iConference AG
AI that lands

AI needs champions in the company

An AI project can be technically complete and still change little in day-to-day work. Assume everything technical is done: AI infrastructure, compliance, data hygiene, AI connected to SharePoint, CRM and ERP. Whether anything changes after that is decided by the people in the company.

Why the fourth pillar is not a project

Four pillars carry every use of AI: processes, data, systems, compliance (AI rarely fails on the technology). Data, systems and compliance can be worked through as a project, with a specification, a deadline and an acceptance test. Processes cannot. New workflows only emerge once the people who run them every day can imagine what AI takes over in them.

Staff lack the imagination

Staff are often overwhelmed by AI. Fears and resistance play a part. Above all, though, they lack the imagination for how AI could change their own work. Anyone who knows AI only as a chat window uses it to speed up the existing work: the request passes through the same stations as before, only faster. Without this understanding, a fundamental redesign of processes is not conceivable.

Three examples to picture

Imagination grows from examples. Here are three from everyday SME life, two of them from our own company.

News without an editorial meeting. Twice a day a workflow collects news from fixed sources. Claude selects items and writes them in German and English, a second model call checks the selection, and then they go live on our website. Nobody approves anything beforehand. How this is built as a graph of nodes and edges, and which nodes need AI, is described in Graph Engineering with AI.

The framework agreement with a large corporation. An SME draws up a framework agreement with a large corporation. Management works on it with Claude: drafting clauses, checking them against its own terms and conditions, formulating variants for the negotiation. There is no fixed workflow for this. Every contract is different, and the judgement stays with management.

The supplier invoice with one confirmation. An invoice arrives by e-mail or as a scan from the mailroom. The clerk checks the extracted details against the PDF and confirms them. The case is then in the Desk, and a draft invoice is ready in the accounting system. Booking and payment are still done by a person. Fixed code reads the invoice's QR code; AI is only needed at individual nodes. This is how our Operating Desk works, in our own company too.

Train and accompany

That is why people belong in the foreground. Training means applying: slides and learning videos convey knowledge, skill only comes from working on real tasks from your own company (Nobody learns AI by watching). Accompanying means someone is there when the first application of your own is built and when it goes wrong for the first time.

Champions at the interface

Every company needs champions who understand both AI technology and the business. They get a time budget and a mandate from management. They translate processes into instructions for the AI, support their colleagues and report back where AI reaches its limits. The classic interface between IT and business has to be staffed prominently for this.

Applications for routine, agents for judgement

A redesigned process needs an application that puts it into practice every day. Champions design it together with the developers. We develop such applications with AI; in operation they run as fixed code. AI sits only where something is read, classified or written: at individual nodes and at edges that decide the next step. The same case thus takes the same path every time, and every step can be traced.

Our hypothesis: the line runs along the type of work. Staff who mostly handle recurring cases, such as supplier invoices, get an application with a fixed workflow. An agent with an open chat window would be asking too much of them: they would have to describe every case themselves, and the result would depend on how well they do it. Open-ended work with an agent like Claude suits tasks that require judgement, typically in management, knowledge work and software development.

Where to start

The first step is an assessment along the four pillars, as provided by our AI audit. In parallel, your future champions can practise on workflows from your own company in our in-house AI training.

Read next

9 November 2026

The work in the Desk, the data in your systems

Writing a quote does not require knowing how to use the accounting software. In the Operating Desk, your sales team works in a single interface. Whatever needs doing in CashCtrl and in SharePoint, the Desk does.

Read →
18 October 2026

Are AI models hosted in Switzerland good enough for day-to-day business?

A second test round with newer models hosted in Switzerland, same dossier, same questions. Again all three confirmed a wrong figure. Claude Opus found the error in every run.

Read →
30 August 2026

Graph Engineering with AI: A Case from Our Own Operations

Our AI news page has published twice a day without human approval since 26 August. Which node needs an AI, which stays deterministic, and what replaced the approval step.

Read →