Case study: a generative AI adoption plan

On November 10, 2026, after the steering committee voted to expand the product description tool, Dana Leclair asked Jules for one more thing before the December board meeting: a single document that shows the board where Birchline's generative AI work is going over the next eighteen months, what it will cost, what it should return, and how the risks are controlled. "Not a vision statement," she said. "A plan I can hold you to." Jules and Farah are writing it this week, drawing on everything in this course. This final lesson walks through that plan section by section, as a worked example of how the pieces fit together, and ends with the practice the Generative AI Leader exam expects: reading a scenario and choosing the strategy that fits.

1. Purpose and principles

The plan opens with a short statement of purpose that the board can repeat: Birchline will use generative AI to give customers faster, more accurate help and product information, and to free staff from repetitive work so they can spend more time on customers and on work that needs judgment. It lists the principles from Birchline's AI policy: benefit to customers and staff, fairness, transparency, privacy, safety and human accountability. And it states the boundaries: no generative AI in hiring, credit or discipline decisions; no payment card data in any AI tool; human approval for anything published or any action with consequences.

2. Where Birchline is today

An honest starting point builds credibility, so the plan summarizes the position in November 2026.

3. The portfolio

The plan presents use cases as a portfolio across three horizons, so the board sees a balance of quick results and longer bets rather than one big project.

Each later item has a decision point rather than a fixed commitment. The shopping assistant launches in spring 2027 only if red-team findings are fixed and the steering committee approves; the email assistant pilot starts only if the shopping assistant's monitoring shows the review and escalation design works with real customers.

4. Value and cost

The plan's value section uses the method from lesson 2: baselines, measurable targets, ranges and named uses for released time. Its cost section lists one-time and yearly costs, including people's time, with assumptions labeled. All figures in Birchline's plan are its own estimates; the ones below are simplified for this lesson.

# Simplified portfolio view (illustrative estimates, in thousands of dollars per year).
portfolio = [
    # name,                   one_time, yearly_cost, yearly_benefit_low, yearly_benefit_high
    ("Product descriptions",   22,        9,          35,                 55),
    ("Review digests",         10,        4,           8,                 15),
    ("Gemini adoption",      6,        0,          40,                 90),
    ("Policy assistant",       12,       10,          15,                 30),
    ("Shopping assistant",     60,       30,          40,                120),
]
one_time = sum(p[1] for p in portfolio)
yearly = sum(p[2] for p in portfolio)
low = sum(p[3] for p in portfolio); high = sum(p[4] for p in portfolio)
print(f"One-time investment: {one_time}k")
print(f"Yearly running cost: {yearly}k")
print(f"Yearly benefit range: {low}k to {high}k")
print(f"Net yearly value range: {low - yearly}k to {high - yearly}k")
print(f"Simple payback range: {one_time / (high - yearly):.1f} to {one_time / (low - yearly):.1f} years")

One-time investment: 110k
Yearly running cost: 53k
Yearly benefit range: 138k to 310k
Net yearly value range: 85k to 257k
Simple payback range: 0.4 to 1.3 years

The Gemini app is already in the IT budget, so the plan counts only the adoption effort and the January Workspace features pilot as new costs, and notes that the wide benefit range reflects how uncertain it is to value general productivity. The shopping assistant has the widest range of all, which is why it has a decision point and a staged launch. Dana's comment on reading the draft is that she believes the low end and hopes for the middle, and that is exactly how a plan like this should be read.

5. How quality is assured

This section explains, briefly, the techniques from lessons 3 and 4: every factual use case is grounded in approved Birchline sources and cites them where users need to check; prompts are versioned; every published output is approved by a person; high-volume uses combine thresholds, sampling and escalation; and every production system has an evaluation set that runs before any change, with safety-critical measures that may not get worse. The board does not need the detail, but it needs to know these practices exist and who owns them.

6. Data

The data section, from lesson 5, commits to three things: a single approved source for every document used for grounding, with an owner and review date; permission-aware retrieval for every assistant; and removal of personal information from documents and logs where the task does not need it. It names the two biggest data risks for the plan, older supplier spec sheets for the description refresh and scattered order history for the email assistant, and the work scheduled to address them.

7. People

The people section, from lesson 6, sets adoption targets and support: champions in every department, role-based training, protected practice time, a prompt library, and updated job descriptions and goals where workflows change. It repeats the CEO's commitment that any future change affecting jobs will be discussed early and openly, and records that HR will consult the union before any warehouse use. It also sets a skills goal: six staff, including Jules, holding the Generative AI Leader certification by the end of 2027, and two engineers with deeper Google Cloud AI skills.

8. Risk, security and responsible AI

This section brings lessons 7 and 8 together in a single risk register, the format boards are used to.

The register lists the privacy duties that apply, from Quebec's Law 25 for Quebec customers to provincial and federal law elsewhere in Canada, and Birchline's privacy officer signs off each customer-facing use. It also records that the company follows the NIST AI Risk Management Framework as a reference and will consider ISO/IEC 42001 certification in 2028 if partners ask for it.

9. Governance and reporting

The final section describes the structures from lesson 9 and commits to what the board will see: a quarterly report with the AI inventory by risk tier, KPIs for each production system, adoption measures, incidents and their fixes, costs against budget, and decisions due in the next quarter. It ends with the decision points and what evidence the steering committee will need at each one.

Reading scenarios like the exam

The Generative AI Leader exam presents short business scenarios and asks which approach, product type or practice fits best. Farah shows Jules a method that works for most of them, and for real decisions too.

• Find the goal. What outcome does the organization want: speed, accuracy, cost, customer experience, compliance?
• Find the constraint. What rules out options: data that must stay private, a need for current facts, no technical staff, a tight deadline?
• Find the risk. What happens if the output is wrong, and who is affected?
• Choose the simplest option that meets all three. Prefer prompting before tuning, grounding for facts, buying for generic tasks, human review where errors matter, and governance proportionate to risk.

Applied to a scenario such as "a retailer's support assistant gives outdated return policies", the goal is accurate answers, the constraint is that policies change, and the risk is customer harm and complaints, so the best answer is grounding on current, approved policy documents with citations, not fine-tuning the model on today's policies. The method is the same one Jules used for the whole plan, at smaller scale.

What makes a plan credible

Farah reviews the draft with Jules against a short list of signs that a board will trust a plan. Does every number have a source or a labeled assumption? Does every use case have a named owner? Are there points where the plan can be stopped or changed, and is it clear who decides? Does it admit what went wrong so far and what was learned? Does it say what the organization will not do? A plan that passes those tests is believable even when its numbers are uncertain, because the board can see how the uncertainty will be reduced.

Jules will present the plan to the board in December. Farah has suggested two additions the board is likely to want: bringing the shopping assistant decision back to the board rather than leaving it with the steering committee, and a short note each quarter on what Birchline has decided not to do, and why. A board that asks what you chose not to do, she says, understands strategy. Whatever the exam result, Jules will have a plan Dana can hold Jules to.