SignalsOperating intelligence
Open navigation

Operating question

Useful AI work becomes easier to trust when sources stay distinct, usage is visible and a technology partnership can be translated into choices a team can test.

Decision Architecture

3 Things AI: The “Show Me Where That Came From” Edition

3 Things AI 5 min3 sources

For

Leaders and workflow owners

You will leave with

3 operating decisions

Reading mode

5 min · 3 verified sources

Reading guide3 decisions · 4 sections+

Decision points

  1. 01Separate internal document facts, external web facts and judgment before accepting a polished AI-assisted answer.
  2. 02Reconcile product metrics, provider invoices and approved local-model use before drawing conclusions about AI cost or adoption.
  3. 03Ask an AI supplier for a one-page dependency map that turns sovereignty language into locations, providers, access and fallback choices.

Three practical tests for separating document facts from web context, seeing hidden model usage and asking what AI sovereignty changes for a Canadian firm.

The harder question in many AI-assisted meetings is no longer whether the answer sounds good. It is whether anyone can show where it came from, which model handled it and what happens if a supplier changes. Three developments this week make those questions more answerable.

1. Keep your document facts separate from outside context

A consultant reviews a client's expansion plan before a bank meeting. It contains company revenue assumptions, while current exchange rates and market figures live elsewhere. An assistant can combine both quickly. The risk is that a web figure and an approved internal assumption arrive in one polished paragraph, leaving the reviewer to guess which is which.

On November 3, Adobe added optional web search to AI Assistant in Acrobat. Adobe says the assistant begins with the user's documents, labels web results with clickable citations and lets the user turn web search off or restrict a question to the documents. It also says documents stay within Acrobat rather than being sent to the search provider. These are Adobe's product statements, not an independent assessment of accuracy, privacy or coverage.

The benefit is an easier comparison between a file and changing outside information. The tradeoff is that citations can make a weak answer look settled. A link may support one number without supporting the conclusion, and a current source may still be unsuitable for a financial or contractual decision.

Try one practical test with a proposal containing five important claims. Mark each as internal fact, external fact or judgment. Ask the same questions with web search off and on, then open every citation and record whether it supports the exact sentence. Keep both source types visible in the final document. The feature can shorten the search, but it cannot decide which source is authoritative.

2. Find the AI usage that your main bill does not explain

A twelve-person development team sees steady Copilot adoption but a cloud-model invoice rising at a different rate. Two developers use local models, another uses a personal provider key and the organization has tested a custom model. The owner has totals, yet cannot tell which requests were hosted by GitHub and which ran elsewhere.

GitHub's November 3 update added bring-your-own-key and enterprise custom-model fields to the Copilot usage metrics API. Reports can separate locally configured keys and models from custom enterprise models routed through GitHub. GitHub notes important limits: local reporting depends on client telemetry, some providers do not return token counts and earlier report days are not backfilled.

The benefit is a better view of where AI activity and cost originate. The tradeoff is incomplete measurement. A dashboard that undercounts local use can still look precise, and user-level data can turn a cost question into unnecessary employee surveillance if its purpose and access are unclear.

For one week, reconcile the Copilot report, provider invoices and a short team declaration of approved local or personal-key use. Compare only what is needed to explain cost and policy exceptions. If the numbers disagree, label the gap instead of estimating it away. GitHub's fields cover its own surfaces and telemetry; they do not discover every AI tool.

3. Turn sovereignty language into one supplier question

A Canadian manufacturer is choosing an AI assistant for maintenance records. One vendor promises a Canadian region, another a European partner ecosystem and a third global resilience. All three use the word “sovereign,” but the contract still does not say where support staff can access data or how service continues during a cross-border disruption.

On November 4, Innovation, Science and Economic Development Canada reported that Canada and Germany are advancing cooperation on AI sovereignty. The release points to a Sovereign Technology Alliance intended to deepen work on common AI models, shared digital infrastructure, capital and research. It describes policy direction and international cooperation; it is not a procurement guarantee for a Canadian business.

The benefit is more attention to resilient, trusted choices beyond one domestic supplier. The tradeoff is that “sovereignty through partnership” can remain too broad for a buyer. Shared infrastructure may reduce one dependency while adding contractual, technical or legal relationships to understand.

Ask one prospective vendor for a one-page dependency map: data storage location, model provider, identity system, support-access countries, backup region, subcontractors and the manual fallback if one service is unavailable. Put the answer beside the business's actual requirement rather than a generic preference for local hosting. Government cooperation may improve future options, but availability, terms and Canadian suitability still need product-by-product verification.

The bigger pattern

The useful shift is from accepting a reassuring label to asking a question that leaves evidence. A citation distinguishes a contract clause from a web result. A usage field distinguishes hosted activity from an outside model. A dependency map distinguishes a Canadian region from a resilient Canadian choice. Each can be incomplete, but each gives a busy team something specific to check before expanding the workflow.

Which supplier claim would become more useful to your team if it had to be shown in one page of evidence?

Verified sources

Continue your decision path

Move from understanding to action.

02 · Go deeper

3 Things AI: The “Does This Hold Up on Monday?” Edition

Three practical ways to compare AI options on finished work, review a useful sample after delivery, and separate an internal rollout from a content agreement.

Read next
03 · Assess

Apply this signal to your architecture.

Identify the workflow, context, and controls to structure first.

Open Architecture Assessment