Service · AI agents and automation
AI agents for SMEs: they do the groundwork, you stay in control
An AI agent takes on one repetitive task: sorting email, reading documents, following up, preparing a report. It works inside the tools you already use, and your team approves the result.
- Human approval by default
- Model chosen to suit your data
- Swiss hosting available
What an AI agent does for a firm of 5 to 50 people
An AI agent is a program that receives a task, reads the information it needs, uses your tools and prepares a result: an email labelled, an invoice entered, a reply drafted. It is part of your process and gets to work every time an email or a document arrives.
In a firm of 5 to 50 people, this work tends to land on people with better things to do: whoever sorts the info@ inbox, the accountant retyping invoices, the project manager chasing quotes. Each agent takes on a single, well-defined task, which makes it easy to test and to check.
We build these agents with language models, n8n (an automation tool that can be self-hosted in Switzerland) and Python code. We connect the tools you already use (email, Microsoft 365 or Google Workspace, your business software) through their API, an export or email. You don’t have to change software.
An agent doesn’t decide for you: it prepares drafts, flags what it isn’t sure about, and a person approves. Nothing is sent automatically unless you choose that, case by case. See also our guide What is an AI agent?
The problem
What gets in the way today
Hours of repetitive work
Sorting email, retyping invoices, chasing clients: the same tasks, every day.
Everyone uses ChatGPT their own way
No shared rules, doubts about which data may go into it, and a way of working that hasn’t changed.
Tools that don’t talk to each other
Email, accounting software, spreadsheets: information is copied across by hand.
Interactive demo
Ask the internal assistant of a fictional SME
It knows a fictional company’s staff handbook, procedures and price list. It answers from those documents, cites the paragraph it used and says plainly when it doesn’t know.
Internal assistant
Answers your team from your own documents, cites the source and says when it doesn’t know.
- 1 Ask a question
- 2 Watch it search
- 3 Check the source
Boisvif assistant Answers from 6 internal documents
Boisvif assistant
Hello! I answer the team’s questions using Menuiserie Boisvif SA’s internal documents: staff regulations, expenses, safety, warranty, onboarding and prices. I always cite my source, and if the answer isn’t there, I tell you.
Pick a question below, or ask your own.
Suggested questions
Agent log
Knowledge base
Menuiserie Boisvif SA (fictional) · 18 staff · canton of Vaud
threshold: score ≥ 2.0 and 50% of the question covered
The assistant only knows these documents. Click one to open it.
Source document
Click a cited source to see the exact passage, highlighted in the document.
Demo: the keyword search really runs in your browser. The answers are prepared in advance from the passages; in production, a language model writes them from the passages found and always cites its sources. Internal rules of a fictional company, not legal texts.
Everything runs in your browser: nothing is sent or stored. Simplified demo: answers are prepared in advance or computed by simple rules in your browser. In production, a language model suited to your data does this work, and someone on your team signs off.
I want this for my businessWhat’s happening, step by step
-
You ask a question Input
In your own words, as you would ask a colleague. The assistant picks out the keywords and their synonyms: “vacation” means “holiday”.
-
Search through your documents Code
A search engine scans every passage of your documents in a few milliseconds, like the index of a giant book.
-
The best passages Code
Each passage gets a relevance score. Only those above the threshold are kept; otherwise, there’s no answer.
-
A grounded answer AI
The AI writes a clear answer only from the passages kept. Nothing reliable enough? It writes nothing.
-
The source is cited Output
Every answer names the document and the section. One click shows you the exact passage, highlighted.
-
Otherwise, HR takes over Human
No reliable source? The assistant says “I don’t know” and offers to pass the question on to HR.
Run the demo: each step lights up here.
ModelApertus 70B · Swiss open source
Data processedA server in Switzerland or yours
Picture a colleague who has read all your internal documents (staff handbook, procedures, prices) and answers every question by pointing to the exact paragraph. If it can’t find the answer, it says so and hands over to HR instead of making something up.
Words to know
- RAG
- Short for “retrieval-augmented generation”: the AI first finds the right passages in your documents, then writes its answer from them alone.
- Index
- Your documents split into passages and organised so they can be found in milliseconds. It updates when a document changes.
- Hallucination
- When an AI makes up an answer that sounds right but is wrong. Here, no reliable source means no answer: the assistant says so.
Why it’s safe
- It answers only from your documents and always cites the exact passage.
- No reliable source, no answer: it says “I don’t know” instead of making something up.
- It sends nothing on its own: passing a question on to HR takes your click.
- People only find the documents they are allowed to read, and you choose where the AI runs.
The agent’s pipeline
- Input Your documents PDF, Word, Excel or intranet pages, from Microsoft 365, Google Workspace or a shared folder.
- Code Chunking and index One passage per section, with document, version and date. Keyword (BM25) and vector (embeddings) index.
- Code Hybrid search Keywords and meaning, filtered by the access rights of the person asking.
- Code Ranking and threshold The best passages are re-ranked. Below the relevance threshold, the model isn’t called.
- AI Grounded answer The model only receives the passages kept and must cite its sources.
- Human HR hand-off With no reliable answer, the question goes to HR. Their reply can be added to the documents.
AI reads and writes. Code checks. Humans decide.
Stack
Guardrails
- No source, no answer: below the relevance threshold, the model isn’t called and the assistant offers to hand over to HR.
- Checked citations: code verifies that every citation points to a passage actually given to the model.
- Access rights respected: people only find the documents they are allowed to read.
- Up-to-date documents: the index tracks versions; a replaced rule is no longer cited.
What the assistant produces (JSON)
{
"question": "How many weeks of holiday do I get?",
"keywords": ["weeks", "holiday"],
"scanned": { "documents": 6, "passages": 18 },
"ranked": [
{ "passage": "r2", "source": "Staff regulations, art. 4.2", "score": 4.5, "kept": true },
{ "passage": "r3", "source": "Staff regulations, art. 4.3", "score": 1.8, "kept": false },
{ "passage": "r6", "source": "Staff regulations, art. 6.3", "score": 1.4, "kept": false }
],
"threshold": { "min_score": 2, "min_coverage": 0.5, "coverage": 1, "passed": true },
"answer": {
"text": "You are entitled to 5 weeks of holiday per calendar year…",
"citations": ["r2"]
},
"escalated_to": null
}Choose the engine
How we set it up
- Day 1 Free audit 30 minutes to choose the useful documents, the questions your team asks most and what must stay confidential.
- Weeks 1–2 Test on your documents We index a first batch of documents and test the assistant on real questions from your team, answer by answer.
- Weeks 2–4 Go-live The assistant is available where your team works (browser, intranet or team chat), with access rights agreed together.
- Every month Care Updated documents, unanswered questions, adjustments and a monthly report.
What we need from you
- Your reference documents: staff handbook, procedures, prices, internal FAQs
- Around twenty questions your team often asks
- Who may read what, and who should receive unanswered questions
What we build
The agents we set up
Each agent follows your rules. You start with one, chosen during the audit or the workshop.
Email triage
An email triage agent labels each message, picks out the deadline or amount, routes it to the right person and drafts a reply in the client’s language.
Document intake
For admin and documents: invoices and receipts are sorted, renamed and filed; amounts, VAT and IBAN are extracted, then re-checked by code.
Follow-ups and prospecting
Unanswered quotes, missing documents, sales outreach to new companies: a few individual messages, each one checked before it goes out.
Multilingual content
One brief becomes multilingual content for LinkedIn, Instagram or your newsletter, in French, German and English, in your brand’s voice.
Reporting
Each month, the agent gathers figures from your tools and writes a readable report for management or clients, which you check before it goes out.
Internal assistant
Your staff ask about a procedure or a price. The assistant answers from your documents, cites its source and says when it doesn’t know.
How we work
From audit to monthly care
-
Free AI audit
A 30-minute call or a 3-minute online quiz: three automations suited to your sector, with an estimate of the time saved.
-
Team workshop
Half a day of training at your office: using AI safely, writing your AI usage policy and choosing the first automation.
-
Automation
We build the agent and test it on your data. We aim for a first automation in service 2 to 4 weeks after the audit.
-
Monthly care
Hosting, monitoring, small improvements and a report every month.
Our commitments
You stay in control
A person approves
Agents prepare, your team approves. Nothing is sent automatically unless you choose that, case by case.
The model suits the data
Leading models or, for sensitive data, open-source models such as Apertus, hosted in Switzerland or on your own servers. You decide where your data is processed.
Built with Swiss data protection (FADP) in mind
A data processing agreement (art. 9 FADP), only the data the task needs, and Swiss hosting if you want it.
Frequently asked questions
What is the difference between an AI agent and ChatGPT?
ChatGPT answers when someone asks it something; an AI agent does a specific task within your process, every time it comes up. It uses your tools, follows your rules, and your team approves the result.
Are you an AI automation agency or a consultant?
Both: NotHuman is a small studio that advises on, builds and maintains AI agents for SMEs. After the audit and the workshop, we put the agents into service and look after them every month. You don’t need a technical team.
Which AI models do you use, and where is our data processed?
The model depends on the task and on how sensitive the data is, and you decide where your data is processed. We use leading models (GPT, Claude, Gemini, Mistral…) or open-source models such as Apertus, the open Swiss language model built by EPFL, ETH Zurich and CSCS. For sensitive data, these open-source models can be hosted in Switzerland or on your own infrastructure.
What if the agent gets it wrong?
It happens, which is why a person approves. We test every agent on your data before it goes live, and uncertain cases are flagged for a human check. Every correction is used to improve the rules.
How much does an AI agent cost for an SME?
It is priced on quote, after a free audit. The price depends on the number of tasks, the monthly volume, the tools to connect and the hosting: a model hosted in Switzerland does not cost the same as an online service. The quote shows a set-up cost and a monthly fee. Request a quote.
NotHuman
Which task should your first agent take on?
In a free 30-minute audit, we pick the tasks to automate first with you and estimate the time you could save.