Everyone uses AI. Operations still run on copy-paste.
Putting AI to work in the business, without a stable workflow
Putting AI to work in the business sounds like strategy. Day to day it often looks like this: ChatGPT or Copilot is open, and the Monday report is still copied together from five files. Everyone is “doing something with AI”. Operations barely move.
That is not a criticism of people who use the tools. Translation and correspondence really are faster there. What is missing is a workflow the company can repeat: same input, same output, a person who owns it.
A typical week in a Swiss SME makes the gap obvious. On Monday someone exports the CRM, opens two spreadsheets, downloads a PDF, and pastes until the workbook matches what management expects at 10:00. On Tuesday they use ChatGPT to translate a supplier mail and to soften a customer reply. Tuesday got shorter. Monday did not. The firm can honestly say it uses AI. It cannot honestly say a recurring workflow has changed.
What “doing something with AI” usually means in an SME
The AXA SME labour-market study, cited by DeepCloud, shows the jump: 22 percent of Swiss SMEs used AI applications in 2024, 34 percent in 2025. The most common jobs: translation (52 percent) and correspondence (47 percent). Automating a work step rose from 23 to 34 percent. Those are their figures, not our survey.
The Lucerne University of Applied Sciences KMU Digital Pulse 2026 is narrower for marketing: people who use AI mostly let it write. Text, translation, images. In marketing, 92 percent of AI users write texts. That job is crowded. Another prompt guide does not help operations.
BDO described the 2026 split this way: large firms embed AI in value creation, many SMEs stay on single apps. Their advice: a few concrete cases, owned by management, not an IT side project.
The adesso GenAI Impact Report 2026 adds measurement: the biggest hurdles they report are data quality and missing processes or governance (32 percent each). Model choice is not the top hurdle in that report. 58 percent see GenAI as a lever for efficiency and automation. What gets measured is often process efficiency (47 percent) or usage rate (45 percent). Whether someone opens ChatGPT says little about whether Monday got shorter. Those are figures from the report, not our own measurement.
ChatGPT in daily work can be the right tool for a one-off sentence. It becomes a dead end when leadership counts seats and logins, then asks why the books, the follow-ups, or the Monday file still depend on one person’s afternoon.
Three signals that usage is not moving the work
No owner
Nobody may say when the workflow is done. AI stays a tab each person opens alone.
No recurring input
Different files and questions every week. Drafting helps. A stable flow has nothing to hold on to.
No measurable output
You track logins, not shorter cycles or fewer errors.
If staff paste customer data into private chats because there is no allowed option internally, that is a different problem: shadow AI without approval. Then you lack both the workflow and the rule.
Why Copilot in a tab does not change Monday
Microsoft Copilot and ChatGPT speed up text in the window where you type. The Monday report comes from the CRM, two spreadsheets, and a PDF. The model does not log into those sources. It does not automatically summarise what is exported each week until someone defines the path: which file, which order, who checks.
Copilot in M365 can attach to documents when your firm configures it that way. Without a named workflow it still stays individual work. The operations question is not “which model” but whether the same job runs next week without the person who opened the tab today.
Pressure from the board or from clients
“What are you doing with AI?” is a fair question. A strategy slide answers it for ten minutes in a meeting. It does not replace a case with input, rhythm, and an owner. If the board wants evidence, login counts are weak. Stronger is one workflow you can name and compare before and after, even roughly, on time or error rate.
Clients sometimes ask about data protection and tools. That is not a call to launch ten pilots. It is a call to know one recurring workflow where customer data does not land in free accounts.
Why the usual first steps fail
A strategy slide for the board names the topic and leaves Monday untouched. Ten pilots at once split a small office until nothing finishes, and nobody remembers which pilot was supposed to change a named workflow. A licence for everyone so “the company has AI” speeds up drafting inside the tab. It does not log into the other four sources. A prompt library never meets the report. People may even use the prompts, which will show up as adoption, while the copy-paste stays intact.
These steps are easy to put in a meeting. They leave the gap between a chat window and a workflow someone else can run when the usual person is away.
Workshop, check, or neither
| You need … | Sensible next step |
|---|---|
| To name one recurring workflow and size it roughly | Process Check (30 minutes, one case) |
| Shared rules and allowed tools for the team | AI in business (workshop) |
| Faster drafting in a tab only | No project required; still set rules for customer data |
The workshop helps when many people need the same language and boundaries. It does not replace the question of which Monday report to automate first. That is the check.
What better looks like
Better, for a first pass, is small. One workflow. One owner who can say it is finished. One input that already arrives (an inbox, an export, a PDF). One output someone can check, on a rhythm you already have. You can tell it worked if that output takes less time or throws fewer errors, not if more people logged in.
That can still include a model. The model might draft a line, or pull a field, inside that path. The path is the point. Translation of a random mail is a fine use of ChatGPT. It is a poor definition of “AI in the company”, because it never had an owner or a repeated output to begin with.
The next step is a workflow, not a model
Filter first: does the work repeat, does it hurt, is it bounded, is the data already there? That is which process you should automate first.
If you want the same filter on your process in 30 minutes, the Process Check is the offer. What the half hour is, and what it is not, is here. The point of this page is the gap. The check is there if you want it applied to one workflow you can name.
Why does ChatGPT use not show up in operations?
In most Swiss SMEs the tool is used to translate and to draft correspondence. Those tasks get faster. A recurring office workflow still needs an owner, a repeated input, and an output you can measure. A licence and a prompt list do not create that.
Written by
Aurum Avis Labs
Builds and ships at Aurum Avis Labs. Writes here about what we learn working with founders and SMEs in the DACH region.
Related Articles
You might also be interested in these articles
Which process should an SME automate first?
Which process should an SME automate first? Four questions on frequency, pain, a clear end and data. The check is free.
What the Process Check is and what it is not
A free Process Check: a 30-minute call on one workflow, a concrete case and a package size. Not a workshop.
Shadow AI at work: your team already uses ChatGPT without rules
Private accounts, customer text in the prompt, no allow-list. Why a ban moves the chat to a phone, and the one-page minimum.