Engagements where language, knowledge and workflow met
These anonymised case stories describe past applied-AI studio work for Canadian and North American clients. Metrics are illustrative results from specific projects—not guarantees for your organization.
A Canadian logistics scale-up — dispatch copilot
Problem: Dispatch coordinators retyped the same exception notes into three systems every morning. A generic ChatGPT tab helped individuals but created inconsistent records.
Approach: Discovery mapped twelve recurring exception types. We built a copilot workflow integrated with their TMS API, grounded on internal SOPs via RAG, with mandatory human confirmation before any customer notification.
Outcome: Pilot teams reported faster note completion in weekly surveys; leadership chose a phased rollout. We do not promise identical results elsewhere.
A national retailer’s support team — triage assistant
Problem: Tier-one agents spent half their shift searching the knowledge base while customers waited. Previous bot attempts deflected tickets without resolving them.
Approach: We redesigned the retrieval layer, added source citations in the agent UI, and built an evaluation suite from six months of anonymised tickets. Escalation paths preserved human judgment on refunds and exceptions.
Outcome: Handle time improved in the pilot queue; the client extended the retainer for governance and prompt maintenance. Adoption required training their internal owners.
A fintech’s risk-ops group — policy Q&A
Problem: Analysts asked the same regulatory interpretation questions in Slack threads that senior staff answered repeatedly.
Approach: RAG over approved policy memos with refusal rules when sources conflicted. Every answer linked to paragraph-level citations; low-confidence responses routed to a compliance reviewer.
Outcome: Repeat questions dropped in internal analytics during the pilot; compliance sign-off remained human. Generative AI did not replace professional judgment.
A professional services firm — proposal drafting copilot
Problem: Partners spent evenings assembling proposal language from old decks scattered across SharePoint.
Approach: Workflow automation pulled approved case blurbs and pricing templates; the copilot drafted sections for partner edit only—no auto-send. PIPEDA-sensitive client names were excluded from training data.
Outcome: Draft turnaround shortened for the pilot group; partners still owned final submissions. Scope stayed narrow to avoid over-promising.
“We show what shipped, where humans stayed accountable, and what we would do differently.”
