4.1 Business Requirements for Grounding Solutions

Why grounding is a go-live decision

A grounding solution retrieves content from sources the organization trusts and requires the generated answer to rely on that content. For an AI Transformation Leader, grounding is not a developer flourish. It is the control that decides whether a fluent paragraph is usable in operations, finance, people decisions, or a customer commitment.

A closed large language model (LLM) is trained on public data that existed at training time. That training set does not include this week's pricing sheet, an unpublished policy, or yesterday's incident report. If a user asks those questions anyway, an ungrounded model can still produce a confident answer. Microsoft's retrieval-augmented generation documentation names the retrieved passages grounding data: content you provide at generation time so the model is less likely to guess. The AB-731 skill is to identify the business requirements that make that pattern safe enough to ship.

Six requirements leaders must be able to name

Treat the following as go-live criteria, not as a backlog of nice-to-haves.

Trusted sources are a governance choice. A help-desk agent might be allowed to use the published knowledge base and the current product handbook, and nothing else. A finance assistant might be limited to the signed close package and the controller's desktop procedures. Leaders name the source systems, the owners, and the exclusion list: payroll sites, legal-hold libraries, merger rooms, and personal OneDrive dumps that were never meant to be organizational truth.

Permissions are not the same as "the file lives in our tenant." Microsoft 365 Copilot work grounding uses Microsoft Graph as the permissioned map of email, files, chats, calendar, meetings, and contacts. Graph honors the same identity-based access boundary as the rest of Microsoft 365. If the user cannot view a SharePoint site, a Teams channel, or a mailbox, Copilot's grounding process must not retrieve that content either. Indexing does not mint new rights. The Semantic Index, which Microsoft enables for paid Copilot licenses, still respects role-based access and whether a SharePoint site remains searchable. Graph API shapes belong in a later chapter. The leader takeaway here is that work grounding is permission-trimmed retrieval, not a tenant-wide dump into every licensed seat.

Citations turn grounding into something a human can challenge. Microsoft Foundry's RAG pattern expects responses that can point back to source titles, URLs, or file names stored on the index. Without citations, a grounded system still looks like a black box in an audit.

Freshness is a service-level choice. Semantic Index updates for Copilot are frequent for many mailbox and SharePoint file types, but a custom Azure AI Search index is only as current as your ingestion job. Leaders ask whether a daily rebuild is enough, or whether this process needs a near-real-time push after each policy publish. Stale grounding is still grounding. It is just grounding on the wrong world.

Confidentiality includes Microsoft Purview sensitivity labels and Information Rights Management. Copilot honors usage rights granted to the user. If a file is encrypted and the user lacks rights, it must not appear as grounding. Custom RAG copies that ignore labels create a second, weaker permission system sitting beside the real one.

Human review for high stakes is the last brake. High stakes include compensation, hiring, credit, clinical or safety instructions, regulatory filings, and customer-facing legal language. Microsoft's Copilot extensibility guidance is explicit that agents connected to untrusted sources or sensitive operations should not complete those operations without careful human intervention. A leader who cannot name the human-in-the-loop step has not finished the grounding design.

Grounding versus ungrounded or web-only answers

Ungrounded generation uses only the model's parameters, plus whatever the prompt happens to contain. Web-only answers add public search. When web search is enabled, Microsoft Copilot can derive search terms from the user's prompt and send a query to the Bing Search service. That path can improve general knowledge. It does not know your unpublished policy, and it does not inherit your SharePoint permissions. Mixing web snippets with work data without a written rule is how confidential facts leak into a public-looking answer, or how a public blog is treated as company policy.

Work grounding is the opposite pattern: retrieve from Graph, or from an approved enterprise index, generate with that context, and show citations. Web search, if the organization allows it, is an explicit extra. It is not a substitute for work grounding.

Microsoft also states that prompts, responses, and data accessed through Graph are not used to train the foundation LLMs used by Microsoft Copilot. That fact matters in an executive conversation, but it does not replace permissions, citations, or human review. "Microsoft will not train on our prompts" is not the same statement as "this answer is grounded in the current policy the user is allowed to see."

When a leader must require grounding before go-live

Require grounding before production when any of the following is true:

• The answer will be treated as operational truth: policy, price, entitlement, configuration, or a script a frontline worker will follow.
• A fabrication would create legal, safety, financial, or reputational harm.
• Reviewers, auditors, or regulators must see sources.
• The knowledge is confidential, labeled, or permissioned across groups.
• The facts change faster than the model's training cutoff.

A consumer-style chat that brainstorms marketing slogans from public brand voice might ship without enterprise grounding. A controller's close assistant, an HR policy bot, or a customer-entitlement agent must not.

In practice: An operations vice president wants a floor bot that tells associates the current refund rule. Legal will not accept "the model is very accurate." The VP's go-live list is the official policy library as the only source, citations on every answer, SharePoint permissions preserved, index refresh after each policy publish, confidential draft folders excluded, and a supervisor review queue when the bot recommends an exception. That list is the grounding requirement. It is not a technical remainder to be scheduled after launch.