Hockey + Technology / Field note
A Practical NHL AI Roadmap: Start with Bounded Workflows and Human Signoff
An NHL organization considering AI does not need to begin with an autonomous decision-maker. It needs a bounded
workflow, a defensible information boundary, and a way to judge whether the tool helps after human review. The
roadmap below is a proposed consulting method, not a description of league AI policy or an account of a client
implementation. It assumes no access to club systems or restricted league data.
Map the work before selecting the model
List repeated tasks, their owners, the information they require, and their existing approval steps. Look for
friction that can be observed: finding the right document, reconciling a definition, or preparing a review packet.
Record the present process before changing it. A fluent answer is not a useful outcome if staff spend longer
checking it than completing the task themselves.
Start with lower-risk work using approved public material or synthetic examples: finding a cited vendor
definition, drafting an internal training outline, or explaining a documented data field. Even these uses require
source checks and rights review. Sensitive workflows include confidential strategy, proprietary scouting notes,
credentials, contracts, and restricted operational records. They require separate authorization and controls;
calling a product enterprise-grade does not make every input appropriate.
Treat access as a requirement, not an assumption
The March 31, 2026 NHL report on the SAP-NHL Front Office salary cap projector describes users approved by the
League Central Registry department. That is an example of a purpose-built tool with an explicit access boundary,
not an invitation to connect a third-party assistant. Public reporting about a system does not grant an
integration right, and a cap workflow is not evidence of a general-purpose AI policy.
Draw the proposed data path from source through retrieval, model processing, logs, and exports. Identify the
owner and permitted audience at every step. Use least-privilege access, approved storage, retention limits, and
tested deletion. Ensure retrieval applies the same permissions as the source; an index or cached answer must not
expose a document to someone who could not open it directly.
Resolve vendor rights before uploading anything sensitive
Ask whether the organization has rights to process the source material with this provider, including video,
excerpts, derived annotations, and embeddings. Then check contractual terms for model training, service
improvement, human access, subprocessors, hosting locations, retention, deletion, and incident notification.
Clarify ownership and export rights for prompts, outputs, evaluations, and configuration. A no-training setting
is not a substitute for a binding agreement or a complete data-flow review.
Assign legal or procurement review and a security owner rather than leaving these judgments to an enthusiastic
pilot user. Keep credentials out of prompts. Treat retrieved documents as evidence, never as instructions that
can expand permissions or cause external actions. Where a right or control remains unclear, restrict the pilot
to approved public or synthetic inputs until the question is resolved.
Evaluate the complete task, including refusal
Build a fixed evaluation set with expected evidence, acceptable uncertainty, and examples where the correct
response is to decline or request clarification. Include outdated sources, conflicting definitions, absent
answers, misleading instructions inside documents, and attempted access beyond the test account. Keep a separate
set of unseen tasks for the final check, and rerun it when the model or retrieval configuration changes.
Measure unsupported assertions, citation correctness, permission failures, appropriate abstention, completion
time, and reviewer effort against the existing workflow. Have the task owner approve the criteria before testing.
Any unauthorized disclosure or unapproved external action should stop the pilot. Do not hide those failures
inside an average quality score. Preserve enough versioned evidence to investigate a failure without retaining
unnecessary confidential content in logs.
Illustrative pilot: a source-grounded reference assistant
Consider a proposed six-week pilot that helps authorized staff find definitions in approved reference documents.
This schedule is illustrative, not a client case or a claim of achieved savings. The opening phase would inventory
rights, choose a small corpus, capture a baseline, and appoint a workflow owner. The next phase would build
read-only retrieval, require source links, and test unsupported questions and permission boundaries.
A small staff group would then use the assistant alongside the existing process, not instead of it. A named
reviewer would approve material before it enters a coaching or operational packet. Participants would record
rejected answers and confusing interactions. The final review would compare total effort and failure patterns,
then choose expansion, revision, or retirement. There would be no automated roster decision, contract action,
or publication.
Make adoption and exit part of the roadmap
The NHL announcement of its Verizon-powered Innovation Lab describes a simulated arena setting for testing
technology. The relevant lesson is the value of controlled evaluation; it is not proof that this proposed
assistant is approved or effective. A club pilot should likewise have an explicit test boundary and rollback.
Compare building and buying on support, integration, security review, and exit costs, not demo fluency. Name an
operational owner, provide role-specific training, and preserve the manual path. Expansion should depend on
acceptable evidence and sustained staff use. A roadmap that can responsibly stop a weak pilot is more useful
than one that treats every experiment as a deployment commitment.
Public sources and scope
Discuss a practical hockey AI roadmap with clear information boundaries,
evaluation gates, and human accountability.