Questions reviewers ask
Answers your security, privacy and procurement colleagues will read
Ten questions that decide whether AI agents go ahead in a Canadian institution. Forward the page to a reviewer as it stands.
The ten questions
Which workflows suit an AI agent?
Workflows where a person already reviews the output before it counts. Examples are retrieval and synthesis across approved sources, a briefing an analyst signs, document triage, and internal policy questions that a service-desk officer then answers. They share a stable set of sources, a named decision owner, a measurable baseline and a review step that already exists. If removing the human is the only way the business case works, start elsewhere.
Can the agents run on Canadian-controlled infrastructure?
Yes. The right setup depends on classification and risk. Options run from a Canadian region of a global provider, through a Canadian-operated sovereign environment, to a model hosted in your own tenancy or on premises. Residency is one of eight control decisions; jurisdiction, administrator access, key custody, telemetry and exit matter as much. We document the options and the shared-responsibility split, and your security authority chooses.
Does a knowledge assistant trigger an Algorithmic Impact Assessment?
Often not, and your institution makes the call. The Directive on Automated Decision-Making applies to production systems that make or support administrative decisions about clients; test and research systems are excluded. A retrieval assistant that drafts for an official who reviews and decides independently frequently falls outside it, though not always. We run the applicability screen with your policy and privacy authorities, record the reasoning and help complete the assessment where it applies. Impact Levels II to IV require peer review.
How do you test French quality and accessibility?
We build bilingual cases into the test set, and a qualified French reviewer scores the French answers for accuracy, terminology and register against your published terminology. Your institution nominates that reviewer by default; if you cannot, we engage one and price it as a variable in the statement of work. Interfaces are built and tested to WCAG 2.2 Level AA. Alexander Leslie, a native bilingual speaker of English and French, leads French engagements.
Who approves what the agent produces?
A named person in your institution, always. The agent retrieves, drafts, classifies or summarizes; a person with the delegated authority reviews and decides. We write the approval authority, the escalation paths and the cases that must never be auto-completed into the charter. We log approvals so a reviewer can reconstruct who saw what and when. If you cannot staff the approval step, narrow the scope; do not weaken the control.
Do answers cite their sources?
Answers come only from an approved set of sources, and each one carries the passages it came from, identified by document, section and version. Retrieval respects identity: a user sees citations only for material they may already read. Where approved sources disagree, the assistant shows the conflict to a person and does not pick one. Citation accuracy is a scored test. You approve the threshold at acceptance.
What happens when the model is unsure?
It declines and hands over. We set confidence thresholds and refusal behaviour during Setup. Below the threshold, outside the approved sources or inside an excluded category, the work goes to a person with the partial result and the reason attached. A rising escalation rate tells you something about the sources or the scope.
Does White Glove Labs hold a security clearance?
In part. We hold Designated Organization Screening at Reliability Status, granted by the Canadian Industrial Security Directorate on 2 September 2026, with a Company Security Officer and an alternate appointed and screened. That covers screened people working on Protected information. Two of our three founders also hold personal Top Secret clearances. We have no Document Safeguarding Capability. Protected material is therefore not stored or processed at our premises; screened staff work on it inside your environment, on your systems. A partner prime carries work that needs a cleared facility or involves classified information. Our Source List qualification is a procurement qualification and not a security screening. Current status, with the organization code and dates, is in the supplier evidence pack.
Which procurement routes are open?
It depends on the value and the vehicles your institution holds. We are qualified in Band 3 on PSPC’s AI Source List (EN578-180001/B), confirmed on 24 August 2026, so a department can solicit us under the source list's terms. Qualification is an invitation to compete, not an award. A subcontract under a prime that holds a federal supply arrangement is also open, with its ceiling set by the prime's vehicle. Where your institution holds a delegated low-dollar authority and an eligible method of supply, a single workflow can move under it. We confirm our standing against any route in writing for your requirement.
What do we get from Setup?
Working skills in a repository your institution owns, each tested against an acceptance test you sign. Your staff get 2 sessions on real work and training in maintaining the skills, so a changed skill does not wait on us. If a skill fails its test, we fix it free. The conditions are on the Consulting offer.
For the security reviewer and for whoever signs off on what an agent may do unattended, the Insights hub lists the pieces we publish on evaluating and deploying agents.