Skip to content
FR

Analyze AI-Powered Business Requirements

Step 3 of 22

AB-100: Business Requirements and AI Grounding Data

AI grounding data and business requirements determine whether an agent can deliver useful, trustworthy results. This AB-100 lesson covers the first skills measured under Plan AI-powered business solutions (25–30%): analyzing business requirements for AI and agents, and reviewing the data that will ground them. A good architect decides whether and where an agent adds value, and whether the data is ready, before choosing any product.

Learning objectives

After this lesson you can:

  • Identify automation, analytics and decision-support opportunities that suit an agent.
  • Distinguish prompt-and-response, task and autonomous agent scenarios and match them to requirements.
  • Evaluate grounding data for accuracy, relevance, timeliness, cleanliness and availability.
  • Organize business data so it is ready for AI and reusable by several solutions.
  • Capture functional, non-functional and compliance requirements for an agent.

1. Where agents add value

Microsoft guidance groups the value of agents into three families of work:

  • Task automation – multi-step processes with rules, approvals and system updates (processing an invoice, onboarding a supplier, triaging a case).
  • Data analytics – gathering, summarizing and explaining information across sources (a sales pipeline summary, a variance explanation for finance).
  • Decision-making support – recommending or taking a next step using context and policy (which offer to propose, whether to escalate).

Indicators of a promising use case include work that is repetitive or high-volume, relies on unstructured information (email, documents, chat), spans several systems, follows describable rules or policies, and has a measurable outcome. It is a poor candidate when the rules are undefined, the data is unavailable or untrustworthy, errors are very costly and cannot be reviewed, or a simple rule-based workflow already solves it cheaply.

Exam tip: do not choose an agent when deterministic automation (for example a Power Automate flow) fully meets the requirement. Use agents where language understanding, reasoning or flexible tool selection is needed.

2. Matching requirements to agent scenarios

Requirement signal Likely scenario
Users ask questions about policies or products Prompt-and-response agent grounded on knowledge sources
A user wants a job done (draft a quote, update a record) Task agent that uses tools and actions on the user's behalf
Work must start from an event or schedule with little human involvement Autonomous agent with triggers, knowledge and action tools, plus rules to escalate
Several specialties must cooperate Multi-agent design with an orchestrating agent

The Cloud Adoption Framework describes five core components: a generative AI model, instructions, retrieval, actions, and memory. For requirements analysis, map these to reasoning capability, role and boundaries, grounding data, authorized tools, and conversation or case state. The four adoption areas discussed in the next chapter are a separate framework; they are not the component count.

For autonomous scenarios, define explicitly when to run, when to stop and when to escalate to a human. Autonomy level should rise only as evidence of quality grows.

3. Collecting requirements

Beyond the business outcome (previous lesson), gather:

  • Functional – tasks, channels (Teams, web, email, phone), languages, systems to read and update.
  • Quality – accuracy targets, acceptable response time, handling of uncertainty ("I don't know" and handoff).
  • Security and compliance – who may use the agent, what data each user may see, data residency, retention, audit needs, regulated content.
  • Responsible AI – transparency that users are talking to AI, human oversight points, fairness and harm risks.
  • Operational – support ownership, monitoring, change management, expected volumes and cost.
  • Constraints – existing licenses, skills, timelines and platform standards.

Interview both the process owners and the people who do the work; hidden exceptions and informal rules are what make agents fail in production.

4. Reviewing grounding data

Grounding supplies an agent with authoritative, current business data so answers are based on facts rather than on model memory alone. Incomplete or stale grounding can cause unsupported answers and wrong actions. Retrieval reduces uncertainty but does not guarantee a correct answer; test citation support and how the agent handles missing or contradictory evidence. Assess each candidate source on five dimensions:

Dimension Question to ask Typical problem
Accuracy Is the data correct and authoritative? Conflicting copies of a policy
Relevance Does it relate to the agent's task and users? A huge repository mostly unrelated to the use case
Timeliness Is it current and refreshed often enough? Outdated price lists, stale documents
Cleanliness Is it consistent, de-duplicated, well formatted, labeled? Duplicates, missing fields, scanned images without text
Availability Can the agent reach it securely, at the needed latency and volume? Data locked in a legacy system without an API or connector

Related checks:

  • Access control – require each retrieval and action to enforce the intended user permissions. Check whether a connector uses the end user's identity or an application/service identity; do not assume the latter automatically trims results per user. Over-shared files can be surfaced more easily, so remediate permissions and test allowed and denied users.
  • Sensitivity – identify personal, confidential or regulated data and apply labels and policies.
  • Structure – structured data (Dataverse tables, SQL, Fabric) suits precise lookups and calculations; unstructured data (documents, SharePoint, websites) suits retrieval of explanations and policies.
  • Ownership – a named owner for each source keeps it current.

5. Preparing data for AI

To make solution data ready for AI and for scalable reuse:

  • Consolidate and standardize definitions, formats and identifiers; prefer a single governed source of truth.
  • Centralize where sensible in a platform such as Dataverse or Microsoft Fabric (OneLake), and use connectors, virtualization or shortcuts rather than copying data when freshness or residency matters.
  • Enrich and label content with metadata so retrieval can filter by topic, region or validity date.
  • Choose a retrieval approach appropriate to the source: connectors and knowledge sources in Copilot Studio, Microsoft 365 data through Copilot, or search and indexing services in Microsoft Foundry for custom solutions.
  • Establish data governance: classification, retention, quality monitoring and lineage.
  • Plan refresh: scheduled sync or event-driven updates so grounding does not drift.

Design data once for reuse: a well-organized knowledge base can serve a Copilot Studio agent, a Microsoft 365 agent and a custom Foundry agent.

6. Worked example

A manufacturer wants an agent to help sales reps answer product-compatibility questions and create quotes.

  • Opportunity: repetitive lookups (data analytics) plus quote creation (task automation). A salesperson reviews every draft quote; quotes above the agreed value threshold also require manager approval.
  • Scenario: a task agent inside the sales team's tools.
  • Grounding review: the product catalog in Dataverse is accurate and current (good). Compatibility notes in PDFs are partly scanned images and contradict each other (poor cleanliness and accuracy). A pricing spreadsheet on a personal drive is not available to the agent and not authoritative.
  • Actions before build: extract and clean the PDFs, designate the product team as owner, move pricing to a governed table, tighten the permissions on the document library.
  • Requirements captured: answers cite the approved source and its version, every draft is reviewed by a salesperson, high-value quotes need manager approval, data stays in the required geography, and activity is auditable.

Work the requirements into a testable acceptance plan:

  1. Establish the baseline. Suppose the team handles 200 compatibility questions a day at eight minutes each. Set a pilot target of five minutes while preserving the agreed answer-quality threshold. A faster wrong answer fails the requirement.
  2. Resolve conflicting evidence. Have the product owner approve one compatibility reference. For scanned pages, verify that the selected ingestion path extracts the required text and tables; add OCR or another supported processing step if it does not.
  3. Define a refusal path. If the approved sources do not establish compatibility, return an uncertainty message and route the question to a specialist. Do not let the model choose whichever contradictory PDF sounds plausible.
  4. Enforce quote approval. The quote service validates permissions and approval state before issuing the quote. A sentence in the agent instructions is not an approval control.
  5. Verify the boundary. Test one salesperson with access to the product range and another without it, including direct tool calls and retrieved citations. Record the source version and action outcome for review.

Use the approved requirements to build an agentic AI strategy and rollout roadmap. Include remediation and refresh work in the AI agent TCO and ROI analysis.

Key takeaways

  • Agents suit multi-step, rule-describable, cross-system work with measurable outcomes; do not use them where simple automation suffices.
  • Match the requirement to a scenario: prompt-and-response, task, autonomous or multi-agent.
  • Define the model, instructions, retrieval, actions, and memory, plus escalation points and enforceable authorization.
  • Evaluate grounding data on accuracy, relevance, timeliness, cleanliness and availability, and respect permissions and sensitivity.
  • Fix data problems before building; organize data once so many solutions can reuse it.

Sources