Custom-built AI agents that take on your work.

Which work is up to you: orders, quotes, suppliers, or something else entirely. Your team steers them from software we build alongside, and gets through many times the work.

Book a free callA prototype on your data in one week. €5,000 fixed.
Agents that work unattended
Tools your team uses every day
Products your customers hold
The systems underneath

What we build

Example: a wholesaler's order desk
  1. Layer 01

    Agents that work unattended

    Every order read, whether it came as an email, a PDF or a photo of a handwritten list. Items, prices and stock checked, the order entered in the ERP. 4,812 today, 110 of them held for your team.

    • Every order, whatever form it comes in
    • Each decision and its reasons on the order
    • Held for your team to review when unsure
  2. Layer 02

    Tools your team uses every day

    The queue the sales desk opens each morning: 110 orders, not 4,812. What the agent checked, what it wasn't sure about, and one button to approve.

    • Only the orders held for review
    • Every check, with its reasons
    • One button to approve
  3. Layer 03

    Products your customers hold

    The tradesperson sees when the order arrives and who approved it. Built to the standard of the best consumer apps.

    • Every step shown, and who took it
    • Ordered in the evening, delivered in the morning
    • Built like the best consumer apps
  4. Layer 04

    The systems underneath

    Evals, monitoring, integrations. What makes it dependable in production, not just in a demo.

    • Connected to the systems you run
    • Every change tested on real orders
    • Watched while it runs

What we build. Systems that run a whole process, from the agents nobody sees to the screens everyone uses.

Take the order desk of a plumbing and heating wholesaler from Freiburg, eleven branches across Baden-Württemberg, 4,800 orders a day, one sales desk that's done reviewing by nine. One system, four parts:

RunsSearch…⌘K

Order intake

Reads every order as it comes in, whether email, PDF or a photo of a handwritten list. Enters what items, price and stock plainly allow and holds the rest for a person. Every decision on record.

Running since 06:00
Read4,812since 06:00, from email, PDF and photo
Entered4,70297.7%In the ERP, for 11 branches
For a person110Waiting in Review
Per order2.1sp95 6.4s · €0.006 each
Log, newest first
EnteredKeller SanitärPDF, 14 lines · all in stock · framework pricesB-48211€2,31017:28
EnteredHeizungsbau MoserEmail, 3 lines · 2026 price listB-48210€89017:28
EnteredBader Bad & HeizungPhoto of a list, 6 lines · stock checkedB-48209€64017:27
For a personWehrle InstallationenOver the €7,500 credit limitB-48208€8,90017:27
EnteredHug HeizungstechnikCall note, 2 lines · in stockB-48207€31017:27

Agents that work unattendedEvery order read, whether it came as an email, a PDF or a photo of a handwritten list. Items, prices and stock checked, the order entered in the ERP. 4,812 today, 110 of them held for your team.

ReviewSearch…⌘K

Review

110 orders the agent wasn't sure about, oldest first.

All 110Items 46Quantity 38Stock 26
Keller Sanitär€1,240100 units, usually 10B-48177
Wehrle Installationen€8,900Over the €7,500 credit limitB-48208
Heizungsbau Moser€1,650No delivery addressB-48102
Bader Bad & Heizung€3,120Item discontinued, successor in stockB-48051
Hug Heizungstechnik€640DN 20 or DN 25, not specifiedB-47990
Gutmann Sanitär€420Not enough stock for tomorrowB-47943
B-48177

Keller Sanitär, Emmendingen

Terms
Framework 2026
Received
17:02
Lines
14
What the agent checked13 lines match the catalogue and stockPriced by the 2026 framework agreement
What it wasn't sure aboutLine 7: 100 press fittings, 22 mm. The last twelve orders had 10 each. Can't tell from the order whether it's a big job or a typo.
It proposesEnter the order with 100 units, delivery tomorrow morning.
ApproveAsk the customerDecline

Tools your team uses every dayThe queue the sales desk opens each morning: 110 orders, not 4,812. What the agent checked, what it wasn't sure about, and one button to approve.

9:41
DDreisamLM
Order B-48177

Job: Hebelstraße

Delivered€1,240Delivered to Hebelstraße 12, Emmendingen. Delivery note in the app.
What happened
  1. OrderedYou17:02
  2. ReadDreisam17:02
  3. ApprovedAnna, sales desk07:31
  4. DeliveredRoute 3, Emmendingen09:48

Products your customers holdThe tradesperson sees when the order arrives and who approved it. Built to the standard of the best consumer apps.

EvalsSearch…⌘K

Evals

Every change to the agent runs against 5,000 past orders before it reaches a customer.

Run now
Passing4,993of 5,000 orders
Accuracy99.9%↑ 0.2 ptsAgainst last release
Regressions0Any one blocks a release
Last run06:00intake-v14 · 11 min
Failed cases, last 30 nights
PassingItem matching
100%
1 failingQuantities and units
99.9%
1 failingPricing
99.8%
PassingPhotos and faxes
100%

The systems underneathEvals, monitoring, integrations. What makes it dependable in production, not just in a demo.

In-house. Thought through, designed and engineered by one team.

Dreisam · Order intake

What we build first.

Read every order, whether email, PDF or photoIn
Enter the clear ones in the ERPIn
Review queue with the agent's reasoningIn
Suggest alternatives when an item is missingLater
Chat with your orderOut

ProductWe find what's worth building, decide what to leave out and in which order. In the room with you, not from a brief.

iPhone 16Order, delivered
9:41
DDreisamLM
Order B-48177

Job: Hebelstraße

Delivered€1,240YouDelivered to Hebelstraße 12, Emmendingen. Delivery note in the app.
What happened
  1. OrderedYou17:02
  2. ReadDreisam17:02
  3. ApprovedAnna, sales desk07:31
  4. DeliveredRoute 3, Emmendingen09:48
AAnna, sales deskCustomers should always see who approved their order.
Amount138 × 44
TypographyInstrument SansMedium44100%−4%
Fill07130E100%

DesignInterfaces people love to use, in a customer's hand or running a department. What someone needs to see to trust what the agents did.

order.tsxreview.ts
1export function Order({ order }: { order: Order }) {
2  return (
3    <Screen>
4      <Label>Order {order.id}</Label>
5      <h1 className="text-[32px] tracking-[-0.035em]">{order.title}</h1>
6      <Card>
7        <Badge tone="delivered">Delivered</Badge>
8        <Amount className="text-[44px] tracking-[-0.04em]">
9          {format(order.total)}
10        </Amount>
11        <Delivery site={order.site} />
12      </Card>
13      <Timeline
14        steps={order.history}
15        // Who approved it, always shown
16        reviewer={order.reviewer}
17      />
18    </Screen>
19  )
20}

EngineeringThe agents, the pipelines they run in, the app on top and the evals that check it all, built to hold up in production at volume.

Your data. We build to the highest security standards. Tailored exactly to your needs.

Your rules

Your sign-in

Your models

Your servers

Selected work. Products we designed and engineered end to end.

  1. AI InbxCore infrastructure · aiinbx.com

    Email infrastructure, built for AI agents.

    Most email services just deliver mail. An agent also has to understand it: who answered, what they said, what to ignore. AI Inbx is the layer that does it, for anyone building agents on top.

    One thread, as the agent sees it

    1. Agent emails Dana at Northwindnew thread
    2. Marco answers instead: new address, new subjectsame thread
    3. Out of office from Danaignored
    4. Pricing.pdf, read as textattachment
    5. Agent replies, held until 9:00 their timesending hours

  2. BrandixCase study: how people steer agents · brandix.ai

    Talk to the agent, steer it with a click.

    Our study of the human–agent interface. Brandix names brands: you describe the project in chat, the agent answers with names whose domain is free. You star what fits, dismiss what doesn't, and every click steers the next round.

    One round, click by click

    1. A calm budgeting app for young familiesbrief
    2. SproutHearth.comstarred · available · $10.46/yr
    3. HomewardCents.comdismissed
    4. Like: two syllablespreference
    5. TogetherTally.comnext round · available · $10.46/yr

Pricing. Start with a €5,000 week. Decide what comes next.

  1. Week one

    5 days

    €5,000fixed

    We sit with your team, map one process as it really runs, and pick the step an agent should take first. Then we build a prototype of it on your data.

    Your teamA few hours

    By FridayWhether it works, what it takes, and your call

    Book a free call
  2. Build weeks

    5 days each

    €5,000per week

    We take the prototype towards production: your systems, evals, the screens your people steer it from. We plan the build together up front, then reshape it as we learn what works. As many weeks as it takes, no more.

    Your teamQuestions as they come up, a look when something's ready

  3. Keeping it running

    Monthly

    On request

    Once it's in production. We watch the agents at work and fix what breaks. When OpenAI, Anthropic and co. release a new model, we run it on your real cases. If it's smarter, faster or cheaper for you, we switch, after proper testing. Your system only gets better, never worse.

    Your teamTell us when something's off or in the way

    Whenever you wantAnother build week, or take it in-house

We don't quote finished systems. Week one is fixed, and whether it goes on is your call. After that we plan the build together and you pay week by week. Prices exclude VAT.

About. You work with the people who build it.

We build systems where agents read, write and decide at volume, so your people get through far more work and spend their day on the part that needs them. We're based in Freiburg, take on a few clients at a time, and the founder is there from the first call to the day it runs in production. You own everything we build, and your data stays where you want it.

Pau Kraft

Pau KraftFounder,

LinkedInX

Pau has been building software professionally for more than six years. Over the last three years he has specialised fully in AI that runs in production. One AI system he built for a client did so well it became its own startup. Handelsblatt has since written about it.