AI ROI: Why the Return Only Appears in the Value Stream

Why AI Is So Hard to Translate into Euros

AI has arrived in breadth, but hardly in results. According to Bitkom (September 2026), 57 percent of companies in Germany use AI, compared with roughly one in five two years ago. According to McKinsey (State of AI 2026), only about 6 percent of respondents attribute at least 5 percent of their EBIT to AI, and that share is stagnating. These are self-reported figures.

The return on AI follows the same principle as any other return on investment: benefit minus cost, divided by cost. That sounds simple. The difficulty lies in deciding what counts as benefit.

In practice, benefit is usually expressed as time saved. A task takes 10 minutes instead of 40, and multiplied by hourly rate and frequency that adds up to an annual sum. The calculation is not wrong, but it is incomplete. Saved minutes are not money as long as nobody decides what the freed-up time is used for.

Then there is the cost side. The calculation has to include more than licences: development, data preparation, training and the effort of adapting processes all belong in it. Anyone who counts only tool costs flatters the business case.

How to Measure AI ROI: The Basic Calculation and Its Limits

The basic formula is: annual value of the benefit minus total cost, divided by total cost. If the result is positive, the benefit exceeds the investment. The point at which both sides are equal is the break-even. To express the result as a percentage, multiply it by 100.

The formula is simple, the inputs are not. Three things make measurement hard in practice.

• The benefit is often an estimate. If nothing was measured before the start, you calculate with assumptions, and those can be bent in any direction.
• Indirect effects are missing from the numbers. Better quality, faster response times, happier customers or lower risks are hard to quantify, yet they noticeably affect the value.
• The time horizon is unclear. Some automations pay off quickly, others only once organisation and roles have been adapted. There is no flat deadline.

Two more points belong in every assessment. First, the initial investment: data and infrastructure often cause high costs at the beginning, before any benefit is visible. Second, scalability: a pilot should not be judged in isolation, but by whether the solution can be transferred to other processes and teams. A structured approach with fixed metrics makes both comparable.

A comparison with the alternative helps as well. Without AI, the company might have to hire additional staff or turn down orders. This comparison shows what the freed-up capacity is actually worth, and it works even when the numbers are still rough.

That is why it pays to add a second question to the formula: where in the process does the benefit arise, and who is accountable for it?

Why Time Saved on the Task Rarely Reaches the Result

The central thesis from the online dialogue "AI Adoption in the Mittelstand" is that time saved on the task rarely reaches the result. The reason lies in how most companies are organised.

A process consists of working time and waiting time. Working time is when someone is actually processing something. Waiting time arises when a case sits in an inbox, or waits for an approval, a query or the next department. AI acts on working time first. Waiting time stays untouched. The whitepaper states that in typical office processes, often less than 10 percent of lead time is actual processing.

Anyone who only speeds up the activity gains hours. Anyone who redesigns the process gains days. McKinsey's analysis shows the same: almost three quarters of the high performers have fundamentally redesigned their workflows with AI, compared with only a quarter of the other respondents.

This explains why many AI projects show no added value on the balance sheet despite visible tool successes. The single task gets faster, but the customer still waits just as long.

From Use Case to Value Stream

A value stream is the end-to-end flow from a trigger to a result, for example from a customer request to payment received. It does not end at the department boundary. That is why many initiatives fail as a collection of single use cases: each department optimises its own slice and meets its own metric, and nobody owns the whole.

Anyone who takes the return seriously changes perspective. Instead of asking where AI makes a task faster, they ask where time and money are lost in the process. This applies to simple writing assistants just as much as to AI agents that handle whole sub-steps on their own. Our whitepaper on agentic AI describes how such agents turn from tools into digital colleagues. They only deliver value where the process has been prepared for them.

The whitepaper Return on AI: Why AI Only Pays Off in the Value Stream goes deeper into the topic. This article shows the basic ideas.

The Four Maturity Levels: Where the Return Begins

Which step comes next depends on the maturity a company has reached. At the start, employees and individual departments use AI on their own initiative, and licenses and usage rates are measured. Next, AI works productively in individual process steps, but the efficiency gains stay within one department and the waiting time moves on to the next. Only when an entire process is redesigned, and finally several processes are connected into one continuous value stream, is the company steered by lead time, margin and cash.

The core message: the return begins with the jump from selective optimization to a redesigned process. It gets expensive when a company stays at a level whose costs it already bears without reaching the benefit of the next.

The jump usually fails because of the organization: the workflow is new, but the responsibilities are still the old ones. That is why it needs a role that mostly did not exist before, the value stream owner. The role leads a value stream from start to finish, with authority across department boundaries and a metric it is measured on. Ideally, a business owner and an IT owner share this task.

Four Steps to a Reliable Business Case

The path from time saved to a real result can be described in four steps. The graphic below gives the overview, and here is the core idea of each step.

1. Choose: one value stream instead of ten use cases

Start with a single value stream that has a clear lever on results and an owner. Without someone who can decide across departments, every selection turns back into an IT project.

2. Measure: baseline first, build second

Measure today's situation, split into working time and waiting time. Without this starting point, you can neither calculate a break-even later nor prove success.

3. Redesign: simplify first, then automate, then orchestrate

The workflow is rethought starting from the desired result. Often the first effect comes from an organizational decision, such as simpler approvals, and not from a model.

4. Capture: calculate in win rate and cash

The business case does not count saved hours but results such as win rate, margin and cash. This requires deciding in advance who books the benefit and what the freed-up capacity is used for.

Return on AI

The return begins with the jump to level 3

Four maturity levels, four steps. What matters is the level at which you redesign the process.

Where the return begins

1. Level 1FragmentedSingle workplaces and pilots
2. Level 2Optimised in placesSingle departments
3. Level 3Process-orientedOne process end to end
4. Level 4Value-stream-orientedConnected processes up to payment

Four steps to the return

1. 1ChooseOne value stream with a clear lever and an owner instead of ten use cases.
2. 2MeasureBaseline before the build, split into working and waiting time.
3. 3RedesignSimplify first, then automate, then orchestrate.
4. 4CaptureCalculate the business case in win rate and cash, reallocate capacity.

An Example from Mechanical Engineering

A typical case from mechanical engineering shows how this can look in practice. Several value streams are candidates, and the biggest lever often sits in the flow from order to payment. That is exactly where an owner is frequently missing. It is then smarter to start with a smaller flow where someone feels responsible and the data sits in the ERP, such as the path from inquiry to quote.

The measurement usually shows that only a small part of the lead time is real work. The rest of the time the inquiry sits in a mailbox, waits for a technical question or is stuck in an approval loop. Anyone who only speeds up the calculation hardly shortens the lead time. The first effect therefore often comes from a decision, such as a higher approval limit, and only then from AI.

In the end, what counts is what happens with the freed-up capacity. If it is reassigned concretely, for example to additional quotes, the win rate rises. If it stays unused, the return stays theoretical.

Where the Return Shows Up Most Readily

The highest potential lies in the core business: according to BCG (2025), 70 percent of the AI potential lies in core functions such as sales and marketing, production, supply chain and pricing. The better return, however, often comes from automation in the back office, for example in quote creation, invoice checking or reporting. Processes there tend to be clearly structured, approvals numerous and waiting times long. The AI ROI is often especially clear there. It pays to count the waiting times first and only then talk about automation.

Similar patterns appear in many industries. In banks and insurers, the coordination between the business unit and compliance often determines lead time. In customer service, queries between the service team and the specialist department prolong handling. In production and supply chains, information flows between purchasing, planning and logistics, and every handover costs time.

Which Metrics Support the Return on AI

A reliable AI ROI needs few but honest metrics. A mix of four perspectives has proven useful: productivity, cost, revenue impact and quality. A single number rarely captures the return.

• Lead time: How long does the process take from trigger to result, split into working and waiting time?
• Win rate and revenue: Does the higher speed lead to more won orders and thus more contribution margin?
• Capacity: What is the freed-up time used for, and who decided that?
• Total cost: Besides licences, are development, data, training and process change included?
• Quality and risks: Do error rates and rework fall, and do new risks arise, for example from unchecked results?

Compare against the right starting point. Anyone calculating the break-even of a project needs the baseline before the start, not an estimate after the fact. Time horizons should also be realistic: a value stream that changes organisation and roles takes more than a pilot month, even if the first measurable effect should appear early.

Typical Pitfalls

Three patterns come up again and again in AI projects. The first is starting with the tool: a solution is bought first, and a use is sought afterwards. The reverse order works better: process, data and responsibilities before tool selection.

The second pattern is islands. Every department tests its own tools, but nobody connects the results into one process. AI initiatives that start in parallel without coordination also raise the number of automated steps, but not automatically the value. That produces many small improvements that do not add up on the balance sheet.

The third pattern is the missing decision. The time saving is celebrated, but nobody decides what happens with the time gained. Without that decision, the effect remains a figure in the project report.

There is also a side effect: where there is no common line, employees often use their own tools outside the official systems, known as shadow AI. That makes measurement harder, because nobody knows which usage contributes to the result at all. The same applies to adoption in daily work: if users do not use the new tools because rules are missing or tasks remain unclear, there is no effect, however good the technology. Clear ground rules before rollout increase acceptance and make usage measurable.

Who Owns the Value Stream

A value stream needs people who own it. That is the real staffing question behind AI ROI. Where a company gets stuck between two levels, CTG Consulting's experience is that almost always one of three things is missing: people who can work confidently with AI and rethink processes, a role that owns the value stream as a whole, or rules on what the AI may decide on its own.

For the jump from level 2 to 3, these are above all the value stream owner and an AI programme lead who steers the portfolio by business value and ends pilots without a business case. According to McKinsey, high performers are twice as likely to report that their top management visibly champions AI initiatives, and twice as likely to report that they have fixed procedures to measure the impact.

Especially in mid-sized companies these roles are often missing. The question then is whether to develop them internally, fill them with permanent hires or bring them in for the duration of the transformation as an interim solution. The answer depends on whether the value stream is a lasting part of the core business or a time-limited transformation.

With AI agents, solid governance becomes more urgent. It answers four questions in writing: who decides on the priority of use cases? Which decisions may an AI system make on its own? Who is liable if it decides wrongly? Which metric measures the contribution, before and after?

Success Factors in Implementation

Whether a value stream leads to results depends on a few factors, whether the matter is a simple automation or a complex AI implementation.

• Data quality: Data must be complete, current and accessible to those involved. Weak data limits every solution.
• Process integration: The AI has to work where the case runs, not in a separate tool that requires extra steps.
• Strategic alignment: Every initiative needs a link to a business goal, not just to a technical possibility.
• Human in the loop: Where results affect customers or decisions, people review and approve until quality is proven.
• Steering models and costs: Anyone tied to a single vendor loses room to manoeuvre. Models and running costs need active steering.

Innovation and change need this framework too. New approaches often emerge outside established IT roles, and they still need a place in the organisation so that experiments become everyday work. Creating that place early prevents good approaches from getting stuck in pilot status.

Frequently Asked Questions

How long does it take for AI investments to pay off?

There is no universal deadline. It depends on whether only one task is automated or a whole process is rebuilt. Plan the break-even per value stream, based on the measured baseline, rather than per tool.

Where should companies start?

With a single value stream that has a clear lever on results, usable data and an owner. A first measurable effect should be possible within four to six weeks.

Is saved working time enough as the benefit in a business case?

No, it is only the starting point. Time saved becomes a result only when the freed-up capacity is put to a specific use and the process as a whole runs faster. So calculate with win rate, lead time, revenue and contribution margin.

What role does IT play?

IT provides data, interfaces and technology, but cannot own the process alone. What matters is an owner who brings the business unit and IT together, ideally as a pair of business owner and IT owner.

How can indirect effects such as quality or customer satisfaction be assessed?

They rarely can be quantified exactly, but they should not be missing. Decide in advance how you will recognise the effect, for example through fewer complaints, faster response times or less rework, and record the starting point. This lets you support intangible benefits as a tendency without squeezing them into a seemingly precise number. Only what you can prove belongs in the ROI calculation itself. The rest serves as a supporting argument.

Which Value Stream Is Yours?

Which process holds the biggest lever for you and who can own it can usually be settled in a few conversations. If the right role is missing, we support selection and placement. The whitepaper "Return on AI: Why AI Only Pays Off in the Value Stream" is available as a compact summary to download.

Your next step

From time saved to results

In a non-binding conversation we place your process and show where the first lever sits.

Book an intro callRead the whitepaper