AI on top of broken workflows.

Organizations are investing heavily in AI to raise productivity, speed up execution and cut operating cost. AI automates repetitive work, writes documentation, processes information and supports increasingly complex workflows. Hours of effort shrink to minutes.

One question gets skipped. What happens when AI is introduced into a workflow that was already inefficient before it arrived? Picture a company where people copy information between disconnected systems, approvals depend on long email threads, responsibilities are unclear and reliable information is hard to find. AI speeds up individual steps. It does not remove the unnecessary hand-offs, the fragmented information, the duplicated responsibilities or the slow decisions. Sometimes it makes them worse: more requests than teams can handle, incomplete information produced faster, more coordination between people and systems.

The organization looks increasingly automated and stays operationally inefficient.

At GrundMind we treat this as an AI fit problem. In Briefing 04, The AI Perception Gap, it appeared as the workflow gap: leadership sees an AI-enabled process, employees see the same hand-offs and bottlenecks as before. This briefing looks at why that happens and what to do instead.

§The assumption behind automation

Most AI business cases assume that speeding up activities improves organizational performance. That holds when the activity is necessary, well defined and part of a functioning process.

Many workflows were never designed. They accumulated. New systems arrived alongside old ones. Approval steps piled up. Teams built workarounds. Responsibilities moved and processes did not. Eventually people spend much of their time coordinating work instead of doing the work that creates value.

AI introduced into that environment tends to reproduce the structure instead of improving it. It drafts an approval request in seconds, and the request still passes through several unnecessary approval stages. It summarizes information from disconnected systems without fixing why the information is fragmented. It automates data entry into a system that no longer serves a clear purpose.

The organization gets faster at work that may not need to exist in its current form.

§Speed moves the bottleneck

Industry research shows that AI-generated speed does not automatically become faster business outcomes.

When AI speeds up one stage, constraints elsewhere become more visible, or more severe. A product team produces research, concepts and specifications far faster. If strategic approval, resource allocation, technical dependencies and go-to-market processes stay the same, delivery does not improve. The problem is often not too little AI capability. It is a mismatch between what AI accelerates and what actually limits performance.

Most organizations evaluate AI at the task level, not across the workflow, so the mismatch goes unseen.

Optimizing one activity does not improve the system it sits in.

§Five ways AI amplifies a broken workflow

§Understand the work before automating it

Good AI implementation starts with how work is actually done, not how the process document says it is done. That means the informal workarounds, repeated activities, system limits and dependencies between teams. We recommend working through six questions in order.

1. What is the intended business outcome?Every workflow should serve a defined result: a served customer, a decision, a delivered product, a met regulatory requirement.
2. Which activities create value?Some steps contribute directly. Some are necessary controls. Some exist only because of outdated practice or system limits.
3. Where does work wait?The longest delays are often between activities, not inside them.
4. Where is effort duplicated?Repeated data entry, rebuilt information and overlapping reviews point to poor integration or unclear ownership.
5. Which activities need human judgment?Some decisions require expertise, accountability, ethical consideration or context that should stay with people.
6. Which activities suit AI?Those where AI's capabilities fit the task, reliable information exists, output can be evaluated, and the surrounding process can absorb the improvement.

Only then is an automation decision an informed one.

§Eliminate, redesign, or automate

AI opportunities are usually framed as manual versus automated. There are three interventions, and we recommend considering them in this order.

The most effective intervention may be removing a step, not automating it.

This matters most where AI programmes are judged by how many activities they automate. Briefing 06, The AI Adoption Trap, looks at why that count is a poor measure of success.

§The approval bottleneck

Take a hypothetical organization that handles internal approval requests by email. An employee prepares a request, sends it to a manager, waits for questions, resubmits missing information and gets a decision after several rounds. The company introduces AI to help employees draft better requests.

Illustrative example, not an observed result.

At task level, preparation time fell by 85%. The request still follows the same route, still waits in the same inboxes and still needs chasing. The organization improved the fastest part of the process and left the main delay where it was.

A workflow assessment would look at why requests stall. It might find that no one has defined what information a request must contain, that approval authority by amount or risk is unclear, and that two of the review stages duplicate each other. The better sequence is to eliminate the duplicate stage, redesign the request as a structured form with required fields and clear approval thresholds, and only then use AI to help complete the form. Applied in that order, AI accelerates a process that can actually move faster.

§The GrundMind view

GrundMind examines AI in the context of the workflow it enters, not the task it touches. A use case is not judged by how much time it saves on one step, but by whether the end-to-end process delivers its outcome with less effort, less delay and acceptable risk.

Two diagnostics carry this briefing. AI Fit Analysis tests whether a workflow is ready for AI: whether the task suits the tool, the information is reliable, the output can be evaluated, ownership is clear and the process can absorb the gain. AI Value Leakage locates where expected benefits disappear: into queues at bottlenecks, duplicated reviews, fragmented data, supervision overhead and rework. Together they separate the processes worth automating from those that need redesign first, and those that need neither.

The result is a recommendation per workflow (eliminate, redesign or automate) with a first move and an owner, instead of a longer list of things AI could do.

§Five questions for the leadership team

1. Do we know where work waits in our key processes?If not, we cannot know whether AI is speeding up the right step.
2. Have we asked whether each automated step should exist?The cheapest automation is the step that is no longer needed.
3. Can our downstream teams absorb more output?If approval and implementation capacity is fixed, faster production just lengthens the queue.
4. Is there a reliable source of truth for the data AI uses?Automating fragmented information spreads the fragmentation.
5. Are we counting automated activities or improved outcomes?Only one of them tells us whether the investment worked.