Businesses are moving quickly to adopt AI Agents. Deploying an Agent, however, is not the same as transforming how a business operates.
As organizations move beyond AI experiments, they face a larger question:
How should work get done when AI Agents can execute alongside people?
Research from Deloitte suggests that realizing the potential of agentic AI will require organizations to redesign work and the systems that support it. The opportunity extends beyond automating individual tasks. It involves rethinking how work moves across people, Agents, data, and business systems.
For businesses exploring agentic AI, the goal should not be to deploy as many Agents as possible. A better approach is to:
Identify where work slows down.
Select a high-volume, rule-based task with measurable value.
Define the Agent’s inputs, actions, and limits.
Keep decisions requiring judgment with the appropriate people.
Expand from one task to a coordinated workflow.
That is how an organization can move from AI-assisted tasks to a redesigned system of work.
The Agentic AI Market Is Still Early
Interest in AI Agents has grown rapidly, but scaled adoption remains relatively uncommon.
Deloitte surveyed 501 U.S.-based leaders directly involved in their organizations’ agentic AI strategies or implementations. The research also included interviews with 20 executives and AI and data science leaders. Deloitte published the research on August 12, 2026.
Area measured | Deloitte finding |
|---|---|
Testing small numbers of Agents or operating only a few deployments | 42% |
Expanding Agent deployment across functions | 43% |
Scaled, orchestrated, multi-Agent adoption | 15% |
Organizations whose processes are prepared for agentic adoption | 16% |
Organizations whose processes are highly prepared | 5% |
Scaled adopters that say their processes are prepared | 46% |
Leaders citing an inaccessible or fragmented data foundation as a barrier | 72% |
Leaders citing trust and Agent governance as a barrier | 70% |
Leaders citing integration cost and complexity as a barrier | 67% |
Leaders expecting nearly half of their processes to be redesigned around Agents within four years | 74% |
Leaders reevaluating their business models because of agentic AI advances | Nearly two-thirds |
The figures show why deploying an Agent is only an early step. Companies increasingly understand that Agents can change how work gets done, while many are still determining which processes to redesign and how to manage that change.
For mid-market organizations, the practical challenge is often choosing the right place to start. The first workflow should have enough volume to produce measurable results, clear rules that an Agent can follow, and a visible connection to business performance.
Deploying an Agent and Redesigning a Process Are Different
Adding an Agent to an existing process usually means inserting automation into one step. Redesigning the process means examining the complete path from the initial request to the final system update or business outcome.
Consider an order-intake process:
An email arrives with an order document.
Someone downloads the attachment.
The information is extracted and matched to a customer.
The employee checks product, account, and order details.
Missing information is requested from the customer or salesperson.
The order moves through an approval path.
An employee updates the ERP or CRM.
Someone checks the status and follows up when work stalls.
An Agent could accelerate document extraction. That may reduce one manual task while leaving the rest of the coordination unchanged.
A redesigned workflow examines the entire sequence:
Workflow stage | Agent role | Human ownership |
|---|---|---|
Email and document intake | Monitor approved sources and collect relevant files | Handle unusual or unrecognized requests |
Data extraction | Read documents and structure order information | Review low-confidence extraction |
Customer and data enrichment | Match the request to the correct customer, account, product, or project record | Resolve ambiguous matches |
Validation | Check required fields, pricing rules, product data, and account conditions | Approve exceptions or unusual terms |
Missing-information follow-up | Identify gaps and send requests through approved channels | Manage sensitive customer communication |
Approval routing | Send work to the correct person based on defined thresholds | Make pricing, credit, contractual, or commercial decisions |
ERP or CRM update | Write approved information to the system of record | Confirm changes when the risk or value exceeds a threshold |
Status visibility | Track progress, surface blockers, and notify owners | Resolve exceptions and make decisions |
The Agent executes repeatable work within defined rules. People retain responsibility for pricing approvals, customer decisions, exceptions, relationship management, and judgment calls.
This is the difference between Agent-assisted work and Agent-enabled workflow design. The first adds automation to a process. The second changes how the process operates from end to end.
Most Business Processes Need Preparation Before Agents Can Operate Reliably
Deloitte found that only 16% of respondents said their business processes were prepared for agentic adoption, while 5% said they were highly prepared. Among organizations that had already adopted Agents at scale, 46% said their processes were prepared.
Agents still operate within the realities of a business:
Data sits across ERP, CRM, document storage, email, spreadsheets, and portals.
Different teams use different approval rules.
Important steps may exist only in employee knowledge.
Processes contain exceptions that standard procedures do not describe.
Records may use inconsistent customer, product, project, or order information.
Employees may update one system while relying on another for status.
Before an Agent can execute work reliably, the organization needs to understand the work itself. That preparation includes:
Mapping the current process. Document each handoff, decision, system, input, and output.
Defining the system of record. Identify where approved information belongs.
Standardizing inputs. Record acceptable document types, fields, naming rules, and required data.
Documenting business rules. Specify validation checks, approval thresholds, and escalation paths.
Classifying exceptions. Separate routine cases from cases that require human judgment.
Assigning ownership. Name the person or team responsible for approvals and unresolved issues.
Establishing access controls. Limit the systems and records each Agent can read or change.
Defining audit requirements. Record the Agent’s inputs, actions, decisions, and handoffs.
This work creates the operating conditions an Agent needs. It also gives employees a clear way to review and improve the process after deployment.
Connecting Fragmented Business Systems
An Agent cannot create a reliable outcome from incomplete or disconnected information. It may need to read an inbound email, extract a document, match a customer in the CRM, check product or account data, apply an approval rule, and update the ERP.
That requires more than a standalone chatbot. It requires a controlled workflow that connects:
Email and communication tools
PDFs, spreadsheets, and document storage
Customer and account records
Product, project, and order data
ERP and CRM systems
Approval queues
Status and reporting views
The Agent should use each system for the information it owns, then write approved results back to the designated system of record. A centralized workflow layer can help coordinate those steps while preserving existing business systems.
Vsimple is a quote-to-cash and process-automation platform that supports ERP API connections and customized workflows for tasks such as order management, service tracking, document approvals, email chains, and spreadsheets.
The Biggest Agentic AI Challenges Are Operational
Deloitte identified three major barriers to scaling AI Agents:
72% cited the lack of a unified and accessible data foundation.
70% cited challenges involving trust and Agent governance.
67% cited the cost and complexity of integration.
These barriers describe the operating environment around an Agent.
An Agent needs access to the right information. It needs clear instructions about what happens next. It needs boundaries that define what it can execute independently and what requires approval. It also needs a dependable way to work across the systems where the business already operates.
Governance should answer practical questions:
Which records can the Agent read?
Which systems can it update?
Which actions require approval?
What dollar, contract, credit, or customer thresholds trigger escalation?
How does the system handle missing or conflicting information?
Who reviews the Agent’s work?
How can the organization reconstruct what happened?
Autonomy should increase only when the process has clear rules, reliable data, acceptable risk, and a review path. An Agent can handle repeatable execution. A person should own decisions involving ambiguous information, material financial impact, customer commitments, legal or contractual interpretation, and sensitive exceptions.
A Practical Framework for Redesigning Work
The path from a Task Agent to an end-to-end workflow can follow a repeatable sequence.
1. Map the current process
Record the work as it happens today. Include emails, documents, spreadsheets, system updates, approvals, rework, and status checks.
The map should show:
The event that starts the process
The information required to begin
Each person and system involved
Every decision point
Common exceptions
The final business outcome
Where delays and manual handoffs occur
This step often reveals that the largest cost sits between systems rather than inside one task.
2. Select one high-value task
Choose a task with:
High or predictable volume
Repeatable rules
Structured or partly structured inputs
A measurable business outcome
A manageable risk profile
A clear owner
Order intake, document classification, data validation, status updates, and approval routing may fit these criteria when the rules are documented.
The first task should be narrow enough to control and important enough to measure. Automating a low-volume task with unclear rules produces limited evidence about the value of agentic work.
3. Define inputs and outputs
Specify what the Agent receives and what it must produce.
For an order-intake workflow, inputs may include an email, an attachment, a customer record, product data, and account rules. Outputs may include a validated order record, a request for missing information, an approval task, an ERP update, and a status notification.
The organization should also define the system of record for each output. This prevents the Agent from creating another disconnected copy of business data.
4. Establish approval thresholds and escalation paths
Write the rules that determine when the Agent can proceed and when a person must review the work.
Examples include:
Route pricing outside an approved range to sales leadership.
Route unusual payment terms to finance.
Escalate a customer match that falls below a defined confidence level.
Hold an order when required information is missing.
Require approval before changing a contractual commitment.
Send unresolved exceptions to the team that owns the process.
These rules give the Agent a controlled operating boundary.
5. Measure the workflow
A useful baseline should exist before deployment. Relevant measures include:
Measure | What it shows |
|---|---|
Cycle time | How long the process takes from intake to completion |
Manual touches | How many human actions the process requires |
Error or rework rate | How often information must be corrected |
Touchless completion rate | The percentage of eligible cases completed without manual intervention |
Exception rate | How often the Agent needs human help |
First-response time | How quickly the process acknowledges or acts on new work |
Status visibility | Whether owners can see the current state and blocker |
Cost per transaction | The operating effort required for each completed case |
The right business result may be faster order processing, fewer errors, better follow-up, improved visibility, or more employee capacity. The organization should connect the chosen measure to that result.
6. Expand from task to role or coordinated workflow
Once the Agent performs a defined task reliably, the organization can connect related work.
For example, an order-intake Agent may begin by extracting information. A broader role-based system could then:
Validate the order against customer and product data.
Request missing information.
Route pricing or credit approvals.
Update the ERP and CRM after approval.
Notify the responsible salesperson.
Track the order until the next handoff.
Surface stalled work to an operations manager.
The people in that role continue to own customer relationships, commercial decisions, exceptions, and accountability. Agents coordinate the repeatable execution around those responsibilities.
This creates a progressive path:
Task → Role → Coordinated workflow → System of Execution
When not to automate
A process may not be ready for Agent execution when it is:
Ambiguous and poorly documented
Too low in volume to justify the setup
Governed by rules that change frequently
Dependent on sensitive personal or contractual judgment
High consequence when an error occurs
Built on unreliable or inaccessible data
Missing a clear owner or approval path
In these cases, the organization can first use AI to help document the process, find information, prepare a draft, or identify missing data. Full autonomous execution should wait until the business can define acceptable outcomes and review requirements.
Start With a Task. Build Toward a Role.
The idea of redesigning an entire organization around AI can sound overwhelming. It does not need to happen all at once.
At Vsimple, the practical approach starts with a specific piece of work where repetitive effort, clear rules, sufficient volume, and measurable impact make automation useful.
A Task Agent might process an inbound document, extract information, validate it against business rules, identify missing details, and update the appropriate system after approval.
Once the organization establishes trust and proves value, it can expand the workflow. Multiple Agents can coordinate the work associated with a role or function while people focus on relationships, decisions, strategy, and exception management.
Those role-based workflows can later connect across functions. Work can then move between people, Agents, and existing systems with less manual coordination and greater visibility.
AI Agents Should Give People More Capacity for Important Work
Agentic transformation changes the relationship between people and work.
Deloitte’s research describes a shift toward Agents handling more autonomous execution while people take on oversight and work requiring judgment, creativity, relationships, and other human capabilities.
The goal is to reduce repetitive digital work that consumes employee capacity. The people who understand the business remain responsible for decisions and relationships.
For example:
A sales coordinator can spend less time copying information between systems, chasing missing details, checking statuses, and routing documents.
An operations leader can see where work is blocked without maintaining another spreadsheet.
A salesperson can spend more time with customers while the workflow handles required administrative updates.
A service or project team can focus on exceptions and customer outcomes instead of repeated status checks.
Agents handle repeatable execution within defined boundaries. People handle customers, decisions, exceptions, relationships, and growth.
From AI Adoption to a New System of Work
The next phase of AI will not be defined simply by which companies have access to Agents. The more useful distinction is how organizations put those Agents to work.
An agentic operating model brings several elements together:
Connected data
Documented processes
Clear approval rules
Governed Agent actions
Existing ERP and CRM systems
Human review and escalation
Measures tied to business outcomes
Visibility into work status and exceptions
Organizations that treat an Agent as another isolated software feature may automate one task. Organizations that redesign the workflow around the Agent can improve how work moves across the business.
The path is practical:
Start with a Task Agent. Solve a defined process and measure the result.
Expand into a Role-Based System. Coordinate related tasks while people manage decisions and exceptions.
Build toward a System of Execution. Connect people, Agents, workflows, and existing systems so work progresses with less manual coordination and clearer accountability.
Frequently Asked Questions
What is the difference between deploying an AI Agent and redesigning a business process?
Deploying an AI Agent automates one or more actions inside an existing process. Redesigning a business process examines the full path from intake to outcome, removes unnecessary handoffs, connects the required systems, defines approval rules, and assigns clear ownership for exceptions.
How should an organization identify its first workflow for an AI Agent?
It should select a high-volume task with repeatable rules, reliable inputs, measurable results, and a clear owner. The task should have enough business impact to justify implementation while remaining narrow enough to govern and review.
What business processes need to be prepared before Agents can operate reliably?
The organization should document the current workflow, standardize inputs, identify systems of record, define business rules, classify exceptions, assign owners, set access controls, and establish audit requirements. These steps help an Agent operate consistently across email, documents, ERP, CRM, and other business systems.
How should an organization decide when a human must approve an Agent’s work?
Human approval should apply to ambiguous cases, material financial or contractual decisions, sensitive customer situations, low-confidence data matches, policy exceptions, and actions that exceed defined thresholds. The organization should define these conditions before deployment and create an escalation path for each one.
How do AI Agents work across fragmented business systems?
A governed workflow coordinates the systems involved in a process. The Agent can collect information from email or documents, match it to customer and product records, validate it against business rules, route approvals, and write approved results to the ERP or CRM system of record. Access controls and audit logs help limit and track those actions.
How should businesses measure the operational impact of agentic AI?
They should establish a baseline before deployment and track measures such as cycle time, manual touches, error rate, rework, touchless completion, exception rate, first-response time, status visibility, and cost per transaction. The organization should connect these measures to a business result, such as faster processing, fewer errors, or greater employee capacity.
What are the main barriers to scaling AI Agents beyond pilots?
The main barriers are fragmented data, limited trust and governance, integration cost and complexity, undocumented processes, unclear ownership, inconsistent rules, and insufficient AI fluency. Addressing the process and operating model is necessary before adding more Agents.
Agents Are the Beginning, Not the Destination
AI Agents represent an important technological shift, but the Agent is only one part of the transformation.
The larger opportunity is redesigning the processes, systems, and roles surrounding it. Businesses that treat agentic AI as another software layer may automate individual tasks. Businesses that rethink how work moves across their organizations can create operations where routine work executes automatically, exceptions reach the right people, systems stay updated, and employees have more capacity for work that requires judgment.
That is the path from adopting AI to building an agentic business.
It starts with a practical question:
If an Agent could execute this work, how should the process work in the first place?
Source: Deloitte Insights, AI agents are only the beginning: The path to agentic transformation, China Widener, Laura Shact, David Jarvis, and Sayantani Mazumder, August 12, 2026
Share this article:
