Why AI Workflow Redesign Must Start With How Work Actually Happens

11 November 2026 · 6 min read

AI workflow redesign should start with the work as it actually happens, not the procedure manual or org chart.

A process may be documented in six steps but operate across eleven, with extra checks happening through spreadsheets, emails and informal agreements. If AI is designed around the documented version, those real steps, handoffs and exceptions are missed.

Map the workflow as it runs before deciding where AI belongs. Within AIVOM™, this sits in the Design dimension and shapes the decisions that follow.

The Gap Between Documented Procedures and Actual Work

What Procedure Manuals Leave Out

Procedure manuals define formal steps, approval authority and compliance requirements. They explain how work is supposed to happen, but leave out the decisions and informal checks that keep it moving.

A finance director may be responsible for approving purchases above a certain threshold. Yet before a request reaches them, it may pass through several informal checks that are absent from the documented process. The approval structure is accurate, but the workflow is incomplete.

Where Operational Knowledge Actually Sits

Experienced teams develop ways of working that rarely make it into formal documentation. They use shared spreadsheets to track exceptions, informal conversations to resolve issues and workarounds to manage gaps in the process.

That informal layer now includes AI. Staff are already drafting responses, summarising cases and checking data with assistants on their own initiative, inside the steps the manual does not record. The actual workflow contains AI before anyone has designed it in, and a redesign that starts from the manual will not see it.

Over time, these practices become essential to how work gets done, even though they remain undocumented. When AI redesign relies only on the procedure manual, it misses the decisions, dependencies and handoffs that determine how the operation functions.

Deloitte’s State of AI in the Enterprise 2026 reports that 84% of organisations have not redesigned jobs or the nature of work around AI capabilities. While this does not directly measure undocumented workflows, it points to a wider challenge: organisations are introducing AI without necessarily reconsidering how work is structured.

What Happens When AI Redesign Starts From Documentation

The Agent Runs the Standard Process but Misses the Exceptions

An agent given the documented six steps will run them end to end, and it will run them well. What it cannot see is the spreadsheet where exceptions are tracked, the message thread where a disputed cost is settled, or the colleague who fills in missing supplier details before anything is submitted. It assumes consistent inputs that the real workflow produces only after informal work it was never shown.

The automation works for cases that follow the documented procedure. Everything else continues to require people, often outside the redesigned workflow and now invisible to the agent that is supposed to own it.

This is a common risk of AI workflow redesign based on formal procedures alone. Individual tasks become faster, but the operation still depends on parallel processes that the redesign has not accounted for.

When the Redesigned Workflow Does Not Fit the Work

The problem becomes more apparent when teams begin using the AI-enabled workflow. They encounter missing steps, unclear responsibilities and exceptions that the new process cannot accommodate.

Adoption suffers because employees must work around a system that does not reflect their responsibilities or the decisions they make. What was intended to simplify the work introduces additional coordination and manual intervention.

Correcting this requires revisiting the workflow design, involving the people who understand the process and addressing gaps that should have been identified earlier. Beyond the additional work, a poorly designed rollout can weaken confidence in future AI initiatives.

How to Map the Workflow Before Introducing AI

Map the Process With the People Who Perform It

Bring together the employees who handle the work, including those responsible for exceptions and approvals. Ask them to walk through recent cases rather than describe how the process is supposed to work.

For the purchase approval example, trace a request from submission to final approval. Where does missing supplier information go? Who resolves disputed costs? Which checks happen in spreadsheets or messages before the finance director sees the request?

Check the answers against the systems. Case records, ticket histories, email threads and the audit trail in the workflow tools show the sequence as it ran, and process-mining or task-mining tools can reconstruct it directly from those logs. The walk-through then validates what the data shows and explains the steps the data cannot: the judgement calls, the phone call to a supplier, the check someone does without thinking.

Compare the Actual Workflow With the Documented Process

The completed map should show the task sequence, handoffs, decision points, exceptions and tools used at each stage. Record how frequently exceptions occur and who resolves them.

Compare this with the procedure manual. A documented six-step process may contain eleven steps in practice, with several handled outside the formal system.

The differences show where the existing process depends on work that has never been formally recognised. They also reveal which steps AI could support, which decisions require human judgement and where the workflow needs redesign before automation.

The result is a working map of the operation, not simply a revised procedure manual.

What the Mapped Workflow Reveals About AI’s Role

Where an Agent Can Run the Work and Where a Person Decides

With the actual workflow mapped, the design question is which steps an agent can run and where a person decides. In the purchase approval example, an agent can assemble the request, check it against supplier records and policy, chase missing details, draft the approval case and route it, acting across several of the eleven steps rather than one. Disputed costs, unusual supplier arrangements and anything outside the agreed thresholds go to a person, and the map is where those boundaries are set, not the agent’s configuration after the fact.

The aim is to reduce routine work without automating decisions that depend on context or judgement.

BCG’s AI at Work report (June 2026), covering 11,749 workers, found that workflow redesign alongside AI deployment was associated with a 24-percentage-point higher share of workers reporting measurable improvement. The finding supports evaluating AI as part of the wider workflow rather than as an isolated tool.

From Workflow Mapping to AI Redesign

A clear view of how work operates gives leaders a basis for deciding where AI can add value, which dependencies need attention and where responsibilities may need to change.

The next step is to assess where the operation stands and identify the workflows to redesign first. On the AIVOM™ Platform, the Designer starts with two activities this article describes: map the current workflow as it actually runs, then design the future workflow with the boundaries between agent and person set deliberately.

Explore the AIVOM™ Platform.