When Code Got Cheap: Why Vertical SaaS Is Losing the German Mittelstand
What Is This About?
Custom software used to be the thing small companies could not afford. Two things changed that: the cost of writing code fell, and software licences did not follow it down. Delia Hallberg, co-founder and Managing Director of Amadeni AI, spends her working days inside German small and medium-sized companies — a bakery, a youth welfare provider, an engineering firm in Berlin, a regional banking association — and she argues that the decisive question is no longer whether a company can afford to build. It is whether anyone in the building can describe how the work actually happens. This is a double feature: part one is the diagnosis, part two is the argument about what happens to software-as-a-service when the cost of code collapses.
Part 1 — E 776 · AI-Supported Custom Software for the Mittelstand
The video goes live on Thursday, September 17th, 2026
The Audio Podcast
The audio podcast goes live earlier the same day.
Subscribe to our podcasts here, or on Spotify and Apple Podcasts.
Part 2 — E 777 · When Custom Software Challenges SaaS
The video goes live on Tuesday, September 22nd, 2026
Placeholder until Tuesday :-)
The Audio Podcast
The audio podcast goes live earlier the same day.
Subscribe to our podcasts here, or on Spotify and Apple Podcasts.
Summary
The binding constraint on software in a small or medium-sized company is not budget and it is not technology. It is describability. A process nobody can explain without naming a colleague is not a process; it is a set of personal habits, and automating personal habits produces a system that stops working the week that colleague takes holiday.
Once a process is describable, the buying decision turns on one mechanical test, and it is not price. It is whether the incumbent software has an open interface. A product without an API cannot be automated, cannot be connected, and cannot be extended — so every future improvement has to be bought from the vendor, at the vendor's pace. Hallberg's clients did not leave their software because it became expensive. They left when a missing interface turned an annoyance into a ceiling. We call that point the API Threshold: the moment at which the cost of staying exceeds the cost of changing, triggered not by price but by the impossibility of building anything on top.
Horizontal software survives this. Everyone still needs email and a browser. The exposure sits in narrow, industry-specific products sold to companies whose business model falls between two industries — the firms with no proper category, which have been making a neighbouring industry's software fit for twenty years.
Who Delia Hallberg Is
Hallberg has been Managing Director of Amadeni AI since January 2026, based in Berlin, with the company registered in Eberswalde, Brandenburg. Her co-founder is Nicolai Hallberg, a software developer with more than a decade building platforms, apps and AI systems.
Her route in is unusual and it matters for the argument. She studied sociology before moving into technology and product work. She spent five years as an IT consultant at Moysies & Partner, implementing low-code systems for public-administration processes, then a year and a bit as a senior consultant at PD – Berater der öffentlichen Hand, running digitalisation projects inside public administration. Before that: product management at Pool Of Stake, and a year and a half at Humboldt-Innovation helping researchers turn ideas into businesses. Her employment history is published on LinkedIn; she does not publish an education section there, and the sociology background is her own account, given in the interview.
The sociology is the point she keeps returning to. “I don't take requirements at face value,” she says. A client who asks for a better documentation tool is describing a symptom. One youth welfare provider asked for exactly that; Amadeni built them an ERP system with voice-first interaction instead, in which documentation is one step in a chain rather than the problem.
Part 1 — The Diagnosis
The test: can anyone describe the process without naming a person?
Hallberg's clearest operating rule is also the easiest to run yourself. Ask someone to walk you through what happens when an invoice arrives. If the answer is “I get the invoice, then Sabine does this, then Thomas does that,” you are not looking at a process. You are looking at habits that happen to be held by two people.
Automate that and you have encoded a staffing arrangement. The moment Sabine is on holiday, the system does not work.
Two further signals sit alongside it. Information typed more than twice is a defect. And shadow IT — the private spreadsheet holding the real numbers, the WhatsApp group that is actually the project management tool — is the strongest signal of all. Not because people are undisciplined, but because they built a workaround for something the official system would not do. Automate around a shadow system without asking why it exists and you ship a tool nobody uses.
Hallberg's note on the German Mittelstand specifically: WhatsApp never left. The startup world cycled through Slack and moved on; the Mittelstand kept the group chat.
What a first workshop actually finds
A bakery producing its own sourdough goods — which, as she points out, makes the production process genuinely harder to model — called because the CEO had a bad feeling. A two-day workshop turned that feeling into roughly 20 systems, about 40 media breaks, and around 35 hours a week of avoidable duplicate work. Her own reaction was that the team was briefly overwhelmed by the result.
That is the argument for diagnosis before code, stated as a mechanism rather than a slogan: put AI on top of that and you automate the chaos. “A bad process with AI on top of it breaks faster, louder, and more expensively.”
The leadership signal is separate and, she says, the more dangerous one. When the managing director describes a workflow differently from the way the team described it five minutes earlier, that gap is the risk — not the software.
The triage order
Ticking time bombs first. Anything that could go off any day and threaten the business goes to the top of the roadmap and gets dealt with regardless of its return. In the bakery's case there were none.
Then the structural foundation. The bakery had no leading system, so data flowed everywhere and nowhere. You do not build on sand.
Then impact against effort. Biggest business result for the smallest investment.
The part operators tend to miss: it is fine not to fix everything. What is not fine is not knowing. “The important difference is that you are aware of them, have quantified them, and you have a plan.” A quantified backlog you have chosen not to touch is a management decision. A gut feeling is not.
The budget line that is in the wrong place
Hallberg's most transferable commercial observation is about accounting, not technology. Small and medium-sized companies treat IT as its own budget. A €25,000 project then looks like it blows the IT budget, and the conversation ends there.
Her reframe: the real comparison is not project cost against IT budget. It is project cost against the personnel cost of running the inefficient process. Not because anyone is let go — her point is the opposite, that a reactive business becomes a planning one and people get to do work that matters. But the number only makes sense on that side of the ledger.
Her thresholds are blunt. Under six months to return, it is a straightforward yes. Twelve to eighteen months is a real conversation, including the risk of not acting at all.
The worked example is the Impuls KBB case, a youth welfare provider. Development reports went from eight hours to two, the calculated saving was €21,000 a month, and the return arrived in four weeks. Hallberg is careful about what produced that, and it was not the eight-to-two step. Three things had to be true:
The client replaced its old software mid-project rather than supplementing it, once it became clear the incumbent had no API. That is what unlocked the automation potential.
The system was built for how social workers actually work. They visit children and young people at home, then sit in the car and type up documentation on a laptop. Now they record a voice message on the way back to the car; it is transcribed, rewritten into official administrative language, and dropped into a workflow. Informal chatter about document status — genuinely expensive, entirely invisible in any process map — was pulled into the software.
It was integrated with the rest of the stack. The separate HR system meant two kinds of time tracking. One interface for the social workers, data flowing behind it.
The four-week return came from the whole chain, not the headline step.
Agents replaced an IT department — and it is not vibe coding
Amadeni ran for a long time on two founders. Hallberg's account of where the agents sit is specific enough to be checkable.
Marketing and content: around 80% handed to AI, with agents holding the strategy, the content plan and the pipeline, and her doing quality control and final sign-off. Research: sector and process deep dives before a first client conversation, because the firm works across industries and regularly starts in one it does not know. Sales: a workflow that captures every piece of data gathered during the sales process, so a converted lead arrives with a project outline already written rather than starting from zero. And coding, where the gains are biggest — multiple coding agents in what she calls a coherent system replacing an entire IT department.
She is emphatic that this is not vibe coding. “It is an elaborate process designed to ensure consistent, coherent code quality.” In part two she extends it: the developer's role has become that of an IT department head, and the agents are the team of programmers being orchestrated. Joe's addition on air was that they will never complain to HR.
Her boundary test is the sharpest line in either episode: you understand your agents well enough if you could do the work yourself should they stop tomorrow. “Agents don't replace competence, they replace capacity.” The judgment, the quality bar and the thinking stay with the human — and the specific danger she names is that AI is good enough at producing plausible output that you can fool yourself for a long time before noticing you have lost the plot.
Three practices hold it together: a human in the loop at every step; agent architectures documented like software architectures, so they are scalable and not living in one founder's head; and honesty about what AI cannot do — client relationships, reading the room, and the judgment call on whether a project needs to change direction.
Her tell for operational fragility: when people talk about AI replacing a department or a person. The useful frame is amplification. The competence was there first; AI makes it scale.
Preventing hallucination by design
The Ostdeutscher Bankenverband case is the one where the stakes are external. The association supplies information its member banks base decisions on, so a fabricated figure is a reputational event.
Hallberg does not claim zero hallucinations. She claims a system that lowers the risk and can catch them. “Hope and prayers are not something that you want in your system design.”
Three design decisions: a human in the loop at every step; different, unconnected language models each given a small, concrete task, on the reasoning that a model hallucinates when it lacks information but still has to answer, so shrinking the task shrinks the pressure to invent; and traceability, so every piece of information can be followed back to its original source. The traceability is there for the humans in the loop as much as for the output — it lets them prove each item rather than trust it.
Part 2 — The SaaS Pressure Test
The API Threshold
The switching threshold, in Hallberg's formulation, is “when the cost of staying exceeds the cost of changing” — and by the time a company calls her, they have usually been past it for a while.
What is useful is what pushes them over. Across her cases it is rarely price on its own. It is the missing interface.
A Berlin engineering firm was happy with its project management software. It asked for one specific controlling feature for more than a year, did not get it, and did not get an API either. That is what ended the relationship — and the firm chose custom development proactively.
Clients arrive asking her to build an API for their existing software so they can connect it to Langdock, the German data-protection-compliant AI platform that has been growing quickly. She has to tell them she cannot build an API for somebody else's product. They want to build their own workflows, customise agents and work with their own data, and a closed product means none of it is possible.
The Impuls KBB replacement was triggered the same way.
Her summary of the mechanism: if a SaaS product has no API, nobody can add automation, and there is no flexibility. Every improvement has to come from the vendor.
The second trigger she names is a management failure rather than a vendor one: companies continuing to pay for products in which the vendor has visibly stopped investing. “Surprisingly, that is the reality for many industry solutions. Clients feel and see that.”
And the third is the workaround that became load-bearing. If the office manager leaves and nobody knows how reporting works, that is not an inconvenience. That is a business risk with a name and a notice period.
Which software is exposed and which is not
Hallberg's read, and she is deliberately narrow about it:
Least exposed. Products that serve a genuine niche and understand that niche's problems well. Products that stayed adaptable, kept open interfaces and kept listening — she has clients eight years into the same product whose vendor has been through waves of change and is still close to its customers, and she considers them fine. And, she notes separately, software in heavily regulated industries is relatively safe.
Most exposed. Products aimed at a very broad customer group, because they “fit everyone and nobody at the same time.” Products whose price rises monthly without a new feature that carries value. Narrow, industry-specific products sold into companies that sit at the intersection of two industries and have been making a neighbouring sector's software fit for years.
Not in the conversation at all. Something a vibe coder built over a weekend. “That's a prototype,” she says, and she hopes nobody is running a critical process on one. Age is not the variable; evolution is.
Her structural line, and the one worth arguing with: horizontal tools stay. Everyone needs an email service and a browser. The pressure is on the vertical layer.
Where SaaS still wins — the 70/30 rule
This is the part most vendors of custom development leave out, and Hallberg leads with it.
Her carpenter clients have nothing unusual about their business model. They are well served by products built for their trade. A standard business model with a good industry product is a case for the standard product, full stop — because a vendor who has spent years in one industry has seen problems no team designing for a single customer would think of. “That's part of being an honest IT consultancy.”
Custom software is an investment project. It takes time and resources, and it is a co-creation process over weeks: she has to learn how the company works and what is genuinely different about it. That is why every conversation runs through return on investment — the process only justifies itself with a real business case.
Her preferred shape is not competition at all. The standard product covers 70% to 80% of the important processes; the company has one process it handles differently; the product has an API; a small add-on covers the rest. “That's really my favourite way of addressing projects.”
Sovereignty stopped being sentiment in June 2026
Asked how much of the German data-protection language on her website is a real buying criterion and how much is comfort for cautious buyers, Hallberg's answer is that the honest answer changed.
In 2023 she would have called it comfort. Now clients tell her they want zero US services in their tooling and the ability to become independent of US products from one day to the next.
The trigger she names is June, and it is verifiable. On 12 June 2026 the US Commerce Secretary issued a letter requiring Anthropic to obtain licences before exporting its Mythos and Fable models to any foreign person worldwide. Anthropic determined it could not block foreign access alone on short notice and disabled the models for all users globally. A second letter on 26 June exempted certain trusted partners for Mythos 5 specifically — largely US critical-infrastructure organisations — leaving most non-US users outside. (Mayer Brown, Forbes)
Hallberg's read of it is the operational one, and it is the reason the episode matters beyond Germany: these were new models, so nothing was running on them yet. “But imagine a model being withdrawn that is already deployed.”
Hence model-agnostic architecture, which Amadeni built from the start and which she now treats as a requirement rather than an option. The target is the ability to switch models within hours. The triggers she expects are not only political: one security report or one government letter can remove a model, and commercial pressure is the other half — she does not believe AI pricing stays where it is. Her preference is European products where they fit the purpose and the quality bar, with the general principle that you do not want a monopoly and you always want somewhere to switch to.
Her framing of the whole thing: it is not verbal comfort, it is business continuity.
The organisational risk nobody budgets for
Amadeni brings an organisational development or change expert into every project, from the start, whether or not it turns out to be needed.
The reason is a real case. Two months into a transformation project, the managing director called to say the leadership was thinking about splitting up. Tensions among three managing directors had escalated because the project had made explicit who was responsible for what — and it turned out they had different understandings of that. The change expert was already in place, already knew the client, and mediated it. The project continued within a week.
Her framing is that a project aiming at a 50% or 60% efficiency increase is not an IT project. If half the team suddenly has time for different work, you are touching how people define their work. And where it had not been needed — a clear, well-led, structured team — the expert simply was not used, which she considers a perfectly good outcome.
Joe's addition on air is worth keeping: three managing directors splitting up is a far bigger risk for the company than for the project.
Who owns it
Her answer is short: the CEO. These are companies of 20 to 250 employees, and a real transformation touches processes, roles, budgets and sometimes identity.
The frontline team has to be in the room, because they know how the work actually happens. The combination she names as the marker of her best projects: strong CEO ownership of the decisions, frontline ownership of the feedback.
On day-to-day AI use, she splits it. Strategy centralised — the C-level decides where to invest and sets the data policy and the guidelines. Usage decentralised, inside those guardrails. “Set the guardrails, then let the people run.”
The belief that a real project disproved
Asked what she believed at the start of Amadeni that a client project has since falsified, her answer is about demand, not technology.
They expected a supply of nice-to-have projects: small side processes that were inconvenient, tidied up. That is not where the good work is. The best projects are the ones where the pain is real and already felt, where the team has looked for an alternative and failed to find one. If she has to convince a team that digitalisation matters, she treats it as a red flag — a signal that the company still has a learning curve to walk before the project can succeed.
What to Do on Monday
For an owner or managing director of a company between 20 and 250 people, the transferable procedure from both episodes:
Run the naming test on one process. Pick invoice handling. Ask a team member to walk you through it. Count the personal names. If there are any, that process is not ready to be automated.
Count the media breaks in that same process. Every time information is retyped or copied between systems. Twice is the limit.
List the shadow systems, without blame. The private spreadsheets, the WhatsApp groups. Each one is a specification for something the official system does not do.
Check the API on your two most important products. Not the price, not the feature roadmap — whether you can get at your own data programmatically. This is the API Threshold test, and it is the one that predicts whether you will be switching in two years.
Move the cost to the right budget line. Price the inefficiency in personnel time, then compare. Under six months to return is a yes; twelve to eighteen is a conversation about risk, including the risk of standing still.
Ask whether a standard product already exists. If your business model is ordinary and a product fits, buy it. Custom development is an investment project with a co-creation cost, not a default.
Ask who signs. If it is not the CEO, and the project touches roles, it will stall at the first conflict.
Hallberg's own closing decision rule is the one to carry out of both episodes. “Can we automate this?” is a technology question and the answer is now almost always yes. The question that decides whether the money works is: do we understand how this work should actually function?
With one caveat she adds herself, and it is the right one: do not use process-first as a reason to wait. “Look at where your team loses time every single day. That is the money you're leaving on the table.”
Selling technology into German, Austrian and Swiss companies? This audience is the one making those buying decisions — operators, founders and investors who listen for work. That is what a Startuprad.io partnership is built on. See how partnerships work.
Episode Guide
Part 1 — E 776 · AI-Supported Custom Software for the Mittelstand
Software, process or leadership — what the first hour reveals
Sociology as a diagnostic tool: symptoms versus requirements
What four people used to do, and which agents do it now
Quality control on agent output, and where the human stays
The decision rule for saying “do not build software here”
Impuls KBB: eight hours to two, €21,000 a month, four-week return
Triage with 20 systems and 40 media breaks on the table
The metric that collapses hours, errors and management firefighting into one number
The naming test, retyped information and shadow IT
Ostdeutscher Bankenverband: designing against hallucination
Part 2 — E 777 · When Custom Software Challenges SaaS
The developer as head of an IT department of agents
What breaks first in SaaS: rigidity, pricing, lock-in, contract length
Which products are most exposed, and which are not
Where standard software is structurally the better answer
The switching threshold and the missing API
Data ownership, German hosting and what changed in June 2026
Model-agnostic architecture and the hours-not-months switching target
When organisational development belongs in the room
CEO ownership, frontline feedback, centralised strategy, decentralised use
The operating belief a client project disproved
Investors and hiring at Amadeni AI
The one question to ask before spending money on AI
Related Startuprad.io Coverage
What We Could Not Verify
Startuprad.io separates what a guest reports from what we have checked at source.
Verified at source. The June 2026 US export-control action on advanced AI models, including the 12 June licensing letter, the worldwide disabling of the affected models, and the 26 June partial exemption. Amadeni AI's company details, founders, stated project range of €15,000 to €50,000, German encrypted hosting and client ownership of the delivered software, from the company's own site. Delia Hallberg's employment history, from her LinkedIn profile. Langdock's existence as a German data-protection-compliant platform with a published API product.
Reported by the guest, not independently verified. All client figures — the bakery's roughly 20 systems, 40 media breaks and 35 hours a week; Impuls KBB's eight-to-two hours, €21,000 monthly saving and four-week return; the Berlin engineering firm's year-long feature request; the leadership conflict at an unnamed client; and the internal split of work between Amadeni's founders and their agents. These are one operator's account of her own projects, given on the record. They are reported as such.
Not a claim we make. That SaaS as a category is ending. Hallberg's own argument is narrower and better: horizontal products are safe, adaptable niche products are safe, and the exposure sits in rigid vertical products without open interfaces.
Corrections and additions: partnerships@startuprad.io.
If your company sells software, services or capital into Germany, Austria and Switzerland, a Startuprad.io partnership puts you in front of the people in this conversation. Start here.
About the Author
Jörn “Joe” Menninger is the founder, host and editor-in-chief of Startuprad.io, the English-language show covering startups and venture capital in Germany, Austria and Switzerland since 2014. He spent more than 15 years in management consulting before turning full time to the show, and he records from Frankfurt am Main. Find him on LinkedIn.
Machine-readable summaries, entity data and retrieval guidance for this and every Startuprad.io episode: startuprad.io/llm.
Created with the assistance of AI.
Comments