AI business cases: why job-cut calculations rarely add up

The wrong question in the business case

As soon as AI comes up in the budget round, the calculation quickly lands on one question: which job does this eliminate?

The question is understandable. Personnel costs are a large, highly visible item, and an eliminated position can be shown neatly in euros.

In mid-sized companies, it still usually leads nowhere. The teams in inside sales, quotation or customer service are already lean. A few people cover many tasks. And skilled workers are scarce anyway: in its analysis of the KfW SME Panel from February 2026 (in German), KfW, the German state-owned development bank, explicitly names the demographic decline in the labour supply as a reason to engage with AI. If you cannot fill open positions, you gain little by cutting filled ones.

Most companies are not planning any cuts at all. In the AI study by Bitkom, the German digital industry association, from September 2025 (in German), 67% of the companies surveyed expect no change in their headcount as a result of AI. Among those already using AI, the figure is 57%. In the calculation, though, job cuts are often still the yardstick.

The return comes where the same team gets more done. More quotes per week, shorter lead times, less rework, growth without additional hires. This calculation is less obvious. But it can be expressed in euros just as well, and it fits the teams you actually have.

Why the replacement calculation rarely adds up

On paper, the replacement calculation looks coherent: one job fewer, licence paid, the rest is profit. In practice, it usually breaks down at the same five points.

1. AI takes over tasks rather than whole jobs. A job in inside sales consists of many activities: sorting enquiries, preparing quotes, following up with customers, clarifying special cases with sales, classifying complaints. AI can take over some of these well. The rest remain. The result is rarely an eliminated job, more often a job with a different shape.

2. Review and exceptions stay with people. Someone has to check every AI result that goes to a customer or feeds into a quote. Then there are the cases without a template: the key account with special terms, the unclear enquiry, the faulty master data record. This residual effort is almost always missing from replacement calculations. And it falls on exactly the people whose jobs were supposedly going to disappear.

3. Knowledge leaves with the person. Which customer has which special arrangement, why an order runs differently: that is in no system. If the person leaves, the knowledge is gone, and the AI never had it. A study by the National Bureau of Economic Research (Working Paper 31161, 2023) with 5,179 customer service employees shows where AI actually has an effect. An AI assistant increased cases resolved per hour by 14% on average, and by 34% for newcomers and less skilled workers. For experienced staff, hardly any effect was measurable. Customer sentiment and employee retention improved. The authors interpret this as the tool passing on the way experienced staff work to newcomers. If that is true, the experienced staff are the source the AI learns from, and they do not belong on the list for cuts.

4. Cut decisions often do not hold. In its predictions for 2026, published in November 2025, the market research firm Forrester expects half of the layoffs justified by AI to be quietly reversed, often through offshoring or at lower salaries. This mainly concerns large companies and the US market. It still works as a warning sign for a calculation that relies solely on headcount cuts.

5. Usage moves into the shadows. AI only pays off if employees use it in their daily work and share their experience. People who fear making themselves redundant will not do that openly. They use it anyway: according to the report "The GenAI Divide: State of AI in Business 2025" by the NANDA research initiative at MIT (July 2025), employees in 90% of the companies surveyed regularly use private AI tools for work, although only 40% of the companies have officially bought corresponding subscriptions. So your team is probably already using AI. The open question is whether this happens visibly and safely. A cut logic pushes it further into the shadows.

Where the replacement calculation fails is rarely the technology. It fails on costs that only become visible after the decision. If you don't price in review effort, loss of knowledge and acceptance, you have left items out.

Replacement and augmentation compared directly

Both calculations promise a saving. They differ in where it arises, when it occurs and what happens if the assumptions do not hold.

Criterion	Replacement	Augmentation
Where the saving arises	Personnel costs of an eliminated position	More cases per person, avoided or postponed hires, less rework
When the effect occurs	Only when the position actually disappears, i.e. after rollout, handover and notice periods	As soon as the first process runs with AI in daily work
Risk	Review effort and exceptions remain, the knowledge leaves. Refilling the position later costs recruiting and onboarding	The saving fails to materialise if the time gained does not go into additional cases. Licences, support and review effort keep running
Effect on the team	The team sees AI as a threat	The team sees AI as a tool that takes work off its hands
Metric	Personnel costs saved	Additional output per person, avoided hires, shorter lead time, each convertible into euros

The replacement column looks harder because it calculates with a fixed salary total. The augmentation column can be counted just as well, as the NBER study shows: cases per person per hour.

The calculation only works with your own values. The three cost levers show how to translate the additional output into euros without counting anything twice.

The mixed team: people and AI with clearly divided roles

Behind the term "blended workforce" is a simple division of labour. The AI takes on the groundwork, drafts and routine. The person checks, decides and carries the responsibility. No quote, no customer reply and no order confirmation goes out unless a responsible person has seen the case or set the rule for it.

In daily work, this fails on one open question: who approves what? So settle two things before the first mistake happens.

First: approval levels. Not every case needs the same control. Three levels are enough to start with:

Level	What the AI does	What the person does	Example
1: Prepare	collects, sorts, summarises	uses the result as a working basis	customer history before a call-back
2: Draft	produces a draft	checks, corrects, approves	quote, reply to a complaint
3: Execute within tight limits	handles clearly defined standard cases itself	checks samples, takes over every exception	acknowledgement of receipt, delivery status information

Start with levels 1 and 2. Level 3 only comes into question once a process has run stably for weeks and you know which errors occur. Anything that touches money, liability or the customer relationship stays at level 2 permanently: prices, discounts, goodwill gestures, delivery date commitments. How tasks can be distributed across tiers is shown in the 3-tier model for AI agents.

Second: responsibilities. Every process in which AI is involved needs a responsible person. They decide which tasks the AI takes on, which level applies and when a case goes to a human. They maintain the work instructions and templates. Without this role, things grow unchecked: everyone uses the AI differently, and nobody can say why a draft looked the way it did.

What changes on Monday morning

Inside sales. Until now, the week starts with typing, chasing and searching. In the mixed team, the AI has pre-entered the orders and flagged discrepancies, such as a missing item number. Staff work through the flagged cases first and check the entries instead of typing them themselves.

Quotation. Until now, every quote is adapted by hand from an old template. Now the AI creates a draft from the enquiry and earlier quotes. The person checks prices, terms and technical feasibility and approves it. The costing stays with the person; the writing work does not.

Customer service. Until now, a service employee reads through the history before every reply. Now the AI summarises the case and suggests a reply. The service employee decides, adapts and sends. Complaints and goodwill questions stay entirely with people.

In all three roles, the activity changes and the responsibility stays where it was. Less typing and searching, more checking and deciding. For this, employees need to know how to check an AI draft and when not to trust it. That is exactly where things fall short: according to the Bitkom study, only 8% of companies offer AI training for their entire workforce, and 43% offer none at all. If you redistribute roles, plan the training at the same time.

Cost lever 1: relieve bottleneck roles

In almost every mid-sized company there are two or three roles on which revenue depends. Inside sales, which sorts enquiries. The person in quotation who pulls together prices, variants and delivery times. Order processing, which turns an order into a clean case in the ERP.

If these roles are fully loaded, enquiries pile up. More demand then leads to a longer pile, and revenue does not grow.

This is where the first lever comes in. The AI reads the enquiry, assigns it to a product, pulls in text templates and earlier quotes and presents a draft. The specialist checks, corrects and approves. Searching, retyping and copy-pasting drop away. The responsibility stays where it was.

The NBER study shows where to look first: where knowledge is missing or still being built up.

Two metrics to collect beforehand

Cases per person per week. How many enquiries, quotes or orders does a person in the bottleneck role complete on average? Count over at least four weeks so that outliers don't distort the picture.
Lead time from receipt to quote. How many working days lie between receipt and dispatch? Take the median rather than the best case.

Without these baselines, all you will have later is a feeling that things are faster.

Calculation logic for your own figures

Step	Formula	Your values
A	Cases per person per week today
B	Cases per person per week with AI (measured in the pilot)
C	Number of people in the bottleneck role
D	Additional cases per week = (B minus A) times C
E	Additional cases per year = D times working weeks

Row E is the additional capacity. It turns into euros by one of two routes, never both:

If demand today goes unserved for lack of capacity: additional quotes per year times win rate times contribution margin per order. This only applies if enquiries really do pile up. If your quotation team is not at capacity, more output per person only brings free time.
If you would have to hire without AI: then you value the same capacity as an avoided hire. That is cost lever 2.

There is no general conversion into euros for lead time. Check in your data whether quotes sent quickly are won more often. If they are, you have a second reason for the lever.

Cost lever 2: growth without new hires

The usual AI calculation compares today's costs with the costs after rollout. What is missing are costs that have not yet arisen: those of the next hire. If you absorb growth with your existing team, you save exactly this amount. It never appears in the replacement calculation, because that only lists existing positions.

What a new hire really costs

The annual salary is only the visible part. On top of it come three items that are rarely recorded properly:

Recruiting: job ads, recruitment agencies where needed, plus the time management and the department spend on selection and interviews.
Onboarding: the months in which the new person is not yet fully productive. At the same time, experienced colleagues lose the time they put into guiding them.
Vacancy: the time between the moment you need the position and the first productive working day. Enquiries wait longer, and the existing team carries the extra work.

The third item gets more expensive the harder positions are to fill. How long a vacancy lasts for you depends on region and occupation. You know this figure from your last few hires better than any statistic.

How to make the item calculable

Put avoided or postponed hires into the business case as a separate line. The values come from your own HR records:

Component	What you enter
Recruiting costs	External costs of the last comparable hire plus internal hours for selection and interviews
Onboarding costs	Months until full productivity × missing share of productivity × monthly cost of the position, plus supervision time of experienced colleagues
Vacancy costs	Average time to fill × lost contribution margin or overtime costs per month
Ongoing personnel costs	Fully loaded cost of the position for the period by which you postpone the hire

Base the calculation on postponement. Do not assume you will never hire. That is more honest and still holds if you do hire in two years. And do not count the same capacity from cost lever 1 again as revenue.

Two conditions for the calculation to hold

First, the additional volume must demonstrably run with the existing team. Without measurement before and after the pilot, the avoided hire is a claim.

Second, the relief must not turn into permanent overload. If the team only manages the growth with overtime, the costs have merely shifted. The AI has to take work away; raising the pace alone is not enough.

Even if you do hire after all, the lever helps: the strong effect on newcomers in the NBER study suggests that onboarding can get shorter.

Cost lever 3: fewer errors and less rework

Errors hardly ever appear as a separate item in a business case. They still cost working time every week, spread across many small cases. A quote goes out with the wrong item number. An order is entered in the CRM and again in the inventory management system, and the two entries differ. The customer follows up because a detail is missing from the order confirmation.

Each case is small. Together they tie up capacity and strain the customer relationship, even though they appear in no cost centre.

The AI checks and prepares, the person approves

Here the AI works as a checking layer. This is what it looks like:

Check completeness: before a quote or order confirmation goes out, the AI compares it with the enquiry and flags missing or contradictory details.
Avoid double entry: data from an email or an order form is read once and transferred to the target system as a draft. The person checks it and types nothing in again.
Prepare replies: for incoming enquiries, a draft reply with the relevant information is ready. The employee checks, adds to it and sends it.

The AI suggests and points out anomalies. Only a person may approve. This rule belongs in the process from the start, because a checking layer that produces unchecked errors creates new rework.

How to measure the effect

Three metrics are enough to start with, collected before the pilot.

Metric	What you count	Where the figure comes from
Rework rate	Share of cases that have to be corrected or redone after dispatch	Corrections in the system, sample over two to four weeks
Complaints	Complaints and queries caused by wrong or missing details	Ticketing system, inbox, notes in the CRM
Response time	Time from receipt of an enquiry to the first usable reply	Timestamps in the inbox or ticketing system

The calculation stays simple: corrections avoided times average time per correction gives the working time regained.

The side effect: knowledge gets documented

For the AI to check anything, someone has to write down what is correct. Which details belong in a complete quote? Which product combinations rule each other out? How do you answer the most common complaint?

Working on the checking layer takes this knowledge out of a few heads and puts it into a form the whole team can use. That makes you less dependent on individuals. It is hard to express in euros. Most managing directors still know exactly which people too much depends on today.

How to calculate your own case

You need five steps and your own figures.

Step 1: choose a bottleneck process

Choose a process in which cases pile up and on which revenue depends, such as preparing quotes, processing orders or handling enquiries.
Take exactly one process. Define its start and end, for example "enquiry comes in" to "quote has been sent".
Name a role that is responsible for the pilot on the business side and makes the decision at the end.
Check whether the cases are recorded at all. If they only sit in personal inboxes, sort out the recording first.

Step 2: measure the current state

Cases per person per week
Lead time from receipt to result
Rework: queries, corrections, double entry
Backlog: enquiries answered too late or not at all
Measure over several weeks and record how you measured. You will need the same method in step 5.

Step 3: set the goal as additional output

Phrase the goal as additional output from the same team: more cases per person, shorter lead time, less rework.
Translate the additional output into euros:

Benefit per year = value of the additional capacity (either additional orders × contribution margin or postponed hiring costs, never both) + rework hours saved × internal hourly rate
Estimate separately who in the team gets how much relief. Newcomers and experienced staff benefit to different degrees, as the NBER study shows. A single average for everyone is misleading.

Step 4: set the running costs honestly

Licences and usage fees: per user or by consumption, projected over twelve months.
Support: setup, adaptation to your processes, maintenance of work instructions, a fixed internal contact person. This effort continues after launch.
Review effort: the time for checking and approving AI drafts. Deduct it directly from the additional output.
Exceptions: cases the AI does not resolve cleanly land back with a person. Estimate their share cautiously.
Training: if you save here, you pay for it later with low usage.

Step 5: take stock after a fixed pilot phase

Set the duration and the criteria for stopping before the start. Do not set them along the way.
Measure the same metrics with the same method as in step 2.
Calculate the benefit from step 3 minus the running costs from step 4. Both sides belong in the same table.
Decide between three outcomes: expand, refine or stop. A stopped pilot is a result too, if it was measured properly. For the way forward, the handover checklist from pilot to routine operation helps.

Why this discipline matters: based on interviews and a survey, the MIT NANDA report concludes that 95% of the organisations studied see no measurable return despite high spending on generative AI. The sample is small and US-heavy. The point still stands: without a measured baseline, you cannot prove any return at the end.

Conclusion: the question for the next budget round

The first question decides the rest of the calculation.

If you start with "Which job disappears?", you are looking for savings that rarely materialise cleanly in mid-sized companies. If you ask "Where is capacity holding us back today?", you end up with the cases that pile up: the quote that goes out three days late, the enquiry without a reply, the order nobody can accept because the one person with the knowledge is on holiday.

You answer this second question with lead times, backlogs and open positions from your own company.

Not every bottleneck is an AI topic. Sometimes what slows you down is missing ownership, data spread across three tools, or an approval route nobody questions any more. Finding that out is a result too.

If you want to work out which bottleneck in your company is suited to a first attempt, the AI Diagnostic is a way in. It does not promise a breakthrough. It creates a basis for the decision before you commit larger budgets.

Which case in your company is currently waiting longest?

Frequently asked questions

Does augmentation mean that AI never eliminates jobs? No. When tasks change, individual positions may disappear or be reshaped. In the Bitkom survey, 20% of companies expect cuts averaging 7%, and 7% expect growth averaging 8%. Augmentation only means that your business case does not depend on a position being eliminated.

How do I take away the team's fear of job cuts? Say openly what your calculation is aiming for: more cases with the same team. And bring the employees who carry the process today into the pilot.

From what company size is a mixed team worthwhile? There is no fixed threshold. According to the KfW SME Panel, 36% of mid-sized companies with 50 or more employees use AI, compared with 19% of those with fewer than 5 employees. That says nothing about the possible benefit. Especially where knowledge sits in a few heads and vacancies take a long time to fill, a mixed team can pay off.

AI in mid-sized companies: the numbers add up through augmentation rather than job cuts