The Process Layering Trap

Familiar Problem or Scene

The operations director has a sensible request. The team spends too much time turning customer calls into follow-up notes, task assignments, and weekly status summaries. The director wants an AI layer that can listen to the call, write the recap, assign the tasks, and push the update into the project tracker.

On paper, this sounds like the right kind of work to automate. The inputs are repetitive. The outputs are predictable. The team is tired of doing the same clerical cleanup after every meeting.

Then the first pilot runs.

The call summary is polished but misses the real disagreement. The task assignments look complete but duplicate work already owned by another team. The project tracker fills with clean new cards that do not match the way managers actually make decisions. Nobody trusts the output, so people keep their own notes on the side. The workflow is now slower than before because the team has to manage the original process, the AI layer, and the cleanup work between them.

This is the process layering trap: adding AI on top of a weak process and mistaking the new surface for a better system.

Hidden Thinking Failure

The hidden thinking failure is assuming the current workflow deserves to survive.

Most organizations approach AI by asking, "Where can we add this tool?" That question feels practical because it starts from the work already in front of them. It is also too small. It treats the existing process as a fixed object and asks AI to make that object faster.

But many workflows are not designed systems. They are sediment. A report exists because someone once needed it before a leadership meeting. A handoff exists because two departments never agreed on ownership. A spreadsheet exists because the official system was too slow or too rigid. Over time, these workarounds harden into ritual. People stop asking why the steps exist and start defending them as the way the business operates.

When AI is placed on top of that kind of workflow, it does not remove the confusion. It formalizes it. The model writes summaries for meetings that should not happen, creates tasks from decisions that were never made, and drafts updates for metrics nobody trusts.

The McKinsey 2025 State of AI work is useful here because it separates AI use from AI value. McKinsey reports that high-performing organizations are more likely to redesign workflows as part of AI adoption, and its earlier work on rewiring for gen AI found that workflow redesign had the largest effect on whether organizations saw EBIT impact from gen AI. The point is plain: value does not come from sprinkling a model across an old operating model. It comes from changing how the work moves.

Better Question or Core Distinction

The better question is not, "Which parts of this workflow can AI automate?"

The better question is, "If this work were designed around judgment from the beginning, what steps would still exist?"

That distinction matters. Automation asks which tasks can be executed by a machine. Redesign asks which decisions need to be made, by whom, with what evidence, and at what point in the process.

An outdated workflow usually contains three kinds of steps:

• Decision steps: moments where someone must interpret context, weigh tradeoffs, or accept risk.
• Translation steps: moments where information is moved from one format to another.
• Residue steps: reports, meetings, approvals, and updates that exist because an older version of the business needed them.

AI is useful for many translation steps. It can summarize, classify, draft, compare, and route. But it cannot decide which residue steps still deserve to exist. That requires human judgment. It requires someone to say, "This status report is no longer used," or, "This approval is a proxy for a risk review we never named," or, "This meeting is where political disagreement hides."

Boston Consulting Group made a related point in its 2025 AI at Work research: AI is changing jobs faster than many companies are redesigning operations to keep up. That gap is where the trap lives. Teams buy the tool, train the staff, and declare momentum, while the old process quietly absorbs the new capability and turns it into more administration.

Concrete Before/After Example

Consider a client services firm that wants to use AI to handle project status reporting.

The weak approach sounds efficient:

Summarize each project meeting, extract action items, update the project tracker, and draft a weekly client status email.

The model can do this. It will produce clean summaries and neat action lists. The problem is that the prompt preserves the whole old reporting chain. It assumes the meeting is useful, the tracker is trusted, the action items are clear, and the weekly email is the right artifact.

The better approach starts with a process audit:

Before drafting or updating anything, map the decisions this status process is supposed to support. Identify which updates inform a client decision, which updates inform internal resourcing, which updates exist only because the team has always sent them, and which missing signals create project risk. Then propose the smallest status workflow that keeps the client informed and gives the internal team enough information to act.

That is a different kind of instruction. It does not ask AI to become a faster clerk. It uses the model to expose the structure of the work so a human can redesign it.

The output may show that the weekly client email should be shorter, that internal risk notes should never be mixed into client-facing updates, and that task assignment belongs after a project lead confirms ownership. It may reveal that the meeting recap is less important than a decision log. It may also show that three tracker fields can be removed because nobody reads them.

This is where the human value sits. The model can help surface patterns, but the professional decides what the workflow is for.

Compression

AI layered onto a bad process turns confusion into faster administration.

Behavioral Takeaway

Before adding AI to an existing workflow, run a process layering audit.

Start with one workflow that currently feels slow or repetitive. Write down every step in plain English. For each step, mark it as decision, translation, or residue. Then ask three questions:

• What decision does this step support?
• Who uses the output, and what do they do with it?
• What would break if this step disappeared for two weeks?

If nobody can answer those questions, do not automate the step yet. Remove it, merge it, or redesign it. If the step supports a real decision, decide where human judgment belongs and where machine translation can help.

The goal is not a cleaner version of the same workflow. The goal is a smaller workflow with better judgment at the points that matter.

AI should not be used to preserve operational habits that no longer deserve protection. It should force the question the old process avoided: what work actually matters here?

Sources: McKinsey, The state of AI in 2025: Agents, innovation, and transformation; McKinsey, The state of AI: How organizations are rewiring to capture value; BCG, AI at Work 2025: Momentum Builds, but Gaps Remain.