Spend less time chasing information. Give agents the information to do the job. A reference architecture for the agentic enterprise.
An agent can gather the records for a customer renewal and draft a reply. With reliable records and clear rules, it can save the account manager some of the work of looking things up and copying them between systems. To prepare a customer renewal, an agent needs the contract and recent support cases, along with permission to read them. It also needs agreed definitions and rules for what it may do next. An agent needs a model, task state, permission checks, data retrieval, tools, and monitoring. A data platform can supply shared definitions and checked records; live state may come directly from operational systems.
People decide what to delegate, set the limits, and remain responsible for the outcome. An agentic enterprise uses software agents to carry out defined jobs. People set the goal and permissions, and decide when the agent must ask for help. Keep permission checks and business rules outside the model. For each request, supply the relevant records, definitions, task state, source dates, and permitted tools.
“Which customers need attention before renewal, and what should we do next?”
What changes for the business?
Choose a job where people spend time searching, copying, or waiting for information. Count the time they spend checking the automated work, and include running costs when you assess the result.
A useful service connects the records, applications, and people needed to finish a job.
People and agents work from the same definitions, current facts, and visible evidence.
An agent can collect records from several systems without asking a colleague to copy each one.
Teams can share datasets, definitions, integrations, and checks across assistants.
People can delegate more work while keeping responsibility for the rules, exceptions, and decisions with serious consequences.
The next workflow can use the same maintained datasets, definitions, and tested action tools.
Agents and models will change. Maintained data, shared definitions, access rules, and tested tools give the next service a useful starting point.
What does the agent need to know?
For a renewal request, give the agent the information it needs for this job: business definitions, the dates of the records, who is asking, what they may do, and the sources behind the answer.
Reuse needs an owner who keeps the records current.
A dataset that works for a monthly report may need more frequent updates before an agent uses it to book a shipment. Check whether its definitions, access rules, and quality suit the new job.
How the request moves through the enterprise
The agent depends on applications, data, tools, and infrastructure. Each needs clear responsibilities and access rules. The data platform helps teams share checked information and agreed definitions.
- Where people workApplications & channels
- ComponentsMicrosoft 365, analytics apps, custom applications, chat, search, notifications.PurposeReceives the request and shows the result in the applications people use.
- DelegationAgents & automation
- ComponentsAgent experiences & automation, custom agents, workflow automation, human checkpoints.PurposeInterprets goals, plans tasks, coordinates tools, and manages state.
- ReasoningAI runtime & orchestration
- ComponentsModels and inference, routing, retrieval, memory, tool calling, evaluations, cost controls.PurposeRuns the model, selects tools, and keeps track of the task.
- Shared contextEnterprise data and intelligence platform
- ComponentsWarehouses and lakehouses, pipelines, curated data products, semantic layers, catalogs, lineage, quality and freshness monitoring.PurposeMakes checked records and business definitions available with access rules, owners, and quality monitoring.
- ExecutionSystems of record & action
- ComponentsCRM, ERP, line-of-business systems, files, operational stores, APIs, event streams.PurposeStores business records and checks requests to change them.
- FoundationInfrastructure & platforms
- ComponentsCloud and data centers, networks, identity, endpoint management, Microsoft licensing.PurposeRuns the services and networks, manages access and licenses, and meets requirements for where data stays.
When every workflow starts from scratch
- Teams reconnect the same sources for each workflow.
- Definitions drift between analytics, workflows, and assistants.
- Missing or stale records can reach the answer unchecked.
- Teams copy permissions and business rules into separate prompts and codebases.
- Teams struggle to reconstruct what happened after an action.
What a maintained platform makes reusable
- The next workflow can use the same checked customer records.
- Shared definitions keep “renewal date” consistent across services.
- Consumers can check quality, update times, and source history.
- Teams can use shared interfaces with access checks.
- Service owners use recorded results to find and fix problems.
Give each capability an owner.
A customer request can involve several teams. Select a capability to see its responsibilities and the information or services it needs from others.
Shared records, definitions, permissions, and sources. The data platform maintains the reusable data; other teams supply live information and controls.
Shared information, named ownersSomeone must correct the records, maintain access rules, fix failed actions, and support the people using the service. Agree these responsibilities before the service goes live.
How much should the agent do on its own?
An agent may help with a draft, work through a task with approval, or act within agreed limits. Give it more freedom only when permissions, monitoring, and a way to recover from mistakes are in place.
The human directs the task and uses the agent to find, summarize, draft, or analyze.
The person checks the answer and decides what to do with it.
The agent plans and executes multiple steps, pausing at named checkpoints.
The agent needs the rules, progress so far, tool contracts, and records that a reviewer can check.
The agent operates within a bounded goal, authority, budget, and escalation path.
The team must keep context current, monitor the work, and provide ways to stop or recover a failed run.
The contracts behind this workflow
The renewal and replacement-shipment services need explicit contracts for data, context, actions, and evaluation. These contracts give the delivery process something concrete to test and operations something to monitor.
Resolve the user, role, goal, workflow, risk tier, and current case.
Identity + task stateSelect domains, time boundaries, policy, memory, and permitted sources.
Context policyQuery semantic data and knowledge, rerank evidence, and expose uncertainty.
Retrieval + semanticsChoose steps and tools, validate parameters, and require approval where the rules specify it.
Orchestration + policyInvoke bounded, idempotent operations with explicit side effects and recovery.
Tool contractsCapture sources, tool calls, outcome, latency, cost, correction, and drift.
Evidence + evaluationMake the interfaces explicit
Write down what each service promises, what it requires, and how it fails. Owners and automated tests must be able to check those promises.
Data product contract
Defines what a governed dataset means, who owns it, when it is fit for use, and how consumers can discover and access it.
Context contract
Specifies how the service combines identity, task state, time, permitted sources, definitions, evidence, and uncertainty for each request.
Tool & action contract
Defines a bounded operation the agent may invoke, including permissions, validation, side effects, failure behavior, and audit evidence.
Evaluation contract
Sets the acceptance criteria for release and daily operation: data quality, relevant search results, permitted actions, and business results.
Observe the whole chain
A correct model response can still produce a poor enterprise outcome. Monitoring must connect the data, retrieval, reasoning, tool, and business layers.
| Layer | What to observe | Typical failure | Primary response |
|---|---|---|---|
| Data | Completeness, validity, freshness, distribution, lineage | Stale or contradictory facts | Stop or degrade the affected data product; route to its owner |
| Retrieval | Coverage, relevance, authorization, source diversity | The search misses relevant records or returns records the user cannot access | Adjust retrieval policy, indexing, ranking, or permissions |
| Reasoning | Instruction adherence, groundedness, uncertainty, model drift | The answer makes a claim the records do not support | Constrain the task, require evidence, change model or review gate |
| Tools | Validation, authorization, side effects, retries, recovery | The tool repeats an action, leaves it unfinished, or cannot undo it | Fix the tool contract and specify who handles failures |
| Outcome | Human acceptance, correction, cycle time, risk, value | The workflow works technically but does not improve the business | Change the workflow, context, ownership, or stop the use case |
Architecture traps to avoid
- Embedding business definitions and permissions inside prompts that no owner governs.
- Letting every agent create its own source integrations and private memory.
- Measuring model quality without testing retrieval, tools, and business outcomes.
- Using vector search as a substitute for curated data products and a semantic layer.
- Granting broad system access instead of exposing bounded, observable actions.
Start with one job worth improving.
Separate assistants can duplicate connections and checks. Fund a useful job and the shared information and tools it needs.
Choose a specific job
Choose a workflow where better information could save time, reduce errors, or improve the result for the customer. Name its owner and decide what needs approval before choosing an agent.
Prepare the information
Agree on the definitions, source records, access rules, and quality checks this job needs. Assign owners and document how other teams can use them.
Try the whole job
Connect the application and agent to the model, records, and tools they need. Test the job from request to result with the people who will use it, including the steps they must approve.
Reuse what works
Reuse data products, semantic models, tools, controls, and evaluations across the next workflows. Let usage and outcomes guide platform investment.
Questions before investing
From renewal reminder to delivery follow-up
The account manager asked about a renewal. Once the agent found the unresolved delivery case, it proposed contacting the service owner first. To make that recommendation and act on it, the service needs answers to four questions:
- Which facts can we rely on?
- What do those facts mean?
- Which actions are permitted?
- Can we check what happened?
The platform keeps the contract and delivery records available for the next request, with the same definitions and access rules. The account manager still approves commitments to the customer, and the service owner remains responsible for resolving the delivery issue.