Which Processes Are Actually Worth Automating, A Swiss SME Guide
Not every process belongs on the automation list. Pick the wrong ones and you burn time, money and your team's trust. A practical filter that helps Swiss SMEs find the right candidates, and spot the risky ones.
Most SMEs automate the wrong end of the problem
When a Swiss SME puts "process automation" on the agenda, the first three weeks usually look the same: the team automates what's most visible. The inbox. The reporting. The weekly Excel refresh.
The result is mixed, for a structural reason: visible doesn't mean expensive. The truly expensive processes are the ones nobody counts, because they're spread across many heads and short moments: daily hunting for the latest version of a document, asking colleagues, double entry between ERP and CRM, cleaning up coding at month-end.
Pareto applies precisely here: roughly 20% of processes generate 80% of friction. Find those 20% and you win weeks. Automate what's visible and you win hours.
Three filters for a good automation candidate
A good automation candidate has three properties at once, drop any one and the ROI collapses.
1. Frequency
The process runs often enough for automation to pay back. Rule of thumb: at least weekly, ideally daily. A process that happens twice a year belongs in the documentation, not in the automation backlog.
2. Structure
The process has structured inputs. A QR-Bill is structured. A free-text job application is not. The more structured the input, the more reliable the automation, and the easier the human-in-the-loop check at the end.
3. Repeatability
The process follows clear rules that can be described without intuition. "If supplier in 'materials' category AND amount < CHF 1,000 → account 4000" is a rule. "We're more careful with this client on variations" isn't a rule, it's experience that stays with humans.
Three red flags, when not to automate
Just as important as the right candidates is the list of processes you shouldn't automate, because they're fragile, relationship-bound, or legally sensitive.
Judgement-dependent processes
Final dunning stages, termination decisions, performance reviews, contract negotiations, anywhere experienced human judgement makes the difference. AI can prepare here (draft, argumentation, fact research), but the decision belongs to the employee. That's not bureaucracy, that's risk management.
Relationship-bound communication
Personal condolence letters, escalations with key clients, sensitive HR conversations. Automation damages exactly what it should protect, the relationship. If a customer notices an automated reply to their letter, that's worse than answering late.
Rare and individual
Special cases that happen once a year. The temptation is huge to give every edge case its own automation. The result: a maintenance jungle that eats the original efficiency gains right back.
The five processes Swiss SMEs should start with
From practice in Swiss SMEs with 10–80 employees, five candidates that almost always score high on all three filters and deliver value quickly.
1. Accounts payable (incoming invoices)
Frequent (daily), structured (QR-Bill, IBAN, reference number), repeatable (coding rules per supplier). Classic biggest pain and biggest leverage. We costed it out here: 16 hours per month at 80 invoices, 12 of which fall away with AI preparation.
2. Expense receipts
Frequent (every employee, every month), structured (receipts with VAT, amount, date), repeatable (categorisation learns from history). AI receipt scanner with suggestion → employee confirms. Detailed logic: Scanning expenses for Swiss SMEs.
3. Knowledge search in the internal wiki
Frequent (every new hire, every stand-in), structured (wiki pages with metadata), repeatable (RAG answer with source reference). The "where do I find the SUVA template again" overhead disappears immediately.
4. CRM hygiene after customer calls
Frequent (every call), structured (notes, tasks, follow-ups), repeatable (standard fields). The assistant creates the CRM entry, the task, the follow-up suggestion, the employee confirms with one click. Prevents the CRM from rotting into a junk drawer.
5. Payment runs (Pain.001)
Weekly or fortnightly, highly structured (ISO 20022 XML), absolutely repeatable. From approved open items, the system generates the Pain.001 file that imports directly into e-banking. Manual retyping, IBAN typos, forgotten payments, all gone.
Human-in-the-loop, where the human stays, period
Even in the five candidates above, the human never disappears. Each case has a decision point where an employee approves, confirms or rejects. That's not bureaucracy, that's architecture.
- Accounts payable: four-eyes approval above CHF 1,000.
- Expenses: manager check before booking.
- CRM: the employee confirms what the assistant prepared, not the other way round.
- Payment runs: signature requirement before banking upload.
These approval points are not a weakness, they are insurance. They cost seconds, but they save the business when an automation does drift.
How hegyon.ai thinks about process automation differently
Classical RPA tools ("robotic process automation") simulate mouse and keyboard, a bot clicks through interfaces. It works, but it's fragile: every UI change at the vendor breaks the bot.
hegyon.ai takes a different path. The assistant works natively inside the business context, it knows customers, projects, invoices, tasks, wiki content. Not through clicks, but through data structures. That makes it stable, auditable and multilingual.
The four structural properties that carry this:
- Multi-provider LLM, the right model per task (OpenAI, Anthropic, Google, Mistral)
- Audit chain, every action logged tamper-resistantly
- Roles & RBAC, the assistant only sees what the employee may see
- Swiss hosting + EU AI Act guardrails, documented in the Trust Center
For a deeper take on the conceptual difference between ERP and AI Operational Assistant, see: AI Operational Assistant vs. traditional ERP systems.
From filter to rollout, a six-week pilot
Approach it soberly, and you go from "we should automate something" to two productive use cases in six weeks. Realistic plan:
- Week 1, process inventory. 60-minute workshop per function (accounting, sales, HR). Everyone writes down their five most time-consuming processes.
- Week 2, scoring. Run the three filters (frequency, structure, repeatability). Pick the top five.
- Week 3, AI policy. One page: allowed data, forbidden data, owners, logging policy. Before the pilot, not after.
- Week 4, pilot team. 3–5 people from the prioritised functions. The two strongest use cases launch.
- Weeks 5–6, pilot live. Real data, small volume. Measure: time per transaction, error rate, employee feedback.
- After 30 days, decision. Scale, adjust, or honestly stop. The worst option is "keep going because we've already invested".
How that rollout works culturally, bringing people in rather than replacing them, is in: Rolling out AI without losing your people.
Bottom line
Process automation isn't "automate everything that can be automated". It's find the right 20%, automate them cleanly, deliberately leave the difficult 80% to humans, and elevate employees from data-entry clerks to decision-makers.
For a Swiss SME that means concretely: away from the tool stack, into the integrated business context. Not "let's buy another system", but "let's have the assistant prepare invoices, expenses, knowledge search, CRM hygiene and payment runs, while we keep control".
Interested in hegyon.ai?
If you want to find out which of your processes are concretely worth automating, book a no-strings 30-minute demo. We run the three filters through your real use cases together, you see immediately where the leverage is biggest.
Swiss hosting. Swiss law. Swiss SME focus. Direct contact: hegyon.ai/en/contact.