Case study: an AI roadmap for a mid-sized company

Andre had promised the board an AI roadmap by the end of the first quarter of 2027. By late November 2026 he had a draft, built from every session of the autumn: the value trees, the prioritized shortlist, the fit screens, the business cases, the sourcing decision, the data readiness assessment, the cost estimates, the pilot gates and the communication plan. This final lesson walks through that draft as a case study, showing how the pieces of AI strategy fit together into a plan a board can approve, fund and hold people to.

Cascade Ridge Veterinary Partners is a typical mid-sized organization for this kind of exercise: about 420 staff, twelve sites, revenue of about 48 million dollars a year, a small IT team, one contract data scientist and a newly built AWS environment. It has enough data and scale for AI to matter, and few enough people that every initiative competes for the same handful of experts.

1. Objectives and principles

The roadmap opens with the board's objectives for 2027, from lesson 1, and the principles that govern every AI initiative. The principles were short and specific:

• Every initiative serves a named business objective and has an owner, KPIs, a baseline and a decision gate.
• Choose the simplest solution that works; AI only where it fits better than rules, existing features or process changes.
• AI may assist, but every clinical decision stays with a veterinarian or a credentialed veterinary technician.
• AI outputs are never used to refuse or delay care or bookings.

Client data stays in Canada, in line with PIPEDA and the Alberta and British Columbia private-sector privacy laws.

• Results are reported honestly, in a consistent structure, including when an initiative is paused or stopped.

2. The portfolio

The prioritization in lesson 2 and the fit screens in lesson 3 produced a portfolio with a place for every idea, including the ones that were not AI projects at all.

Including the decisions not to use AI was deliberate. They showed the board that the roadmap was not a wish list, and they freed budget and attention for the initiatives that mattered. The two "not AI" items were also the cheapest wins on the list.

3. The timeline

The timeline sequenced initiatives so that the same people were not needed for everything at once. Noor's time went first to the no-show pilot, then to the summarizer evaluation. Kwame's time went to integration for the no-show model and to working alongside the partner on the assistant. Data foundations ran throughout, because every later initiative depended on them.

4. The budget

The first-year budget of 450,000 dollars was built up from the business cases and checked against the investment guide in lesson 11, where it came to about 9.4 dollars per 1,000 dollars of revenue.

Lorraine questioned the size of the non-initiative share and then accepted it. "If we don't fund the foundations, every project pays for them separately and badly," she said. She added one condition: each initiative's running costs had to be tagged and tracked against its own line from the first day, using cost allocation tags and AWS Budgets alerts as lesson 8 described.

5. Expected value

Andre presented the value of the portfolio at full run rate, counting only realized value that had an owner and a measure. Each figure was an expected case with a lower conservative case beside it, and the intangible benefits were listed separately.

The total of about 190,700 dollars a year at full scale was well below the first-year budget, and Andre said so plainly. Only two initiatives paid back on counted cash value alone: the no-show reminders, in about 8.5 months, and the lab-report extraction, with about 45,000 dollars of one-time cost and about 20,000 dollars a year of net benefit, in a little over two years. The call-note summaries roughly covered their own running costs in overtime savings, and the client-support assistant cost slightly more to run than the call time it saved. Those two were justified mainly by benefits that were listed but not counted in money: safer, less exhausting overnight shifts and lower technician turnover for the summaries, and parity with rising client expectations for the assistant. The data foundations and governance were justified by the 2028 pipeline they made possible. Lorraine preferred this honest picture to a portfolio that claimed to pay for itself in year one.

6. Governance and decision gates

Each initiative had a decision gate at the end of its pilot, with criteria agreed in advance, as lesson 9 described. Above them sat a portfolio review every quarter by a small AI steering group: Andre, Lorraine, Dr. Kaur, Hannah and Noor. At each review the group looked at every initiative's KPIs, leading indicators, guardrails and costs, and made one of four decisions: scale, continue, pause or stop. Any initiative touching clinical work needed Dr. Kaur's approval; any new use of personal information needed Hannah's.

The governance detail, including risk reviews, model documentation and incident handling, belonged to the responsible AI course, and the people and change side belonged to the transformation course. The roadmap referred to both rather than repeating them, but it made clear that no initiative would move from pilot to scale without passing both the value gate and the governance review.

7. Communication plan

The roadmap ended with a communication plan built on lesson 10: a one-page summary for the board each quarter, a monthly value dashboard for managers built in Amazon Quick from aggregated figures, briefings for staff at each clinic before any pilot reached them, and client-facing notices for anything that changed how clients were contacted. Reception leads and emergency technicians were named as champions, with time set aside for them to collect feedback.

8. Risks to the roadmap itself

Finally, Andre listed the risks to the plan as a whole, not just to individual initiatives:

The recommendation

The board will meet in December to consider the draft. Andre's recommendation was to approve the 2027 budget in tranches: release funding for data foundations, the no-show pilot and the partner build of the client-support assistant at once, and release the rest after the first gate in May 2027, depending on its results. Tranche funding tied money to evidence, which was exactly what the roadmap promised.

"It's not a technology plan," Lorraine said when she read the final draft. "It's a business plan that happens to use technology." Noor considered that the best review the roadmap could get, and Andre added it, unattributed, to his exam notes as a summary of Domain 2.

Lessons for any mid-sized organization

Andre's roadmap was specific to Cascade Ridge, but Noor pointed out that its shape would fit most mid-sized organizations starting their AI programs. Start with the business objectives and a few well-chosen initiatives that can show measured value within a year. Fund data foundations and governance visibly from the start. Prefer buying or partnering for parity uses and building, or co-building, where the organization's own data offers an advantage. Tie money to evidence through decision gates and tranche funding. Communicate honestly and often, including when things go wrong.

She also warned against two opposite failures she had seen elsewhere. One was the organization that launched dozens of small AI experiments with no portfolio view, no measures and no owners, and two years later could not say what any of them had achieved. The other was the organization that waited for a perfect strategy before trying anything, and found that its competitors and its own staff had moved on without it. A roadmap like Andre's avoids both: it is focused enough to measure and flexible enough to change at every quarterly review.