Most automation projects begin with a process map.
That is often where the first mistake is made.
The map describes how the process is supposed to work. A request enters through the correct channel. The required information is present. Each system contains an accurate record. The case moves through its approvals and reaches a clear conclusion.
The real process is usually less orderly.
Information arrives by email instead of through the form. An employee notices that two systems disagree. A manager approves an exception in a message that is never attached to the case. Someone maintains a private spreadsheet because the official system does not reveal which requests are urgent. An experienced operator knows which colleague can resolve a problem that does not belong cleanly to any department.
These actions may not appear in the procedure.
They are still part of the work.
When a company automates only the process it has documented, it risks building for an organization that does not exist.
The official process is not the operating process
A standard operating procedure normally describes the expected path.
It identifies the intended sequence, assigned roles and formal controls. This is useful for training, compliance and consistency.
But most procedures are written from the perspective of a successful case.
They say less about what happens when information is missing, two rules conflict, a customer's situation is unusual or an upstream team has made an error. They rarely capture the waiting, verification, reconciliation and informal coordination that surround the visible steps.
This hidden work accumulates in places that are difficult to observe:
- Personal inboxes and direct messages.
- Local spreadsheets and employee-created databases.
- Notes added inconsistently to customer records.
- Meetings held to resolve individual cases.
- Knowledge retained by experienced employees.
- Manual comparisons between systems.
- Unofficial approvals and escalation routes.
The organization may know which applications it operates and which department owns each stage. That does not mean it understands how a unit of work travels from beginning to end.
Microsoft's 2026 Work Trend Index offers a revealing distinction. Its most advanced AI users were considerably more likely than other respondents to report that agent workflows, human handoffs and quality standards were documented and repeatable—not merely within one team, but across functions and the wider organization. Microsoft: 2026 Work Trend Index
The significance is not that documentation itself produces better AI use.
It is that dependable automation becomes easier when the organization can describe what is being delegated.
Digital activity is not the same as operational visibility
Companies often assume that work becomes visible once it happens inside software.
It does not.
A customer request may pass through a website, customer relationship management platform, billing system, ticketing application, email exchange and finance spreadsheet. Every system may record part of the activity while none provides a complete view of the case.
The company has data.
It does not necessarily have a coherent account of the work.
Financial systems organize activity around transactions and cost centres. Customer-service platforms organize it around tickets. Project tools organize it around tasks. Communication platforms organize it around people and conversations.
The workflow may cut across all of them.
Even when records can be assembled chronologically, sequence alone does not reveal intent. A system may show that an employee opened one application, changed a field and then contacted a manager. It does not explain whether the message was an approval, a request for missing information or an attempt to correct a policy conflict.
Research from IBM has highlighted this limitation in process discovery. Many process-mining methods infer relationships largely from the order in which activities occur, even though temporal sequence does not necessarily identify the true causal dependency between them. IBM Research: Causal execution dependencies
Process mining can reveal how cases move across enterprise systems. Task mining can expose desktop actions that do not appear in central records. Both are useful because they replace assumptions with evidence. But neither removes the need to speak with the people performing the work and understand why variations exist.
A complete operational view requires three forms of evidence:
- What the organization says should happen.
- What the systems show did happen.
- What the people involved explain was necessary.
The gaps between those views are often where the most important work is found.
Employees are frequently the integration layer
Many apparently inefficient human activities exist because the underlying process is incomplete.
An employee copies information between systems because no reliable integration exists. Another checks the same record twice because the source cannot be trusted. A manager attends a recurring meeting because no system shows whether a case is becoming risky. A customer-service representative contacts operations because the formal status does not reflect what is physically happening.
The employee is not merely performing a task.
They are connecting systems, interpreting ambiguity and repairing the process while it is running.
Consider a company automating customer refunds.
The written policy may appear straightforward: confirm the purchase, check whether the request falls within the return period and issue the approved amount.
The actual decision may involve more.
Was delivery delayed? Has the warehouse received the item but failed to update its status? Is the customer reporting a defect that should be escalated to the product team? Has the same account made repeated claims? Would a rigid denial damage an important relationship over a minor amount?
An experienced employee may resolve these questions quickly because they understand both the policy and the commercial context.
If the company automates only the written rule, it may produce decisions that are technically consistent but operationally poor.
If it simply reproduces every employee workaround, it may encode years of accumulated inconsistency into software.
Visibility must therefore extend beyond identifying what people do.
The company must understand why they do it.
Hidden work is not always waste
When organizations begin mapping a workflow, they often treat every deviation as inefficiency.
Some deviations are unnecessary.
Others are controls.
A second review may exist because the first system regularly receives incomplete information. A manual reconciliation may prevent duplicate payments. A senior employee may intervene because a low-frequency exception carries unusually high financial or reputational risk.
Removing those steps without understanding their function does not simplify the process.
It removes the mechanism currently keeping it safe.
The reverse is also true. A control may have been introduced years ago to address a problem that no longer exists. An approval may continue because no one has the authority to remove it. Employees may re-enter data because the organization has never revisited a workflow after two systems were connected.
The purpose of visibility is not to defend every existing activity.
It is to distinguish among them.
Some steps should be eliminated. Some should be standardized. Some require better information. Some can be automated safely. Others should remain under human judgment.
Without that distinction, automation becomes a faster way of preserving the past.
Agents make invisible work more consequential
Earlier automation technologies generally performed predefined actions within a narrow environment.
AI agents can operate across a wider boundary. They can interpret unstructured information, retrieve records, compare conditions, invoke tools, update systems and decide what to do next.
That flexibility allows agents to address work that previously resisted conventional automation.
It also makes an incomplete understanding of the process more dangerous.
An agent needs to know more than how to perform an action.
It needs to recognize the current state of the case, determine which information is authoritative, understand what it is permitted to change and identify when the work is complete. It must know which exceptions it can resolve, which require confirmation and where responsibility moves when it escalates.
If those conditions are unclear to the organization, they will also be unclear to the agent.
The result may still look productive. The system generates responses, updates records and moves cases forward. But activity can increase while unresolved work accumulates elsewhere.
Recent evidence increasingly connects AI value with workflow redesign rather than tool access alone. BCG's 2026 workforce research found that respondents in companies redesigning workflows were 24 percentage points more likely to report measurable business improvement than those that were not. BCG: AI is reshaping jobs faster than companies are reshaping work
A separate 2026 field experiment cited by McKinsey gave identical AI tools to two groups of start-ups. The group that redesigned work across handoffs generated 1.9 times more revenue and was 11 percentage points more likely to acquire paying customers than the group using AI only within existing steps. The context was early-stage companies, so the exact effect should not be generalized to every enterprise, but the distinction is important: value appeared at the interfaces between activities, not simply inside the activities themselves. McKinsey: Cutting the coordination tax
Agents make it possible to automate more of those interfaces.
First, the company must be able to see them.
Discovery is not writing a larger procedure manual
Operational visibility does not require documenting every click performed by every employee.
The objective is to create a sufficiently accurate model of how an outcome is produced.
For each important workflow, the company should be able to identify:
- The outcome the process exists to produce.
- The unit of work moving through it, such as a request, order, claim or account.
- The states that unit can occupy.
- The information required to move between states.
- The decisions made and evidence used at each decision point.
- The people, systems and permissions involved.
- The principal exceptions and their destinations.
- The time, cost, failure rate and quality associated with the current process.
- The person accountable for the end-to-end outcome.
This model should be built from real cases.
Ordinary cases reveal the common path. Delayed, reopened and escalated cases reveal where the process depends on knowledge or coordination that the formal design has not captured.
The work should also be observed at the handoffs.
What does the receiving team need before it can act? How often is that information missing? What does it verify again? How long does the case wait? What happens when the two sides disagree?
These questions often uncover more valuable opportunities than timing the visible task itself.
A five-minute activity may sit behind three days of waiting. Automating the five minutes changes very little. Removing the uncertainty at the handoff changes the process.
The first automation decision may be not to automate
Once the real workflow is visible, the organization may discover that automation is not the first intervention required.
The process may need a shared status field, a clearer policy or an integration between systems. An approval may need to be removed. Two teams may be performing the same verification independently. A structured intake form may eliminate more work than an AI system added after the fact.
This leads to a more useful sequence of decisions:
Eliminate activities that no longer serve a purpose.
Standardize variation that does not create customer or commercial value.
Expose missing information, ownership and process state.
Automate frequent work with sufficiently stable inputs, rules and outcomes.
Assist people where context is useful but judgment should remain human.
Escalate cases whose rarity, ambiguity or consequence makes autonomy inappropriate.
The distinction matters because the largest manual task is not always the best automation opportunity.
A repetitive activity may be easy to automate but have little effect on the final outcome. A less visible problem—such as missing information at intake or repeated reconciliation between teams—may cause most of the delay and error.
Automating an unnecessary step makes waste faster. Automating a workaround can make fragility permanent.
Automation should make the work more visible
Visibility should not end when the system is deployed.
A production automation should produce a clearer operational record than the manual process it replaces.
The organization should be able to determine:
- Which cases the system handled.
- Which information and rules it used.
- What actions it performed.
- Where it encountered uncertainty.
- Which cases were escalated and why.
- When a person changed or rejected its decision.
- Whether the intended outcome was ultimately achieved.
This creates a feedback loop between execution and improvement.
If escalations rise, the company can identify whether the cause is poor input data, a policy change or an automation boundary that was set too broadly. If employees repeatedly override one category of decision, the system's instructions or authority may need to change. If a workflow develops new variants, the organization can decide whether to support them or restore standardization.
Microsoft reports that active agents in its ecosystem grew fifteenfold between March 2025 and March 2026. Its research also found that more advanced organizations capture the signals produced by those agents—what worked, what failed and where outcomes drifted—and turn them into shared routines. Microsoft: 2026 Work Trend Index
As automation expands, this form of operational observability becomes more important.
A human employee may notice that a process is deteriorating and compensate informally. Software will usually continue executing the process it has been given until someone can see that the result is changing.
The company needs a map of the work, not a map of its software
An application inventory shows which tools the company owns.
An organization chart shows who reports to whom.
A procedure shows how work was designed.
None necessarily reveals how the business produces an outcome under real conditions.
That requires following the work across departmental and technical boundaries. It means identifying the informal decisions, repeated verifications, invisible queues and employee knowledge on which the formal process depends.
At BraidNest, this is where we begin when improving a workflow.
Before selecting an automation approach, we reconstruct how the work currently moves. We combine system evidence with the knowledge of the people operating the process. We identify which variations protect the company's distinctive value, which compensate for weaknesses and which simply persist through inertia.
That visibility allows the technology to be applied selectively.
Common mechanics can be automated. Important judgment can be preserved. Exceptions can be given clear destinations. The resulting system can produce an operational record that makes the next improvement easier to identify.
The company cannot automate work it cannot see. Once it can see the work clearly, it may discover that the greatest opportunity is not to automate the existing process. It is to redesign it.


