AI tools for SMEs: SaaS, RAG, agents or custom applications?
A practical build, buy and integrate comparison for Swiss SMEs, covering data, costs, controls and maintenance.
By Ralph Heezen
The best AI tool is the smallest solution that reliably improves a process and has clear ownership of data, costs and operations. Existing SaaS often suffices; sometimes only integration is missing. RAG, agents and custom applications need justification through workflow or control requirements.
Start with the problem#
- A standard tool handles the essential workflow: buy and introduce it properly.
- The tool works but lacks system data: integrate it.
- You need answers from versioned internal documents: assess RAG.
- Unstructured inputs require several controlled actions: pilot a narrow agent.
- Workflow, interface or governance is specific: consider custom development.
- Benefit, owner or data quality is unclear: do not build yet.
Build, buy and integrate#
| Option | Good fit | Data and cost | Maintenance | Main risk |
|---|---|---|---|---|
| AI SaaS | Common standard tasks | Provider data and recurring licences | Provider plus internal adoption | Unclear use and price changes |
| Workspace bot | Instructions and approved files | Low/medium setup plus subscriptions | Prompts, files and access | Limited enforceable controls |
| RAG | Sourced answers from internal documents | Index and model costs | Versions, permissions and deletion | Convincing use of outdated sources |
| Integration | Connect useful tools to systems | Build plus operations | APIs and error handling | Invisible process failures |
| Agent | Language-guided multi-tool tasks | Pilot plus variable usage | Models, tests and boundaries | Incorrect or unauthorised action |
| Custom app | Distinctive workflow or governance | Highest initial responsibility | Product team required | Building what standard tools already solve |
These options can combine: a custom app may use a purchased model; an agent can use existing integration; RAG can remain read-only.
1. Buying SaaS#
Meeting transcription, editing, common support drafts and code assistance often fit standard products. Check purpose, permitted inputs, training settings, contracts, roles, retention, export and replacement cost.
Subscriptions may become expensive with many users or modules. Custom software can also be locked to proprietary model features. Compare total dependency and cost.
2. Configured assistants#
A Claude, ChatGPT or workspace assistant can follow style instructions, templates and approved files. It can help with drafts, checklists and summaries.
Prompts do not enforce permissions and uploaded files are not automatically current. Live ERP data, role-specific access and controlled writes need additional integration and policy layers.
3. RAG for knowledge#
Retrieval-Augmented Generation finds relevant sections of authorised documents before generating an answer. It fits current quality procedures, approved terms and complaint instructions.
Maintain source, version, validity, permission and deletion down to document sections. A vector database alone is insufficient. For current outstanding invoices, a direct ERP query is usually clearer than document retrieval.
4. Integration#
Keep a useful interface and model while a controlled bridge supplies ERP, CRM, shop or email data. Examples include status lookups, CRM lead creation, consolidated reports and PDF-to-draft-order workflows.
Revenue and margin should be calculated deterministically. AI may explain deviations afterwards, but should not invent underlying metrics. See system integration and data/BI.
5. Agents#
An agent can interpret a request, identify missing details, search records, ask about ambiguous matches and prepare a draft for human review.
Use narrow tools, server-enforced approvals, error paths, spending limits and audits. Agentic AI and MCP explains architecture. Our copilot demo shows direct tools; the Bexio MCP server provides a working reusable catalogue.
6. Custom applications#
Custom development fits distinctive workflows, guided interfaces, multiple systems and roles, field filtering, enforced approvals or frequent use justifying maintenance.
The valuable custom elements are process, data contracts, interface, tools and governance. The model usually remains a specialist provider’s product. ‘Our own AI’ alone is not a business case.
Data residency#
Swiss application hosting does not mean all context stays in Switzerland. Map:
| Location | Questions |
|---|---|
| Sources | Which ERP, CRM, email and document fields? |
| App/gateway | Where are permissions and filtering enforced? |
| Model | What is transferred, retained or used for training? |
| Logs/backups | What is stored and for how long? |
| RAG index | Where are text, embeddings and access metadata? |
Tokenisation can reduce exposed identifiers but does not automatically anonymise data. See the architecture.
Five forms of lock-in#
Assess data export, workflow ownership, tool contracts, model portability and operational handover. MCP can improve tool reuse, but tests, small interfaces and separation of business logic remain necessary.
Total costs#
Include selection and training; APIs and cleansing; licences, tokens and indexes; hosting, support and updates; model/API changes; tests and approvals; and exit migration.
A free prototype tests a small flow. Existing Bexio modules start at CHF 5'900 excluding VAT; custom apps receive scoped fixed quotes. Model and operational costs remain separate.
Maintenance#
Test normal cases, ambiguity, missing data and system failures. Monitor model behaviour, API fields and limits, staff roles, stale documents, unexpected usage and expanding permissions. Assign an owner, budget, incident response and shutdown method before approval.
When not to build#
Avoid custom AI when use is rare, existing software suffices, data or ownership is unclear, fixed rules work better, or success cannot be measured. Process cleanup, dashboards or a small integration may be the better investment.
Four practical steps#
Describe one process with frequency and effort. Identify risks and decision ownership. Compare the smallest options using the same cases. Pilot with permission to stop if standard software wins.
Our AI agency assesses fit and builds controlled solutions; AI agents and MCP covers specific technical implementation.
Decide between build, buy and integration
We first assess whether custom development is appropriate.
View AI agency