Pattern 4: Core business process transformation

Agents run core enterprise processes end-to-end across multiple systems. In this pattern, agents make decisions within defined boundaries, not just surface recommendations, with direct impact on revenue, cost, and customer experience.

In this pattern, agents orchestrate complex workflows autonomously and escalate exceptions to humans. Humans shift from doing the work to governing the system.

Important

Accountability in this pattern stays with the business, not IT. Before pursuing this pattern, confirm that the responsible business owner (not a technology team) is prepared to own outcomes.

What this pattern is

Core business process transformation rebuilds a business-critical process around agents. The agent orchestrates work across systems, decides the routine cases within set limits, and routes the exceptions to a human. People move from doing the work to governing the system that does the work.

This pattern is firmly on the execute side of the assist-to-execute shift, where the stakes are the highest. The agent acts with real financial and customer impact, so the four new demands of execution apply with full weight: named ownership, a defined risk response, lifecycle management, and explicit authority limits. Accountability stays with the business, not with IT.

What agents do

Agents in this pattern orchestrate complex, multi-step workflows across systems. For example:

  • Claims processing: Intake, verification, risk scoring, payment
  • Order-to-cash: Order validation, fulfillment coordination, invoicing
  • Financial close: Reconciliation, accrual calculations, variance flagging
  • Manufacturing quality: Inspection triage, defect classification, routing
  • Supply chain: Demand sensing, inventory optimization, supplier coordination
  • Customer onboarding: Know-Your-Customer verification, account setup, welcome workflows

Agents make routine decisions autonomously within defined boundaries and escalate exceptions that fall outside those boundaries to human reviewers.

What humans do

Humans shift from doing the work to governing the system:

  • Define autonomy limits: specify which decisions agents can make and which decisions require human approval.
  • Review exception cases that fall outside agent boundaries.
  • Monitor business outcomes against defined key performance indicators (KPIs).
  • Continuously improve the process based on performance data.
  • Own accountability for business results.

This shift is a fundamental role transformation. People in these processes need new skills, new mental models, and explicit support for the transition.

How the operating model works

Successfully deploying this pattern requires changes across four dimensions:

Dimension Before After
People Doing work Owning performance and governing the system
Agents Executing steps Orchestrating end-to-end across systems
Governance Tool-use policies Autonomy limits and decision rights framework
Metrics Productivity gains Cycle time, throughput, accuracy, and business value

Target maturity profile

This pattern requires depth across all five capability drivers. You can't skip any drivers.

Capability driver Target level Why
AI strategy and experience 500 (Efficient) Run the process as an adaptive, AI-first capability, not a project.
Business strategy 500 (Efficient) The scale-breaker. Formally redesign the process and link it to KPIs.
Governance and security 400 (Capable) Proactive governance, autonomy limits, and automated monitoring for financial and customer-impacting actions.
Technology and data 400 (Capable) Scalable foundations and reliable integration across multiple systems.
Organization and culture 400 (Capable) People take on a new role governing the system. Their jobs change, not just their tools.

Key insight: This pattern demands depth everywhere. At 400–500 level maturity across all five drivers, embed agents in core business operations with direct impact on profit and loss. Don't deprioritize any drivers.

Scale-breaker: Business strategy, specifically formal process redesign and KPI-linked orchestration. Redesign the process before you automate it. Otherwise, you merely automate a broken process faster. Discipline comes before technology.

Business units own agent delivery and outcomes. The CoE provides standards, enablement, and governance by exception. This structure requires mature teams in the business that own the full agent lifecycle: not just build agents, but also run, monitor, and improve them.

Tip

The make-or-break role for this pattern is the business owner. Accountability must sit in the business, not IT. If the business owner treats this role as an IT project, the pattern stalls when it matters most.

Learn more about structure, roles, and risk-tiered governance in Build an agentic Center of Excellence.

What you need and don't need

You need:

  • Process redesign with human accountability: Define who owns the outcome before you design the agent. Ownership of outcome is separate from ownership of the technology.
  • Decision rights framework: Document which decisions agents can make autonomously and which require approval. The business owner must formally endorse this framework.
  • Autonomy limits: Define clear boundaries for agent actions, especially those with financial or customer impact. Set these limits conservatively at launch and expand them as trust is established.
  • Production-grade monitoring: Track SLAs, accuracy, and drift, not just uptime. Agents drift silently. Monitoring catches drift before it affects customers.
  • Federated governance: Business units own agents and outcomes. The CoE provides standards and oversight.
  • Change management: The humans in this process need a new role, not just a new tool. Invest in helping them understand and embrace the shift.

You don't need:

  • Individual enablement programs: This pattern isn't about personal productivity.
  • Lightweight governance: This pattern has profit-and-loss impact and requires formal controls.
  • Centralized IT delivery: The business must own outcomes. IT enables, it doesn't own outcomes.
  • Incremental automation: You're redesigning the process, not patching it with automation.
  • Skipping change management: People's roles fundamentally change in this pattern. Change management isn't optional.

Value and success metrics

Value shows up as step-change improvements in core operations and in direct gains in business KPIs. Measure the process, not the tool. Track how much runs without human input, how fast it runs, how accurate it is, and how it affects the business metrics you already track.

What value looks like

  • Step-change improvements in cycle time, throughput, and quality.
  • Lower operating cost through fewer handoffs and less waiting.
  • More resilient, predictable core operations.
  • Direct impact on business KPIs such as revenue, cost, and service levels.
  • Teams that shift from doing the work to owning performance.

Success metrics to track

Category Example measures What it tells you
Process performance Straight-through-processing rate, cycle time, throughput, exception rate How much of the process runs end-to-end without a human, and how fast.
Quality and risk Accuracy and error rate, model and decision drift, audit-readiness Whether autonomous decisions stay correct and compliant over time.
Economics Cost per transaction, hours and cost saved, capacity rebalanced Whether the redesign lowers unit cost and frees people for higher-value work.
Direct business KPIs Revenue accelerated or protected, working-capital and cash-flow gains, service level Whether the process change moves the numbers the business already tracks.
Governance Autonomy-limit adherence, exception-review SLA, agent inventory coverage Whether people are governing the system effectively as it scales.

How to measure

  • Production-grade monitoring tracks straight-through processing, accuracy, and drift, not just uptime.
  • Microsoft Agent 365 provides the registry, identity, and observability to see and govern the agents running the process.
  • Microsoft Purview and Dataverse give auditable decision logs and a system of record for process data and business rules.
  • Tie agent metrics to the existing process KPIs so value is reported in the language the business already uses.

Tip

Don't automate a broken process. It scales the breakage. A poorly designed flow can seem to save time, but it can hide rework, exceptions, and downstream risk. First, redesign and instrument the process. Set explicit autonomy limits for financial decisions and decisions that affect customers. Watch for model drift, not just availability.

Learn more:

Common anti-patterns

The failures here are expensive because the process touches revenue, cost, and customers.

Anti-pattern What it looks like What to do instead
Process redesign is skipped The team automates the existing process step by step instead of redesigning it for agent-first execution. The result is faster execution of a broken process, with added human-agent handoff complexity. Always redesign the process for agent-first execution before automating, and link it to KPIs.
IT owns the outcome The technology team builds and runs the agent, but the business doesn't own the outcome. When performance drifts or errors occur, no one in the business is accountable, and the agent becomes an IT project rather than a business capability. The business owns the outcome. The CoE sets standards and governs by exception.
Lightweight governance Assist-grade controls are applied to a process with profit-and-loss impact. Use formal controls: decision rights, autonomy limits, and production monitoring.
Autonomy limits aren't defined before deployment The agent goes live with no documented boundaries for what it can decide on its own, and edge cases cause unexpected outcomes. Define the decision-rights framework before deployment. Start conservative and expand autonomy as trust is established.
Monitoring is reactive, not proactive The team monitors uptime but not accuracy or drift. The agent slowly degrades, giving wrong answers with confidence. Make production monitoring proactive: track accuracy and detect drift and anomalies, not just availability.
Skipping change management People's roles change and no one prepares them. The new model stalls. Manage the change: people move from doing the work to governing it.
Big-bang cutover The whole process flips to agents at once, with no fallback. Stage the rollout, keep exception paths, and expand autonomy as confidence grows.

Customer stories

These published Microsoft customer stories show agents running core processes across systems, with humans governing outcomes.

Customer Reported outcome
EY Rebuilt finance processes on Power Apps and a Copilot Studio agent integrated with SAP, reducing general-ledger lead times by 95% and saving 120,000 hours a year on payment clearing.
Lifetime Products Dynamics 365 finance agents automate reconciliation and order processing, with a 95% increase in e-commerce order processing and the ability to rebalance 20% of the knowledge workforce to higher-value work.
Cineplex A Copilot Studio refund agent orchestrating Power Automate across systems cut refund handling from 5 to 15 minutes down to under a minute, part of more than 30,000 hours saved a year.
Dow Inc. An autonomous Copilot Studio "Freight Agent" that analyzes more than 100,000 freight invoices a year, identifying significant savings.
University of Leicester Saved 312 hours a year handling 30,000+ student queries with Dynamics 365 AI agents.
TotalEnergies Improved fault-cause identification by 25 to 45% with autonomous agents (MARGE for maintenance, BuyerCompanion for procurement).
Signetic Serves 800,000+ patients with Power Platform and Copilot Studio agents automating prior authorization and claims.
KONE Processes 54,000 contracts a year with handling time cut by roughly two-thirds, using a Copilot Studio agent and AI workflows.
KPMG Cut 18 months from compliance timelines with an enterprise ESG agent automating the policy and compliance lifecycle.
Banco Ciudad Freed 2,400 hours a year at 90% approval with 10 internal Copilot Studio agents built in six months.

Microsoft agentic capabilities for this pattern

The following examples highlight capabilities that are particularly relevant to Core business process transformation. Use these examples as a starting point for matching capabilities to your scenarios and user segments, and then explore Microsoft's agent ecosystem for the complete picture.

Orchestrate the process

  • Copilot Studio autonomous agents act on events and schedules within scoped permissions so the process runs without a person prompting each step.
  • Multi-agent orchestration lets a lead agent coordinate specialized agents across the steps of a process.
  • Workflows add deterministic, rule-based steps for the parts of the process that must be predictable.
  • Dataverse provides a secure, governed system of record for process data and business rules, with row- and column-level security, role-based access control, field-level encryption, audit logging, and built-in compliance controls that make every agent action traceable and auditable.
  • Connectors and the Model Context Protocol reach line-of-business systems such as SAP and Dynamics 365.
  • Dynamics 365 prebuilt agents, such as account reconciliation, supplier communications, and procurement agents, run specific finance and supply-chain processes.

Build pro-code when you need to

Microsoft Foundry builds and hosts custom pro-code agents and models for processes that exceed low-code scope.

Keep humans governing the system

  • Human-in-the-loop approvals and decision boundaries require sign-off above set risk thresholds and keep agents inside their authority limits.
  • Microsoft Agent 365 registers agents, gives each a Microsoft Entra Agent ID, enforces least-privilege access, and provides observability across the estate. Agent 365 is generally available for commercial customers with some scenarios still in preview. Confirm coverage for your scenario.

Evaluate and test before you ship

For a pattern with profit-and-loss impact, evaluation isn't optional. Test agents against realistic scenarios before they touch live processes, and continue evaluating in production as the process evolves.

  • Agent evaluation in Copilot Studio runs structured evaluations against your agent, scores responses across quality dimensions such as accuracy, groundedness, and task completion, and surfaces regressions before they reach production. Use it to validate that autonomy limits and routing logic behave as designed across a range of inputs, including edge cases and exception scenarios.
  • Copilot Agent Kit (formerly known as Copilot Studio Kit) extends test coverage with bulk testing, automated regression runs, and a scoring dashboard. It's particularly useful for core business process agents, where you need repeatable, auditable evidence that the agent behaves correctly across the full range of inputs the live process will encounter.

How to know you're ready

Start this pattern if most of these statements are true:

  • You've formally redesigned the process and linked it to KPIs.
  • The business owns the outcome, with named accountability.
  • You've defined decision rights and autonomy limits, especially for money and customers.
  • You can run production-grade monitoring, including drift detection.
  • You have mature teams that can run and improve agents, not just build them.

Next steps

Take the Agent Readiness Assessment

Or explore the full Agentic AI adoption maturity model.

Explore other patterns in more detail:

  • Last updated on