Which generative AI use cases actually pay back for a small company?

For a company of 10 to 200 people, the generative AI use cases that pay back first are the unglamorous ones that sit inside work your people already do and already check. The visible ones, such as a website chatbot, an AI strategy or content that publishes itself, usually disappoint in the first year.

The difference is rarely the model; every mainstream tool can draft a decent email. Payback depends on whether the task is frequent, whether a person checks the output before it matters, and whether anyone knows how long the task takes today.

This article is about which use cases to pick first. A pilot that already stalled is a different problem, covered in why AI pilots stall in Swiss companies.

Which generative AI use cases pay back first for a small company?

Five use cases tend to pay back first in companies of 10 to 200 people: drafting quotes and proposals from past ones, triaging customer email with draft replies, searching internal documents in plain language, summarising calls into the CRM, and extracting fields from invoices and delivery notes. Each one is frequent, keeps a person checking the output, and replaces work someone already does by hand.

Quotes and proposals. A 60-person distributor writes similar quotes every day. Given the three closest past quotes and the customer's request, a model drafts in the house style. The sales lead still sets price and terms.

Email triage and reply drafts. Inbound mail is sorted by type and urgency, and routine questions get a draft answer that nobody sends unread. For a 40-person consumer goods brand selling online with a three-person service desk, the first gain is usually faster first response, not fewer people.

Internal knowledge search. Plain-language questions over your own handbooks, contracts and product sheets. It works when the documents are current and sensibly permissioned. If the answer lives only in someone's head, no tool finds it.

Call summaries into the CRM. The note nobody writes after a client call, filed against the contact with next steps. The value is the record that exists when that salesperson is on holiday.

Document extraction. Supplier invoices and delivery notes turned into fields for accounting or the ERP. A mature category with a measurable error rate, so the easiest of the five to baseline.

Which generative AI use cases usually disappoint?

Three kinds usually disappoint early: customer-facing chatbots with no curated knowledge base behind them, an AI strategy written before anyone has chosen a process, and fully autonomous content published without a person reading it. Each puts unchecked output in front of customers or produces a document instead of a changed task, so the cost arrives before any benefit does.

A chatbot without a maintained knowledge base answers questions on delivery, warranty and price with whatever sounds plausible, and your customer reads that as your word. An AI strategy with no process attached produces slides, not saved hours; a one-page digital strategy tied to a single process is enough. Autonomous content raises volume and lowers trust, because readers recognise generic text quickly. None is permanently wrong, only wrong as a first move.

Score it before you spend anything

Three things decide whether a use case can pay back: how often the task happens, how long it takes each time, and how much error the process can tolerate. Multiply the first two to get the hours at stake each month. Then ask who notices a wrong output, and what it costs when they do.

Take a 25-person fiduciary. Extracting fields from receipts and invoices happens daily, a few minutes each time. Drafting a client letter about a tax change happens a few times a year. The first wins comfortably, even though the second feels more like AI.

Tolerance for error is where human-in-the-loop comes in. For every use case on the payback list, a person reads the output before it leaves the building, by design rather than as a temporary crutch. The saving is the gap between writing and checking, so if checking takes as long as writing, the use case does not pay.

The use cases side by side

The ratings are typical observations from the engagements we have run, not benchmarks. Low tolerance for error calls for validation rules or a review step, not necessarily a no.

Can we paste customer data into an AI tool under the revised FADP?

Yes, provided the tool is treated like any other provider that processes personal data on your behalf. Under the revised Swiss Federal Act on Data Protection you need to know where the data is processed, whether the provider trains on it, whether a data processing agreement is in place, and whether your privacy notice covers the use. A free personal account opened by an employee usually answers none of those questions.

The FDPIC has stated that existing data protection law applies directly to AI-supported processing, so there is no AI exemption to wait for. The text of the Act is on Fedlex.

In practice the revised FADP, or nFADP, means three things for a small company. Use business plans with a data processing agreement and training on your data switched off. Know where processing happens, because transfers abroad need adequate protection or safeguards. And list what may never be pasted: health information, anything under professional secrecy, salary data, ID numbers. Our revised FADP checklist covers the wider obligations.

Run a four-week trial with a baseline

A trial without a baseline produces opinions. Before switching anything on, time the task on 20 to 30 real cases and write the numbers down.

• Week 1. Set up a business account, choose two or three people who do the task daily, and agree in writing what may not be pasted and what result means stop.
• Weeks 2–3. Use the tool on real work. Log time per case, including checking, and every correction.
• Week 4. Compare against the baseline. Extend, adjust or stop, and record why.

If time per case did not fall, or the share of outputs needing heavy rewriting did not settle, stop. A stopped trial costs four weeks; a tool rolled out on a feeling costs far more.

To turn a good result into a business case, use the structure in how to measure ROI on a digital transformation. Once drafting works, the next question is whether a system should act on its own, covered in what AI agents can realistically automate in an SME. Our AI and machine learning practice can help structure the first trial.

Related reading

• Why AI pilots stall in Swiss companies, and what gets past that
• The revised FADP: what Swiss companies actually have to do
• What can AI agents realistically automate in an SME in 2026?
• AI and machine learning consulting