AI does not automatically improve a bad workflow. Before choosing an agent, automation platform, or model, organizations need to understand the work itself: its inputs, steps, decisions, exceptions, and outputs.
Key Takeaways
- Automating an unclear process can accelerate inefficiency instead of removing it.
- Jobs should be broken into individual tasks before deciding what to automate.
- Predictable, repetitive work is often better suited to deterministic automation than to an AI agent.
- Agentic systems are most useful when workflows involve variability, tool use, or changing conditions.
- Human review, access limits, and escalation paths should be designed before deployment.
Why It Matters
The excitement around AI agents creates an understandable temptation: find a process that feels slow or expensive, connect an agent to it, and let the system take over.
But technology cannot compensate for a workflow that nobody fully understands.
If a business process contains unclear ownership, inconsistent inputs, undocumented exceptions, unnecessary handoffs, or conflicting rules, adding AI may simply make those problems move faster.
The better starting point is operational, not technological.
Before asking, “Which AI tool should we use?” organizations should first ask:
What exactly happens in this process today?
That question sounds basic. It is also where many successful automation efforts begin.
Start With the Work, Not the Job Title
One of the most useful ideas in the source material is that organizations should avoid treating an entire job as a single automation target.
A job is usually a collection of very different tasks.
A customer-service role, for example, might include:
- looking up account information
- updating records
- sending routine confirmations
- investigating unusual problems
- calming frustrated customers
- deciding when an exception should be escalated
Those tasks do not all have the same automation profile.
Some are repetitive and predictable. Others depend on judgment, empathy, or unusual circumstances.
The work-decomposition framework discussed by Ravin Jesuthasan and John Boudreau evaluates tasks along dimensions such as:
- repetitive versus variable
- independent versus interactive
- physical versus mental
That kind of decomposition helps organizations see that automation is rarely an all-or-nothing decision.
Good Automation Starts With Clear Inputs and Outputs
Once a process is broken into tasks, the next step is mapping how the work actually flows.
A useful process map should identify four things:
Inputs
What information enters the process?
This might include:
- customer messages
- forms
- transaction records
- documents
- database entries
- system events
Steps
What actions happen, and in what order?
For example:
- receive a request
- identify the customer
- retrieve account information
- evaluate eligibility
- create a response
- update the system
Decision Points
Where does the process branch?
A workflow may behave differently depending on:
- missing information
- customer status
- amount involved
- approval thresholds
- policy exceptions
Outputs
What counts as a completed result?
That could be:
- an updated record
- a confirmation email
- a report
- a shipping label
- an escalation ticket
This kind of mapping makes hidden complexity visible. It also reveals whether automation is actually solving the right problem.
Sometimes a Simple Workflow Is Better Than an Agent
Not every automation problem requires agentic AI.
Traditional step-based automation is often extremely effective when the workflow is predictable.
A rule such as:
When a customer submits this form, validate these fields, update this record, and send this message.
does not necessarily require an AI agent.
A deterministic workflow can often execute that process more consistently and with less complexity.
Agentic systems become more useful when the process requires the system to interpret changing information, choose between actions, use multiple tools, or adapt its plan when conditions vary.
That creates a useful design principle:
Use the simplest architecture that can reliably handle the work.
An agent should not be chosen because it is more advanced. It should be chosen because the workflow actually requires the flexibility it provides.
The Cost of Automating a Bad Process
Poorly understood processes create several problems when automated.
Exceptions become failure points
If the normal workflow is documented but exceptions are not, automation may work well until the first unusual case appears.
Bad handoffs become faster bad handoffs
An unclear transition between departments or systems does not disappear because AI is involved. It may simply happen faster.
Errors can propagate
If one incorrect input influences several later steps, a small mistake can affect the rest of the workflow.
Complexity becomes harder to see
Automation can hide weak process design behind a polished interface.
The result may look sophisticated while remaining operationally fragile.
That is why process clarity should come before tool selection.
Where Human Judgment Still Belongs
Workflow mapping also helps determine which parts of a process should remain under human control.
Some tasks involve:
- high financial consequence
- legal or contractual decisions
- sensitive customer situations
- unusual exceptions
- ambiguous information
- reputation risk
- significant security permissions
Those tasks may benefit from AI assistance without being appropriate for full autonomous execution.
A strong workflow can allow AI to gather information, prepare an action, or recommend a decision while still requiring a human to approve the final step.
That is not a failure of automation.
It is good system design.
Five Guardrails to Build In Early
The source material supports several practical controls that should be considered before an AI system is given operational authority.
1. Least-privilege access
Give the system only the permissions it actually needs.
If it only needs to read information, it should not automatically receive delete or administrative access.
2. Pause rules
Define conditions that cause the workflow to stop rather than continue automatically.
3. Data boundaries
Identify sensitive information and control what enters model context, memory, or connected systems.
4. Human approval zones
Require explicit approval for actions with significant financial, legal, security, or public consequences.
5. Escalation paths
When the system encounters something outside its expected scope, it should have somewhere to send the work.
A workflow without an escalation path is not fully designed.
Build the Business Case Before the Automation
Process clarity also makes it easier to determine whether an automation project is worth doing.
Before building, an organization should be able to answer:
- What problem are we solving?
- What does the current process cost in time or effort?
- What outcome should improve?
- How will we measure that improvement?
- What tools and integrations are required?
- What will the system cost to implement and maintain?
- Where will human involvement still be required?
If those questions cannot be answered, the project may still be an experiment—but it is not yet a strong business case.
A Practical Readiness Checklist
Before automating a workflow, ask:
Process
- Is the workflow documented?
- Have we broken the work into individual tasks?
- Do we understand the handoffs?
Inputs
- Are required inputs reasonably consistent?
- Do we know where the data comes from?
- Are sensitive-data boundaries defined?
Exceptions
- Do we know what common exceptions look like?
- Is there a fallback path?
- Can unusual cases be escalated to a person?
Architecture
- Could deterministic automation solve the problem?
- If we need an agent, what flexibility does it actually provide?
- What tools and permissions does it require?
Governance
- Are high-risk actions identified?
- Are approval checkpoints defined?
- Does the system know when to stop?
If several of those answers are “no,” the next step may not be building an agent.
It may be fixing the process.
What to Watch Next
The most useful AI systems will probably not be the ones with the most autonomy.
They will be the ones that fit the work.
Organizations that understand their workflows—how tasks connect, where judgment is required, what exceptions exist, and how success is measured—will be in a much better position to use AI effectively.
The technology matters.
But the process comes first.
Sources & Further Reading
- Google, “Agents” whitepaper — Julia Wiesinger, Patrick Marlow, Vladimir Vuskovic, 2025.
- Harvard Business Review / USC, “How to Break Down Work Into Tasks That Can Be Automated” — Ravin Jesuthasan and John Boudreau, 2019.
- Moveworks, “Agentic AI vs. AI Agents: Key Differences Explained” — Amy Brennen, 2026.
- Microsoft Copilot, “AI Automation for Modern Business,” 2026.
- Zapier, AI automation and orchestration guidance, 2026.