AI potential analysis: what pays off

Why most AI ideas start at the wrong end

When companies talk about AI, the list usually starts with what is in the press: a chatbot for the website, an assistant for sales, automatic quotations. These are not bad ideas. But they start from the tool, not from the work. The workflows that pay for themselves fastest are rarely spectacular. They are the tasks that come up in large numbers every day, that nobody likes doing and that still require expertise: reading documents, transferring data, sorting enquiries, adapting texts to rules.

That is why our workflow analysis does not begin with a brainstorm but with the people who do the work today. We watch, ask about volumes and times and look into the systems. The result is a list of all the workflows examined, sorted by effort today and saving tomorrow. And a column that hardly any consultancy proposal includes: "does not pay off", with reasons.

The criteria we use to fill this list are no secret. You can apply them to your own workflows.

The ten criteria we use to assess every workflow

No workflow meets all the criteria. What matters is the overall picture: a workflow with high volume, clear rules and good checkability pays off even if the data first has to be cleaned. A workflow that occurs only twelve times a year rarely pays off, no matter how elegant the technology would be.

The ten criteria we use to assess every workflow
Criterion	In favour	Against
1. Volume	Dozens of cases a day or a dataset of thousands of records	Individual cases, a few times a month
2. Input format	Free text, emails, PDFs, data sheets: a person has to read to understand	Already structured data that only has to get from A to B (conventional automation is enough for that)
3. Rule-based	Two experienced colleagues would decide most cases the same way	Every case is a judgement call, and the rules exist only in one person's head
4. Checkability	The result can be checked by a rule, a comparison or a sample	Whether the result is right only becomes clear months later
5. Cost of errors	An error is found in the check or is easy to correct	A single error has serious consequences and cannot be caught
6. System access	The systems involved have interfaces or export routes	Data exists only on paper or in a system without access
7. Data quality	Master data is maintained, mandatory fields are filled	Duplicates and sprawl: not a reason to rule it out, but a step that has to come first
8. Stability	The workflow has run the same way for years	It is being reorganised or depends on an open decision
9. Who does it today	Skilled staff who should really be doing something else	The work is also on-the-job training that nobody wants to lose
10. Measurability	Volume, time and error rate can already be counted today	Nobody knows how long it takes, and nobody can find out

Two criteria deserve an explanation. The second, input format, is where AI parts ways with conventional automation. If an order arrives from the shop as a structured record, you do not need a language model to transfer it to the ERP; an interface is enough and is more reliable. Language models play to their strengths where a person used to have to read: an enquiry in which the customer describes her problem in three sentences, or a data sheet from which dimensions and materials have to be extracted.

With the seventh, data quality, many companies give up too early. Poor data is rarely a reason not to start at all. It is a reason to start tidying up. At a spare parts dealer with around 50 employees, the CRM data contained placeholders, internal identifiers and discontinued products. Today a filter keeps them out of everything generated from that data. What this looks like as a separate service is described on the Putting data in order page.

What ends up in the "does not pay off" column

The honest column is the part of the analysis that managing directors ask about most often. Five patterns come up most often in our analyses:

Too rare. The year-end report that costs a week once a year is annoying. But the build costs more than the manual work would cost over many years.
Actually an organisational problem. If quotations sit waiting because approval is stuck with the head of sales, no AI will help. A clear rule on who deputises will.
Better without a language model. Many workflows can be handled with an interface or a rule in the system. That is cheaper, faster and more robust. We then say so in the list, even if we could build it ourselves.
Standard software fits. If there is a good product for a workflow that works exactly the way your company does, connecting is better than building. The analysis states for each workflow whether connecting or a custom tool makes more sense.
Decisions about people without human review. We do not build an AI that pre-sorts applications and sends rejections on its own. AI systems for selecting applicants count as high-risk systems under Annex III, point 4 of the EU AI Act, Regulation (EU) 2024/1689, for which extensive obligations apply from 2 December 2027 (Article 6 et seq.; deadline postponed by the Digital Omnibus in 2026). Workflows that file applications in a structured way and put them before a person for review are something else.

A line in this column is not a failure. It saves you the money you would otherwise have put into a build that would never have paid for itself.

How to estimate effort and saving yourself in advance

Before you talk to a provider, you can draw up a rough calculation for each candidate. It needs no software, only honest figures from your business:

Count the volume. How many cases a week? Do not estimate; have them counted for a week or look in the system.
Measure the time per case. From opening to completion, including follow-up questions and rework. The figure is almost always higher than the first estimate.
Work out the hours per year. Volume times minutes, divided by 60, times working weeks. At 200 cases a week at 6 minutes each, that is 20 hours a week, or around 920 hours a year over 46 working weeks.
Deduct the share for checks. A workflow never takes over 100 percent of the cases. Expect some cases to be put before a person and samples to be checked. For a first estimate, 20 to 30 percent remaining effort is a cautious assumption.
Value at full cost. Not gross salary, but salary plus employer on-costs and the cost of the workplace. And ask yourself what the freed-up time is worth if it goes into work that is currently left undone.

This calculation does not replace an analysis, but in an hour it separates the candidates that deserve a closer look from those that do not. What a build costs and how to set it against the saving is described in the article What an AI project costs a mid-sized company.

How the three days of analysis work

Our workflow analysis takes three days, on your premises or remotely. The time is divided like this:

Day 1: with the people who do the work. Conversations with clerical staff, sales, marketing, service. We watch how work actually happens, not how it appears in the organisation chart. Volumes and times are noted.
Day 2: in the systems. With your IT we look into CRM, ERP, PIM and mailboxes: are there interfaces, what does the data look like, which permissions are needed. And for each workflow we clarify which data is involved and which operating mode suits it.
Day 3: the list. Every workflow with today's effort, expected saving, fixed price for the build, data flow sketch and model recommendation. Plus the "does not pay off" column and, for each workflow, the recommendation to connect standard software or build a tool.

Your internal effort: the people who do the work need one to two hours each for a conversation and for letting us watch. IT needs half a day for a look at the systems. A project team is not necessary, but a contact with decision-making authority is. What the analysis costs and what comes afterwards is set out openly on the Approach and prices page.

"Our workflows are too specialised"

We hear this sentence in almost every first call. It is usually true, and that is precisely why an analysis pays off. Standard software and off-the-shelf AI products often fail at mid-sized companies because they assume an average workflow that exists in no company in that form. A workflow built for the individual company captures its particularities as rules.

An example: at a technical distributor with around 2,000 employees and twelve locations in Europe, the aim was to route enquiries to regional sales teams by market and country. What sounds simple had edge cases, such as a customer in one country with delivery to another. These rules had to be adjusted twice after go-live. Today every enquiry reaches the right team. The details, and what else went wrong along the way, are in the case study on the distributor.

If a workflow really is too specialised, because there is no rule, only experience in one person's head, then it goes in the honest column. That is a result too.

When employee data is involved

One criterion is missing from the table because it does not rule anything out, but it has to be considered early. If a workflow is capable of monitoring employees' behaviour or performance, the works council has a right of co-determination under section 87(1) no. 6 of the German Works Constitution Act (BetrVG). This also applies to workflows whose aim is not monitoring at all, for example when a log records who handled which enquiry and when. We therefore define in the analysis what a workflow logs. Involve the works council early. This note is not legal advice.

What this means for you

Whether AI pays off in your company is decided not by the technology but by the work. Take the ten criteria, go through the workflows that cost the most time with two or three of your clerical staff, and calculate volume times time for the best candidates. If two or three workflows are left that have volume, rules and checkability, an analysis is the sensible next step. If not, you have saved money.

Your next step: book a 30-minute call and bring your candidates. We will tell you openly which of them we think are worthwhile and which are not.

AI potential analysis: which workflows pay off with AI, and which do not