The AI product can fail by not selling — and by selling.

The company

Klarschicht sells software to German mid-market companies. 135 subscriptions, 50 enterprise accounts, and a product that took eight years to make boring.

Boring is the business model. Renewals arrive. The sales cycle is understood. Nobody has had a bad quarter in a while.

Then the quotes started coming in lower.

Not from the incumbents — from four-person shops shipping in a quarter what used to take a roadmap, because they let a model write most of it. Klarschicht's product is still better. Their customers still renew. Nothing is on fire.

But the floor under the price is moving, and everyone in the building can see which way. And the fund that led at year four is closer to selling than buying — it would rather take an AI company to market than a software company with good renewals.

So they did what every SaaS company is doing this year: an AI add-on, bolted onto the base they already own. Usage-priced. The customer pays for what they consume.

Which makes it a different animal from everything else Klarschicht sells. Their software has the economics of software — build it once, sell it two hundred times, and the two hundredth customer costs roughly what the first one did, which is nothing. The add-on does not work like that. Every unit consumed burns compute someone has to pay for. Every account switched on is paid for in consultant days long before it returns a cent. For the first time in eight years, revenue and cost travel together.

And the pivot is already in the cost base. The teams were hired, the roadmap was re-cut, the year was planned around it. That spend lands every month whether the add-on attaches or not.

There is also a convertible loan in the works. Not because anyone thinks they need it. It is the sensible thing to do — a buffer, cheap insurance, the kind of decision a good CFO makes on a quiet Tuesday while the numbers still look fine. Sign it, park it, forget it.

Hold that thought.

What the board wants to know

Which is where the questions start.

• Do we actually need the loan, or is it just tidy?
• How early would we have to raise?
• What is the cash floor we do not go below — and at what number do we stop spending?
• How much do renewals actually matter?
• How much can inference cost before this stops working?
• Are we charging enough?
• Does onboarding have to get cheaper, or shorter, or both?
• And: how fast can we sell this thing?

Eight of those have the answer you would guess.

The last one does not.

How you answer a question like that

You could build a spreadsheet. Somebody at Klarschicht already had: three columns wide, base, bull and bear. Three numbers, picked by the person who built the model, each one presented with a straight face.

That is not what we did.

We rebuilt Klarschicht as a working machine — the subscriptions, the enterprise book, the add-on, the payroll, the cash account, every rule about who pays whom and when. Then we took every assumption that is not actually a fact and replaced it with a range. How many accounts attach in a month. What inference costs. How many customers renew. Whether the loan lands at all.

Then we shook it. A thousand times.

Each run draws its own values out of those ranges and plays the company forward, day by day, to 31 May 2028. A thousand futures, every one of them plausible, not one of them the official one.

And a thousand futures can be searched. You stop asking what will happen, which nobody can answer, and start asking questions that have answers. In how many of these futures does the company run out of cash? In those futures, what was different? Which assumptions separate the ones that made it from the ones that did not?

That is the whole move. A forecast hands you a number. A thousand futures tell you which assumptions you are actually betting on. It is conditional on the model, and a human still has to read it. How Brenwick examines a company walks through the method in full.

So. Back to the board's questions.

The question nobody worried about

Start with the loan, because it was the least interesting thing on the list. It was going to be signed. It was prudent. It was, in the language of the board pack, a buffer.

Across the thousand futures, 220 of them ran out of money — a month-end with the cash account below zero before we get to 31 May 2028. Roughly one in five. Bad, but survivable-sounding. The kind of number you put on a slide with a mitigation plan beside it.

Then split those futures by one thing: whether the loan landed.

Where it landed, 15% ran out of cash. Where it did not, 80%.

That is not a buffer. That is the hinge. In the futures where the loan arrives, Klarschicht is a company managing a risk. In the futures where it does not, Klarschicht is a company with a four-in-five chance of running out of money, holding an AI product it has already paid for.

Nobody had this as their first question. It was the first answer.

Renewals come second, and a distant second — but they are the only other input that pulls the survivors apart from the rest with any real force. Which is worth saying plainly to a board about to spend all of its attention on the new product: the two things that decide whether you make it are the loan you thought was optional and the customers you already have.

Neither of them is the AI.

The failure everyone can name

Now the AI. Start with the failure the board already has a slide for: it does not sell.

The teams are hired. The cloud contract is signed. Payroll lands every month regardless. If attach stays thin, none of that comes back — the new product never earns out the bet that created it, and the company ends up smaller and poorer for having tried.

That failure is real, and it happens in plenty of these futures. It is also the one failure the board does not need us for. They can feel it coming. They have a name for it. They can tell you roughly which quarter it would become undeniable.

So we asked the other question. Not what happens if nobody buys it.

What happens if everybody does.

How fast can we sell it

This was the last question on the list, and the one they expected the least trouble from. Faster is better. Faster has always been better. Every instinct in an eight-year-old software company points the same way: get it attached, get the logos, get the usage curve into the next deck.

Across the thousand futures, 538 of them — more than half — closed 31 May 2028 with blended gross margin below 86%.

That number is the floor the plan is built on. Below it the company still functions. It just is not the company in the pitch.

Then sort those futures by how fast the add-on attached.

Where attach stayed near 1% a month, the floor was never in danger — not one future broke it. At 4%, roughly what the plan assumes, it broke in 63% of them. At 9%, it broke in all of them.

That is the article. The better the AI product sold, the more reliably the company slid under its own margin floor. Not because selling is bad — because every unit consumed burns compute, every account switched on is paid for months before it pays back, and the add-on simply does not have the economics the rest of the business has.

Inference cost pushes the same way. Where compute ran cheap, the floor mostly held. Where it ran expensive, it mostly did not. Price and cost are walking toward each other, and only one of those two numbers is set in Klarschicht's own pricing meeting.

None of this is an experiment. These are futures sorted by what happened inside them, not levers we pulled to see what moved.

So: the AI product can fail by not selling. And it can fail by selling. Those are not two risks to weigh against each other. They are one product seen from both ends. Sell too little and the pivot never earns out. Sell enough and the margin goes.

Which changes the job. The board is no longer there to push the thing as hard as it will go — it is there to steer between two walls: a pivot that never earns out, and one that earns out at a margin nobody would underwrite. The examination can show them both walls. It cannot drive.

What we would do

Secure the convertible loan before committed pivot spending keeps landing. Protect renewal: it is the second runway dependency once the CLA is set aside. Put a ceiling on AI inference cost, and reprice the add-on if that cost stays high. Shorten onboarding and cut what it costs. Stage the add-on rollout instead of attaching everyone the model will take.

Hold at least nine months of forward runway. Treat the next two rungs as escalation, not as commentary.

1. Nine months Target. Below that, start financing and preparation.
2. Six months Freeze new commitments. Move to weekly 13-week cash control.
3. Three months Crisis actions. Do not wait for the account to print zero.

These are Brenwick's recommendations, not results we have already tested. A later intervention study will put them back inside the model.

The exact reports

The complete report set is not published here. Request the exact reports on a confidential diligence call.

Book a confidential diligence call