Skip to content
Chat with AI Agent
Business & governance

Why Bolting AI onto Old Processes Plateaus, and What Redesign Pays Instead

Orange ITS — AI engineering team 8 min read

A 40-person firm in the Zurich area (an illustrative case) gives its sales team an AI assistant to draft customer replies. Writing time per reply drops from ten minutes to four. After a quarter, the owner checks the numbers and finds that quotes still go out in the same five days. The replies got faster. The quote didn’t. It still waits for a price check, a manager signature and a trip through the ERP.

This is the pattern behind most disappointing AI results: the tool improves one step inside a process that nobody redesigned, so the process barely moves. This article lays out what the survey evidence says about redesign, why bolt-on gains plateau, and how to choose one process to rebuild first. If you already know which process you want to start with, our process optimization service covers how we scope that work.

For related background, see AI Agents for Business: Where the ROI Actually Is on choosing candidates and The AI Last-Mile Problem on why working models stall before production.


What bolting AI onto a process looks like

A bolt-on adds an AI feature to an existing step. A chat assistant drafts the email, a summariser condenses the call, a copilot suggests the next line of code. The process around it keeps its sequence, its approvals and its owners.

Fine as a start. It needs little change management, the risk is low, and individual staff often like it. The trouble is what happens next. Task-level studies summarised in the Stanford AI Index 2026 report output gains of roughly 14 to 15% in customer support and 26% in software development. Real gains, at the level of a task. They don’t say whether a company’s cycle time, cost per case or margin changed.

The gap between task gains and business results is what this article covers.

Why the gains plateau

Four mechanisms explain most of the plateau, and each is visible in the quoting example above. None of them is a flaw in the model, and they aren’t fixed by a better one. They’re properties of the process the model was dropped into.

The bottleneck moves

If drafting takes ten minutes and approval takes two days, halving the drafting time changes almost nothing for the customer. The constraint was never the step you sped up. After a bolt-on, it sits one step downstream, and the queue there now fills faster.

Saved time gets absorbed

Minutes saved per person are spread across hundreds of small tasks. Nobody can hand them back as a cost line. Nobody does. They disappear into checking the AI output, into waiting, or into work that expands to fill the day. Without a redesigned process there’s no point at which the saving becomes a smaller team, a faster delivery or a lower price.

Old controls stay in place

Approvals, sign-offs and double entry were built for a slower, error-prone, human-only process. If they all stay, the process keeps its old cost floor. If they are removed without anything replacing them, risk goes up. Redesign is the work of deciding which control still earns its place and which one a test or a log can replace.

Ownership stays fragmented

An AI feature belongs to whoever bought it. A process belongs to nobody in particular, so nobody is accountable for the end-to-end result. The Stanford Digital Economy Lab’s study of 51 successful deployments puts it bluntly in its foreword: the difference “was never the AI model. It was always the organization.” That sample was chosen for success, so it describes what winners did and not how common they are. It still points to the same place.


What the survey evidence says about redesign

The most cited evidence comes from McKinsey’s annual State of AI survey. Three points matter for an owner.

First, financial impact is thin. In the 2025 edition, about 39% of respondents attributed any EBIT impact to AI, and most of those said under 5%. The roughly 6% classed as high performers attributed 5% or more of EBIT to AI. McKinsey’s 2026 edition, as reported by secondary coverage, shows the picture essentially unchanged: about 37% with any EBIT impact and again around 6% high performers.

Second, the high performers behave differently, and the difference is large. In 2025 they were nearly three times as likely to have fundamentally redesigned workflows, and McKinsey called workflow redesign one of the strongest contributors to meaningful impact.

Third, the limits. These are survey responses from executives, and the attribution of EBIT to AI is self-reported. The pattern is an association. A company that is already well run may both redesign more and earn more, and the survey cannot separate the two. It’s fair to say that redesign is the practice most associated with financial impact. It isn’t fair to say it guarantees it.

What this means in practice

Treat the finding as a hypothesis worth testing on your own numbers. A single redesigned process with a before-and-after measurement tells you more about your company than any survey. For the experimental evidence on how peer examples change what firms do, see how companies reorganise work around AI.

The Swiss SME picture

Swiss data shows the same split between using AI and changing the business. The AXA and Sotomo study from March 2026 surveyed 336 SMEs in German- and French-speaking Switzerland. It found that 74% use or test AI in 2026, essentially level with 71% in 2025 and well up from 55% in 2024. Across all surveyed SMEs, 35% have integrated AI deliberately and 39% are testing it. The authors conclude that AI optimises existing processes but does not fundamentally call the business model into question. The full picture of Swiss adoption is in Swiss SME AI adoption in 2026.

Read that as an open position rather than a verdict. Most Swiss SMEs have the bolt-on. Few have rebuilt a process around it, though that is our reading of the authors’ conclusion, since the survey did not ask about redesign directly. For a firm with 20 to 200 staff that moves quickly, redesign is an advantage that competitors have mostly not taken yet. The window won’t stay open.


What a redesign changes

A redesign works on five things at once, which is why a feature purchase cannot deliver it.

  • Steps. Remove steps that existed only because a person had to move data from one place to another.
  • Routing. Decide which cases run straight through, which go to a person with the work pre-assembled, and which are refused.
  • Controls. Replace blanket approvals with checks that fire on risk, such as amount, customer type or confidence.
  • Ownership. Name one person accountable for the whole process, including the AI part.
  • Measurement. Record cycle time, error rate and cost per case before and after.

The agentic workflows guide explains how an AI agent fits into this kind of flow, including where a human checkpoint belongs. Redesign doesn’t always need an agent. Sometimes it needs a simpler form, a shared inbox and a rule. The test is whether the process end to end got shorter.

How to pick the first process

Choose a process that passes these checks.

  1. It runs often enough to measure within a month.
  2. One person can be named as its owner.
  3. A large share of cases are routine.
  4. Its output can be counted: quotes sent, invoices posted, tickets closed.
  5. It is not in the middle of another change.

Quote-to-order, supplier invoice handling, support triage and employee onboarding usually qualify. Write down the baseline first, before any tool is switched on. If you can’t say how long the process takes today, you can’t show that the redesign paid.

Keep the scope small. Fix a review date. A redesign that is allowed to run without a measurement date turns into another tool that people use a bit.

A worked illustration

Take the quoting example again, with illustrative numbers. The firm sends 60 quotes a month, and each takes five working days from request to dispatch. The redesign starts by mapping where the five days go. Perhaps one day is drafting, one is a price check against the ERP, two are waiting for a manager signature, and one is waiting for the customer’s clarification. For the practical side, see mapping a workflow before you automate it.

A bolt-on attacks the first day. A redesign asks different questions. Can the price check run automatically against the price list and flag only the exceptions? Can the manager sign off only quotes above a set value or discount, and let the rest go? Can the assistant ask the customer the missing question in the first reply instead of the third? If the answers are yes, the same model that once saved six minutes on a draft now removes two or three days from the cycle. The model didn’t change. The process around it did.

Notice also what the owner can now measure: quotes per month, days to dispatch, win rate on quotes sent within two days. Those figures belong to the business, and they are the ones a board or a bank will actually ask about.


When redesign is the wrong move

Redesign has costs: time from the process owner, a period of parallel running, and some disruption. It’s the wrong move when the process is rare, when the volume is too low to repay the effort, or when the real problem is data quality or missing systems. In those cases a bolt-on tool for individuals is the sensible choice, and the redesign can wait.

It is also worth being honest about the evidence. The studies cited here come from large firms, from executive surveys or from success-selected case sets. We found no published controlled test on Swiss SMEs of 20 to 200 people. Your own pilot is that test.

Frequently asked questions

What is AI process redesign?

It means rebuilding how a piece of work gets done around what AI can do, instead of adding an AI tool to the existing steps. The redesign decides which steps disappear, which run automatically, where a person reviews, and who owns the result. A bolt-on keeps every step and speeds up one of them. The two can use the same model and still produce very different business results.

Is the evidence that redesign beats bolt-on AI strong?

It is consistent but not conclusive. McKinsey's surveys associate workflow redesign with larger EBIT impact, and the high performers are far more likely to have done it. Surveys can show an association but cannot show cause, and firms that already perform well may be the ones that can afford to redesign. Treat the finding as a strong reason to test redesign on one process and make no assumption of a guaranteed return.

Why do bolt-on AI gains stall?

Time saved on one step rarely reaches the bottom line, because the steps around it keep their old speed, queues and approvals. The bottleneck moves rather than disappears. The saved minutes get absorbed by waiting, rework or checking the output. Gains compound only when the surrounding steps, handoffs and controls are changed as well.

Should a small company redesign processes or just buy AI tools?

Do both in sequence. Tools are fine for personal tasks such as drafting or summarising. For a recurring process that touches customers, money or compliance, redesign one process first, with a named owner and a baseline you can measure. A Swiss SME with a few dozen staff can usually map and rebuild one such process in weeks, and then judge the result on its own numbers.

How do I choose the first process to redesign?

Pick a process that runs often, has a clear owner, produces a measurable output, and has a large share of routine cases. Quote-to-order, invoice handling, support triage and onboarding are common candidates. Avoid processes that depend on rare judgement or sit in the middle of a reorganisation. Measure cycle time, error rate and cost per case before changing anything.

Insights

Put these ideas to work

A 30-minute call is enough to find out whether an AI agent fits your workflow — and what it would return.