Working with vendors and partners

Almost nothing in Cascade Ridge's AI portfolio would be built entirely in-house. The models behind the call-note summaries and the client-support assistant would come from foundation model providers through Amazon Bedrock. The lab-report extraction would use a managed AWS document service. The assistant itself would be built by a partner firm. The practice management vendor was adding AI features to VetLedger whether Cascade Ridge asked for them or not. "We're going to be governed by our contracts as much as by our policies," Hannah said at the start of the seventh session. This session was about leading those relationships.

Why vendors matter to the transformation

The strategy course treated build, buy and partner as a sourcing decision for each initiative. This lesson looks at what happens after the decision: due diligence, contracts, shared responsibility, ongoing management and exit. The exam guide's governance tasks ask you to identify and address regulatory compliance risks in business processes that use AI and to choose appropriate access controls and data security for AI systems. When a third party provides part of the system, much of that work is done through the vendor relationship. An organization can outsource the work, but it cannot outsource accountability: if the assistant gives a client a wrong answer, the client blames Cascade Ridge, not the partner.

Kinds of AI suppliers

Built-in AI deserves particular attention, because it can arrive without any decision. When VetLedger announced an AI feature that would suggest diagnosis codes from visit notes, Hannah found out from a vendor newsletter. Under Cascade Ridge's new policy, any AI feature in an existing product counted as a new AI use: it needed an owner, a risk tier and approval before it was switched on. Andre asked the vendor for documentation on what data the feature used, where it was processed and how its suggestions could be reviewed, and the feature stayed off until the steering group had seen it.

Due diligence

Due diligence is the work of checking a supplier before relying on it. For AI suppliers it covers the usual commercial and security questions and several AI-specific ones. Andre and Hannah built a short questionnaire that every AI supplier would answer, so responses could be compared side by side.

Some of this evidence is easier to get from large providers than from small ones. AWS, for example, publishes AI Service Cards describing the intended uses, limitations and design choices of some of its AI services, and makes compliance reports available through AWS Artifact. Amazon Bedrock's terms state that customer inputs and outputs are not used to train the underlying foundation models and are not shared with model providers. A small implementation partner may have none of this documentation, which does not rule it out, but means more of the assurance has to come from the contract and from testing.

The shared responsibility model

Cloud providers describe security as a shared responsibility: the provider is responsible for the security of the cloud, such as the data centres, hardware and core services, and the customer is responsible for security in the cloud, such as who has access, how data is classified and how applications are configured. The same idea extends to AI. The model provider is responsible for how the model was built and the protections it offers. Cascade Ridge is responsible for which model it chooses, what data it sends, how it configures guardrails and access, how outputs are reviewed and how the system is monitored.

Getting this division wrong in either direction causes problems. Assuming the provider handles everything leads to systems with no guardrails, wide access and no monitoring. Assuming the provider handles nothing leads to duplicated effort and missed protections that are already available. Each AI contract should make the division explicit.

Contracts for AI

A good AI contract does more than set a price. Noor and Hannah listed the terms they wanted in every significant AI agreement, and Lorraine added the commercial ones.

• Data use and protection: no use of Cascade Ridge data to train the supplier's general models without explicit agreement; processing in named Regions; deletion on request and at contract end; breach notification within a set time.
• Service levels: availability targets, support response times and service credits if targets are missed.
• Model change notice: advance notice of model updates or retirement, with time to retest before a change takes effect where possible.
• Performance and acceptance: agreed evaluation tests that must pass before each stage is accepted and paid for.
• Ownership and knowledge transfer: Cascade Ridge owns prompts, configurations, evaluation sets and integration code built for it, and the partner documents and trains Kwame as it goes.
• Responsible AI and compliance: the supplier follows Cascade Ridge's AI policy for this work and helps with privacy reviews, audits and incident investigations.
• Exit: data and configurations exportable in usable formats; reasonable help with transition to another supplier.

Reading a service level

Service levels are often written as an availability figure, such as 99.9 or 99.5, meaning that the service should be available for that much of the time out of every 100. A service at 99.9 should be unavailable for no more than 0.1 out of every 100 minutes. A small difference in the figure can mean a large difference in downtime, and the strategist should translate it into terms the business understands.

Managing the relationship

Signing the contract starts the relationship rather than ending it. Each significant AI supplier would have a named relationship owner at Cascade Ridge, who would meet the supplier each quarter to review performance against service levels and acceptance tests, incidents, model changes announced or planned, costs against budget, and progress on knowledge transfer. For the partner building the assistant, Kwame would join every working session for the first stage, and the partner would hand over documentation at each stage gate before the next payment was released.

Lock-in was the risk Lorraine raised most often. Some dependence on a supplier is the price of using its strengths, and avoiding all lock-in would mean building everything in-house, which Cascade Ridge could not afford. The aim was to keep the cost of switching reasonable. Portable evaluation sets, prompts and guardrail rules kept in Cascade Ridge's own systems, standard integration patterns, and an exit clause did most of the work. So did an annual check of at least one alternative for each major supplier, so that a switch, if needed, would not start from nothing.

Compliance through the supply chain

Regulatory duties do not stop at the organization's boundary. Under PIPEDA and the Alberta and British Columbia private-sector privacy laws, an organization remains responsible for personal information it transfers to a third party for processing and must use contracts or other means to provide comparable protection. Some provinces also have rules on public bodies storing data outside Canada, though those do not apply to a private veterinary group. For each supplier, Hannah recorded which obligations applied and which contract terms met them.

The interim AI guideline also had a supply chain dimension. Staff using a free public chat tool were, in effect, using an AI supplier with no contract at all. That was the strongest argument for providing an approved generative AI tool for everyday drafting, governed by Cascade Ridge's contract and settings, rather than simply banning the public ones. People who are given a safe tool use it; people who are only told no tend to find another unsafe one.

Working with AWS Partners and programs

Organizations building on AWS can draw on more than a single partner firm. AWS Partners with relevant specializations and competencies have been reviewed by AWS for experience in particular kinds of work, which gives a first filter when choosing an implementation partner. AWS Marketplace lists third-party models, data products and AI applications that can be bought through an existing AWS account, so that their costs appear on the same bill and some procurement steps are simplified. AWS also offers programs such as the AWS Generative AI Innovation Center, which works with customers on generative AI use cases. None of these removes the need for due diligence, but each can shorten it.

Andre used the partner specialization list to draw up the shortlist for the client-support assistant, and he asked each shortlisted firm to answer the same questionnaire. The firm chosen in the strategy course had relevant experience and references, but it was small, so Hannah asked for more evidence on security practices and wrote the knowledge-transfer obligations in detail. The choice of a reviewed partner did not change who was accountable; it changed how much evidence was already available.

Vendor scenarios on the exam

Exam questions on vendors often describe a supplier that is cheaper or faster but weak on one governance point, such as unclear data use terms, no change notice, or no way to export configurations. The best answer usually addresses that point through due diligence, contract terms or design, rather than either accepting the risk silently or rejecting the supplier outright. Questions about built-in AI features in existing software usually expect them to be treated as new AI uses with their own review. And questions about accountability expect the organization to remain accountable to its customers whoever supplies the technology.