The process exists nowhere else
Your working sequence is a competitive advantage or a regulatory constraint specific to your industry. No vendor has modelled it.
Tailored AI systems · Switzerland
Custom AI agents: when standard is no longer enough.
Beyond Sania, we design AI agents and automation systems built around your processes, your tools and your validation rules. From Geneva, for Swiss companies.
Direct answer
A custom AI agent is a system able to understand a business request, consult only the authorised tools, apply the company's rules, then prepare or execute an action. Swiss Agent Network builds these systems for Swiss companies when an off-the-shelf product does not cover the process: specific business logic, internal tools, compliance constraints or a high volume of exceptions.
Two offers, two needs
Sania on one side. Custom systems on the other.
We are not selling the same thing under two names. The choice depends on the nature of the process, not on company size.
| Criterion | Sania (product) | Custom system |
|---|---|---|
| Target processes | Emails, quotes, calendar, follow-ups | Any documentable process |
| Setup | Guided, fast | After scoping |
| Connected tools | Existing integrations | Including internal and industry tools |
| Decision rules | Configurable | Specifically modelled |
| Cost | Monthly subscription | Project + operations |
| Good signal | "Admin work is slowing me down" | "Nobody sells what I do" |
If your need matches the left column, start with the AI assistant for Swiss SMEs: cheaper, faster and sufficient in most cases.
Decision criteria
Five signals that justify a tailored system.
Data lives in internal tools
In-house ERP, industry database, structured files, sector application without documented public API: access has to be built.
Decision rules are numerous
Pricing conditions, special cases, approval thresholds, client exceptions. A configuration form cannot express them.
Traceability is a requirement, not a comfort
You must be able to explain every action to an auditor, a client or an authority, with a usable log.
Volume makes manual validation expensive
The process repeats often enough for partial automation to genuinely change the workload.
Counter-signal: only one of these
If you tick just one criterion, a pilot with a standard product is almost always the right first step. We will tell you so.
Method
How we build a custom AI agent.
Every stage produces a decidable deliverable: you can stop the project at the end of any of them.
Process scoping
We observe the task as it is actually performed, not as it is described in a procedure. Deliverable: scope, required data, decision points, success criteria and what we recommend not automating.
Architecture and permissions
Which tools to connect, minimal access levels, autonomy per action, human validation points and log format. Deliverable: technical diagram and permission matrix.
Minimal scope development
A single use case, connected to real tools, with systematic validation at the start. Deliverable: a system usable by someone on your team.
Controlled go-live
Real usage over an agreed period, recording correctly handled cases, exceptions and validation time. Deliverable: measurements from your activity, not from a benchmark.
Expansion or shutdown
Based on measurements: more autonomy, a second process, or a deliberate stop. A project that demonstrates nothing should stop. See our ROI calculation method.
Our position
What we refuse to automate.
A provider who accepts everything does not protect your company. Here are our limits, stated before the quote.
Irreversible decisions without validation
Sending a contractual commitment, a payment, a termination, an external publication: the agent prepares, a person decides.
Processes nobody can describe
If two employees state two different rules, the problem is organisational. Automating it freezes the ambiguity.
Judgements about people
Screening applications, evaluations, decisions with individual impact: we do not build systems that decide in a manager's place.
Inventing missing data
A missing price, an unclear clause, an unknown contact: the system flags the gap, it does not fill it.
Rare processes
Three times a year, automation costs more than it returns, maintenance included.
System examples
What a custom agent can take on.
Design scenarios, presented as such: they describe what we know how to build, not published client deployments.
Multi-channel inbound request handling
Qualify a request arriving by email, form or transcribed call, enrich it from the CRM, route it to the right person and prepare the first reply.
Business file preparation
Gather expected documents, check completeness, flag what is missing and assemble a file ready for human review.
Deadline and obligation tracking
Monitor dates from several sources, prepare internal and external reminders, escalate what has not moved.
Reconciliation between two systems
Compare data from two tools that do not talk to each other, produce qualified discrepancies and propose corrections for validation.
Monitoring and opportunity detection
Watch defined public sources, filter against your criteria and deliver a sourced, verifiable summary.
Internal assistant over a document base
Answer team questions from your authorised documents, citing the source and flagging uncertain areas.
What drives the budget
A custom project is quoted after scoping.
We do not publish a flat rate, because a price displayed without a scope is a decorative number. Here is what actually determines the cost.
Design and scoping
Time spent observing the process, modelling the rules and defining success criteria.
Integrations
Number of tools, presence or absence of a documented API, data quality, access management.
Usage costs
Model consumption depending on volume handled and task complexity, re-billed transparently.
Operations and maintenance
Hosting, monitoring, third-party tool changes, rule adjustments over time.
FAQ
Frequently asked questions about custom AI agents.
What is a custom AI agent?
A system built around one specific process in your company. It receives an objective, consults only authorised data and tools, applies your rules, then prepares or executes an action depending on the validation level chosen. Unlike a standard product, its business logic, integrations and safeguards are specific to you.
When should I choose custom AI over a standard product?
When the process is specific to your company, depends on internal tools not covered by market products, involves many decision rules, or when compliance demands fine-grained control of access and traceability. If your need is emails, quotes, calendar and follow-ups, Sania already covers those cases at a far lower cost.
How long does it take to build a custom AI agent?
It depends on the number of integrations, data quality and the level of validation required. We work in stages: short scoping, a first narrow scope in production, then progressive expansion. The precise timeline is confirmed after scoping — not before, because a date announced without a scope commits nobody.
How much does a custom AI agent cost in Switzerland?
Quoting happens after scoping. Budget items are design, development, integrations, model usage costs, hosting and maintenance. Our page on AI agent pricing in Switzerland details each item and the orders of magnitude to anticipate.
Who owns the system that is developed?
Rights, hosting and reversibility terms are set contractually before development. We recommend covering them during scoping: access to configurations, data export, conditions for stopping the service.
Is a custom AI agent compatible with the Swiss FADP?
Compliance depends on how the project is framed rather than on the technology: documented purpose, data minimisation, information of data subjects, sub-processor management, logging. These are defined during scoping. See our guide on AI agents and data protection in Switzerland.
Do you work with our existing tools or do we have to replace everything?
We start from your tools. Replacing a working system in order to install AI is almost always the wrong sequence: it adds a migration project to the risks of the automation project. See AI integration in companies.
Do you work outside Geneva?
Our team is based in Geneva and works with companies across French-speaking Switzerland, including Lausanne and the canton of Vaud. Scoping work happens on site or remotely depending on the project.
Editorial transparencyThis page describes the design method of Swiss Agent Network Sàrl, publisher of Sania. The system examples are design scenarios presented as such, not published client deployments: we will publish named case studies when clients authorise us to. No duration, price or quantified gain is announced on this page, in the absence of a defined scope. Reviewed by Ethan Avon, CEO. Last updated: 30 July 2026.