Hundreds of duplicate clinic records in a central directory, cleaned up by hand.
Healthcare · Delivered and handed over
An AI proposes each merge, an official register checks it, and a person approves.
The problem: A central directory of healthcare providers had collected duplicates over the years, each with conflicting addresses and identifiers. Cleaning them meant comparing records by hand, checking identifiers against official registers and writing database scripts manually. One wrong merge would corrupt data that other systems depend on.
What we built: A review application that groups related records and lets an AI model propose how to merge each group. Every identifier is checked against an official register before a reviewer approves, edits or flags the proposal. Unclear cases go to manual review. The AI never touches the live database: the output is a cleanup script the client runs themselves.
Where it stands: Delivered as a working application and handed over to the client's team, who now own the code.
Campaign reports arrived as spreadsheets by email and were checked one at a time.
Media and marketing · In use with the client
Reports are read automatically, compared with the whole portfolio, and outliers raise an alarm.
The problem: Publisher reports landed in a shared mailbox as Excel attachments. Someone opened each one, compared it with similar campaigns from memory and wrote a note to the account manager. Nothing was benchmarked systematically, and nobody saw the portfolio as a whole.
What we built: An automation watches the mailbox, reads each report, benchmarks it and drafts an analysis email. A dashboard shows the portfolio by category, country, client and campaign, and alarm rules flag the campaigns that need attention.
Where it stands: In production. The daily benchmark and the alarms run unattended, operated by the client's own analytics team.
Customers struggled to build dashboards from thousands of data parameters.
Software and data · Proof of concept
Describe the dashboard in one sentence and get a valid configuration back.
The problem: The platform offers thousands of data parameters and layers. Customers found dashboards hard to configure, and solution engineers repeated the same setup work. Nobody knew whether a language model could produce valid configurations reliably enough to build into the product.
What we built: A workspace for the company's engineers. They describe a dashboard in words, choose prompt versions and models, and generate configurations side by side. Every result is validated or repaired automatically and stored with its cost, speed and a rating.
Where it stands: Two proof-of-concept rounds gave the team a working generator and the evidence to decide whether to build the feature into the product.
Teams had AI licences and curiosity, but no path from an idea to a working result.
Team enablement · Delivered several times
Hands-on sessions where people work on a real problem from their own job.
The problem: Generic tool training does not stick, because nothing built in the session belongs to anyone. Teams know they should use AI, but they do not get from there to a result.
What we built: Hands-on formats for different audiences: a prototyping workshop in which participants built a first version of their own idea, a short practical session with a small professional-services firm, and a live tour of a working agent setup for an airline's product community. The checklist and prompt library stay with the team.
Where it stands: The first of these workshops grew into an ongoing enablement partnership.
For every campaign, sales reps hunted for fitting articles and took screenshots by hand.
Media and marketing · In use with the client
The system finds and ranks fitting articles and captures the screenshots. The rep chooses.
The problem: For every campaign, sales staff searched publisher sites for a fitting article, captured desktop and mobile screenshots and put together a proposal. The same manual routine, campaign after campaign, with inconsistent results.
What we built: A service that picks up the campaign from the CRM, searches the approved publisher sites, filters out unsuitable pages and lets an AI model rank the rest against the brief. An automated browser then takes clean desktop and mobile screenshots and prepares the delivery email.
Where it stands: In use with the client. The rollout is deliberately cautious: nothing is sent automatically until a supervised test phase has passed.
Most product features shipped without an explainer video, because each one took days.
Healthcare · In use with the client
Drop in a screen recording and a brief. Out comes a branded video with narration and captions.
The problem: Product and marketing teams needed tutorial videos for every release. Each one meant an editor, a narrator, a brand review and days of turnaround, so most features never got one.
What we built: A video pipeline run by AI agents. It analyses the recording, writes a scene plan, generates German narration in the brand voice, adds branded cards, captions, cursor highlights and music, and checks the audio and the finished video before delivery.
Where it stands: In use for the client's product explainers, in German.
Support staff rewrote similar replies all day, and generic AI tools could not see the tickets.
Healthcare · Proof of concept
An AI agent reads the ticket and the knowledge base, then drafts a reply for a person to review.
The problem: A support team spent much of its day on repetitive work: similar replies, moving information between tickets and the knowledge base, documenting solutions. Generic AI chat tools could not see the tickets, and patient data made any cloud tool a hard sell.
What we built: A branded workbench running in a sandbox on the client's side. The agent reads the ticket and the relevant documentation through a standard interface, then drafts a German reply for a person to review and send. Built-in rules forbid patient data and require human review of every output.
Where it stands: Deployed for the support team's evaluation.
Estimators measured quantities from plans by hand, position by position, for days.
Construction · Proof of concept
The system proposes quantities, shows where each number comes from and flags what it is unsure about.
The problem: Estimators measure quantities from plans by hand and then price them. It takes days per project, is hard to check, and the first cost estimate arrives late. Plain AI extraction is not trusted, because it cannot show its work.
What we built: A system that reads the PDF plans, proposes walls, openings and quantities, and marks on the plan where each number comes from. Rule-based checks compare every quantity with the printed dimensions, uncertain items get a second, independent check, and the estimator validates everything in a review screen.
Where it stands: Tested blind against an engineer's own takeoff of a real house: close enough for a first estimate, with every unexplained area of the plan listed rather than hidden.
Support emails full of personal data could not go anywhere near an AI tool.
Healthcare · Proof of concept
Five independent checks remove personal data before any text reaches AI.
The problem: Support emails in healthcare contain patient names, social security and phone numbers, bank details and internal identifiers. The client wanted AI help in support, but first needed a clear answer to what exactly leaves its own systems.
What we built: A small service that takes text and returns it anonymised, in five steps: pattern checks for structured Swiss identifiers, name and place detection with a German language model, an AI model hosted in Switzerland for personal data that only shows in context, a random sample for a person to check, and a final rule-based sweep.
Where it stands: On the client's real test emails, the final check found nothing left to remove. A test report came with the handover.
After every go-live, staff typed dozens of billing rows into the ERP by hand.
Healthcare · Proof of concept
A browser extension calculates the rows, shows a preview and writes only after approval.
The problem: When a clinic went live, customer-success staff created billing rows by hand, one per doctor and module. Prices, sponsorships, fees and duplicates each had their own rules, kept in heads and spreadsheets. Mistakes showed up weeks later on invoices.
What we built: A browser extension that opens on the project page in the ERP and applies rules the client maintains in a spreadsheet. It shows exactly which rows it will create, with warnings and duplicate checks, and waits for a click. Afterwards it reads the project back and checks every new row independently.
Where it stands: Supervised test writes were verified end to end on the test system. The production rollout is documented and waiting to be scheduled.
Sales wanted to know what competitors advertise, but the facts were scattered.
Media and marketing · Proof of concept
Enter a brand and a country, get a factual snapshot of live ads and a one-page brief.
The problem: Before a client meeting, sales staff want to know what the client's competitors are advertising. The information is public but scattered, and tools that claim to know competitors' ad spend mostly guess.
What we built: A tool that reads a public ad library for a brand and a country, extracts the active ads with advertiser, start date, landing page and text, and writes a short brief for the sales conversation. The brief keeps what can be verified, the live ads, apart from what cannot, such as spend or future plans.
Where it stands: Tested on several Swiss, German and US brands. Each run produces a brief with the ads found and who is running them.
Process discovery meant interviews, hand-drawn diagrams and debates about what is right.
Our own tools · In use in client discovery
A narrated screen recording becomes a process map, every step linked to its evidence.
The problem: Discovery usually relies on interviews and hand-drawn diagrams. They miss exceptions, lose the evidence and take a workshop day to produce. Afterwards people argue about whether the diagram is right.
What we built: Someone records their screen and explains how they do the job. An agent turns the recording into a structured workflow: steps, owners, tools, delays, exceptions and open questions, each linked to the moment in the recording that shows it. A web app displays it as an interactive map with views for roles, risks, quality and automation potential.
Where it stands: Validated in real discovery calls with two clients, where every step on the map was linked to evidence. It is now part of how we start an engagement.
Ideas from the business reached developers as one sentence in a meeting.
Our own tools · Live, free to use
A guided conversation turns an idea into a brief a developer or a coding agent can build from.
The problem: Ideas for new tools reach developers as a sentence in a meeting, and weeks go by in clarification. AI coding tools make building cheap, but only once someone has written down what to build.
What we built: aiidia, a free web tool. A guided chat, in German or English, by text or voice, walks through the idea, the problem, a quick look at the market, the users, feasibility and scope. The result is a structured brief with instructions for handing it to an AI coding agent.
Where it stands: Live and free to use. It is the first step in our workshops and the entry point for companies that want their people to specify before they build.
A slide deck cannot answer the question its reader actually has.
Our own tools · Live product
Upload a PDF and share a link: readers talk to an AI tutor that knows the document.
The problem: A deck or a document cannot answer the question its reader has. The expert who could is not in the room, and cannot be there for every prospect, learner or new colleague.
What we built: VoiceDeck, a web product. A creator uploads a PDF, sets up the tutor and its voice, and shares a link. The reader has a spoken conversation with the tutor while seeing the pages; the tutor explains, answers questions and moves through the slides in the reader's language. Transcripts show creators what readers asked.
Where it stands: A live product with paying and pilot users. We use it ourselves for onboarding material and sales follow-ups.
Our knowledge lived in meetings and inboxes, and nothing built on what came before.
Our own tools · In daily use at Lailix
Every meeting, email and voice note becomes sourced knowledge our agents can use.
The problem: In a consultancy, knowledge lives in meetings and inboxes. Preparing a call means re-reading notes, and writing a proposal means remembering what was said months ago.
What we built: A company memory. Meetings are transcribed, emails and voice notes are read in, and a nightly process extracts individual insights, links them to clients, people and decisions, and compiles topic pages with their sources. Any AI agent can query it through one interface, with confidential material filtered out by default and a source attached to every answer.
Where it stands: Runs every night at Lailix. Meeting preparation, follow-up drafts and proposals all draw on it.