The Cloud Group logo
Contact
logo

AI doesn't save costs if you don't redesign the work.

October 14, 2026

Artificial intelligence doesn't reduce costs on its own; it reduces the effort required for certain tasks. This freed-up effort only translates into savings if someone redesigns how the work is done around the new capability. Without this redesign, the time saved is lost, and the bottom line shows no improvement.

McKinsey measured this precisely in its 2025 global survey: the factor most strongly correlated with EBIT impact is having fundamentally redesigned workflows, and only 211% of organizations that adopted generative AI had done so. High-performing companies are about three times more likely to have undertaken this.

It is, by far, the most actionable conclusion of the entire cycle.

Why does freed-up time evaporate?

A business process is not a sum of independent tasks: it is a sequence with waiting periods, coordination, and validations between them. Speeding up one task within that sequence does not speed up the sequence as a whole, because the bottleneck is usually elsewhere.

Three specific mechanisms eliminate savings:

The subsequent wait absorbs the improvement. If drafting the report goes from two hours to twenty minutes but the review still takes three days, the cycle time hardly changes.

Time is fragmented. Twenty minutes saved at five different times of the day do not become twenty productive minutes: they become five breaks.

The job is expanding. If writing is cheap, more writing gets done. Reports appear that weren't done before because they weren't profitable, and the overall workload doesn't decrease.

None of the three is a failure of the technology. All three are the result of improving a task without changing the process.

What does it really mean to redesign?

Level

What changes

Cost impact

Assist

People do the same thing, just faster

Perception of agility, with no measurable effect

Automate the task

The system performs one task

Partial savings, absorbed by waiting times

Redesign the flow

Steps, validations, and waiting times disappear

Change in cycle time and cost

The difference between the second and third rows is what determines whether there's a return. And the third row requires something the second doesn't: making decisions about how people work, which makes this a management project and not a technology project.

The four questions that define a real redesign:

  1. What validation exists simply because there was no way to verify it at the source before? Many revisions are relics of a limitation that no longer exists.
  2. What information is manually re-entered? Each manual copy between systems is a step that can disappear.
  3. What do you expect in between steps and why? Most of the cycle time is usually spent waiting, not working.
  4. What decisions could be made sooner, with the information that is now available? Making a decision in advance usually eliminates two subsequent steps.

The real obstacle is not technical.

Redesigning a workflow means someone has to stop doing what they were doing. Even if no positions are eliminated—and in most cases they aren't—responsibilities do change, and that generates legitimate resistance.

Three conditions make the redesign possible:

A person in charge with authority over the entire process. If the workflow passes through three departments and the person in charge only has authority over one, the redesign will get stuck at the boundaries. That's why we insist that Every project needs a business owner [internal link].

A clear destination for the freed-up capacity. Absorb growth, improve service, reduce lead times. When no one says what will be done with the time saved, the organization assumes the worst and defends itself.

Participation of the person who carries out the process. Not because of political correctness: because only they know the exceptions and the unwritten rules, and without that information the redesign fails in production.

How to handle the operation without stopping

A complete process redesign is a large and risky project. The working version is limited.

  1. Choose a process with recognized volume and pain. Just one.
  2. Measure the baseline for two or three weeks. Cycle time, number of hands, waiting times, corrections.
  3. Map the actual process, with exceptions. What really happens, not what the manual says.
  4. Redesign on paper and validate with the person who implements it. Before we touch on technology.
  5. Implement in stages, measuring as you go. And compare against the baseline.

This order has a decisive advantage: steps 2, 3, and 4 produce value even without implementing any new technology. It's common to discover improvements in step 3 that can be implemented within a week without any new software.

The question that sets the tone for the conversation

When faced with any proposal to reduce costs using AI, one question sets the whole picture: What will stop being done?

If there's a concrete answer—this validation disappears, this report is no longer generated, this data re-entry is eliminated—then there's a redesign and there will be savings. If the answer is that everything will continue to be done but faster, then there's assistance, which improves the work experience but doesn't change the outcome.

Both options could be worthwhile. Only one of them can be presented as a business case.

Frequently Asked Questions

Why doesn't AI automatically reduce costs?

Because it reduces the effort required for specific tasks, not the overall process time. The savings evaporate in three ways: subsequent waiting times absorb the improvement, the time freed up is fragmented into breaks, and the workload expands because what is made cheaper is done more often.

The goal is to eliminate steps, validations, and waiting times, not just to speed them up. This means removing validations that existed due to outdated limitations, eliminating manual data re-entry, reducing waiting times between steps, and accelerating decisions that can now be made earlier.

According to McKinsey (2025), only 211% of organizations that adopted generative AI had fundamentally redesigned any workflows, despite this being the factor most strongly correlated with EBIT impact. High-performing companies are about three times more likely to have done so.

It's not a technical issue: it's simply that someone stops doing what they were doing. Three conditions are necessary: a person in charge with authority over the entire process, a clear destination for the freed-up capacity, and the participation of those who execute the process, as they are the ones who know the exceptions and the unwritten rules.

Choosing a single process with recognized volume and pain, measuring the baseline for two or three weeks, mapping the actual process with its exceptions, redesigning on paper with those who execute it, and implementing in parts against the measured baseline.

What will stop being done? If there's a concrete answer—a validation process that disappears, a report that's no longer produced, a data reintroduction that's eliminated—then there's redesign and savings. If everything continues to be done but faster, then there's assistance.

Do you really want to reduce costs, not just go faster? We measure the baseline, map the actual process, and redesign the flow before proposing technology. Two hours of analysis, no obligation. Let's talk →

Data sovereignty and control of business information
Software engineering as a competitive advantage for companies
Discover how to apply FinOps to Artificial Intelligence to control tokens, infrastructure, models and agents, measure return on investment and avoid difficult-to-sustain technology costs.