

Yes. With the tool, about 15 more leads in every 100 got a first reply within 15 minutes. The first comparison anyone would run says 18, and it is wrong for a reason that will come up again at Ford: the dealers who volunteer for AI pilots are already the stronger digital dealers.

Share of web leads answered within 15 minutes, by week

Wave 1 (lottery winners)
Waitlist (volunteered, not yet live)
Did not volunteer

Four ways to read the same pilot

Dealer-level means, percentage points, 95% confidence intervals. "True effect" is known only because the data is simulated.

What each comparison gets wrong

Pilot dealers vs everyone else after go-live gives . Before go-live the same dealers were already ahead, because volunteers are the more digital dealers.

Before vs after at pilot dealers gives . It credits the tool with the network's slow improvement over the summer.

Wave 1 vs the waitlist compares volunteers with volunteers, split by lottery, and subtracts each dealer's own starting point. It gives . The true effect in the model is , inside that range.

Check: run the same comparison on weeks 1 to 12 with a fake go-live at week 7 and it finds , which is what no effect looks like.

81% of wave 1 switched the tool on, so the effect at a dealer that uses it is about .

Guardrails and data checks

What has to be true before wave 2

Speed is the reason to run the tool. These are the reasons to stop it: wrong answers to customers, people unsubscribing, and dealers who cannot use it in their language or their CRM.

Audited AI replies with an error, by week

QA reviews about 10% of AI replies each week. Errors: wrong price or stock, wrong language, a promise the dealer can't keep. Small weeks swing a lot; the line is the 2% target.

Data tests

15 SQL tests run on every load. Errors stop the readout; warnings are explained.

Dealer feedback · week 22 pulse survey

What dealers said

60 of 80 wave-1 dealers answered a pulse survey with a 1-to-5 rating and open comments, coded into themes. Dealers who were live rated the tool 3.7 out of 5. The 9 still waiting rated it 2.1; six of them were French-speaking or on CRM-C.

Enablement

What dealers and staff get

The posting asks for playbooks, training content and AI literacy material. These are the short versions; the full drafts are in the project's docs folder.

Dealer playbook

AI first reply: the first 30 days

Before go-live. Connect the CRM, set store hours, name the person who takes hand-offs, and approve the reply templates in English and French.
Week 1. The BDC lead reads every AI reply the same day and flags anything wrong in the review queue.
Weeks 2 to 4. Review ten replies a week. Check four numbers every Monday: leads answered in 15 minutes, appointments set, minutes to a human hand-off, opt-outs.
Day 30. Call with the AI Strategy team. Keep, adjust or switch off.

AI literacy · 45-minute session

Using the AI assistant at work

10 min. What the assistant is good at, and where it makes things up.
15 min. Live: summarise a dealer bulletin, draft a reply, build a slide outline.
10 min. The safe-use card: no customer personal information in unapproved tools, check every number, you own what you send.
10 min. Each person writes down one task they will try this week. We check back in two weeks.

Governance

The intake template

Owner and problem. Who asked, and which number gets better.
Data. What it reads, and whether any of it is personal information or from Quebec customers.
People. Who reviews the output, and how a customer reaches a person.
Measure. The baseline, a comparison group, and what would make us stop.

The checks are questions for Ford's privacy and legal teams, not legal advice. Laws named: PIPEDA, Quebec's Law 25 (privacy impact assessments and notice for fully automated decisions), CASL (consent for commercial electronic messages) and Quebec's French-language rules.

Leadership brief

Status report, week of October 26

Dealer AI first reply

Amber: working, not ready for all dealers

Stage

Functional Program

What happened

80 dealers went live in July. Where the tool is on, about 15 more leads in 100 get a reply within 15 minutes, and appointments are up about 2 points. French-speaking dealers and dealers on CRM-C fell behind for reasons we can fix.

Decision needed by October 30

Start wave 2 on November 2 for the 80 waitlisted dealers, with three conditions:

Dealers on CRM-C wait until the vendor ships the integration.
A qualified translator signs off the French templates before any French-speaking dealer goes live.
Audited reply errors stay under 2% for two weeks in a row. Today: 2.3%.

Next 30 days

Keep the 10% reply audit. Add a per-message opt-out log so the opt-out guardrail can be read properly. Re-run this readout with wave 2's first four weeks, then decide on an Enterprise Bet proposal.

How it was built

Method and limits

Pipeline

A seeded Python model writes five source extracts: dealers, pilot enrollment, weekly leads, reply audits and the survey. SQLite builds staging tables and marts, and 15 SQL data tests run on every load. The analysis is standard-library Python with dealer-level difference-in-differences. Two problems are planted to test the checks: duplicate rows in week 20 and a CRM export outage in week 17.

Limits

Everything is simulated, including dealer behaviour, effects, error rates and survey answers, so the page shows a method and no facts about Ford's dealers. The use-case scores are my judgement. The French findings come from the model's assumptions and no French-speaking dealer was asked. A real French playbook needs a qualified translator.

Who made it

Abdullah Al Mneni, Bachelor of Commerce in Business Technology Management at Toronto Metropolitan University, graduating Spring 2028, based in Oakville. I wrote the code with Claude Code. I chose the scope, the stage gates, the scoring rules, the planted problems and the decision, and checked every number here against the query output.