AI application
A product whose core feature relies on a model: summarise, classify, extract, generate, understand an image or a document (multimodal).
Build With Me / AI
Build your AI product with me.
An assistant that answers from your documents, an automation that removes a repetitive task, an AI feature in your SaaS: useful when the problem is real and the result can be checked.
To build an AI product, start by checking that the problem needs AI: rules, search or a well-designed form are often enough. If AI helps, call the model from your server, give it the context it needs (document retrieval, limited tools), require structured output, measure quality on a set of examples before every change, cap the costs, and keep human approval before any irreversible action. An AI agent does not replace human judgment: it equips it.
Checked on
In brief
- For
- Teams and founders who want an AI feature or an automation that holds up in production.
- Covers
- Assistants, agents, RAG and search, document processing, customer support, process automation, AI-powered SaaS.
- First question
- Does this problem really need a model, or rules and good search?
- Guardrails
- Evaluations, cost caps, least-privilege tools, human approval before irreversible actions.
- Stated limit
- No Bynilson product in production depends on an LLM today.
Three ways to build
Learn, build it together, or have it built.
01 · Learn and buildResources available — A guide on this site and the official documentation.
Learn and build it yourself
This page gives the architecture, the guardrails and a staged path, with the model providers’ official documentation. To decide whether you can build your product with AI tools or should delegate, the free build, buy or delegate diagnostic (in French) helps decide. Build School teaches building a web product with a coding agent — an AI tool is one of its product types.
02 · Build it togetherOn request — Reviewed case by case, with no price or date promised.
Build it together
You build; Bynilson reviews the architecture, the evaluation set, the costs and the risks, in writing or in a session. For a product centred on an LLM (agent, chatbot), a request is reviewed case by case: no package, no promise before seeing the problem.
03 · Have it builtGuided offer — An existing Bynilson product, on published terms.
Have it built
No Bynilson product in production depends on an LLM today. Delivery builds on what is proven — rules engines and workflows, like those behind Atamatax, Horvera and Vellaci — and adds AI where its result can be measured. The business process automation engagement exists; an agent or a chatbot centred on an LLM is reviewed case by case, through the project qualification.
No Bynilson product in production depends on an LLM today: delivery builds on proven rules engines and workflows, and adds AI where it can be measured.
The products
Which AI product are you building?
Six product shapes. For each, the sensible way to build it, and the trap that makes it fail in production.
AI agent
A model that chains steps and calls tools (search, read, draft) to reach a goal, orchestrated by your code.
Assistant, chatbot and search over your content
Answer questions from your documents or knowledge base — for your customers (support) or your team (internal assistant).
AI-powered SaaS
A regular SaaS — accounts, data, subscription — with an AI feature that adds measurable value.
Document processing
Extract fields from an invoice, classify documents, check that a file is complete, compare a document against rules.
Process automation
A sequence that runs on its own across your tools: an order creates an invoice, a request is sorted and assigned, a report goes out every Monday.
The real problem
Does this problem need AI?
The question that saves the most money. A model is useful for understanding or producing language; for the rest, simpler tools are more reliable and cheaper.
| Your need | Often enough | When AI becomes useful |
|---|---|---|
| Apply known conditions (eligibility, calculation, pricing) | A deterministic, tested rules engine | Almost never for the decision; sometimes to read free-text input |
| Find information | Full-text search and good filters | When questions are phrased freely and the words do not match |
| Collect information | A well-designed form | When information arrives as free text, emails or documents |
| Classify or extract | Rules, if the format is stable | When documents vary and human review is planned |
| Draft a text | Text templates | When every case differs and a person reviews before sending |
Write the problem in one sentence, with the person who has it and what the current situation costs them. If a written rule solves 90% of cases, start with it: AI can handle the rest later, on a measurable basis.
AI, or conventional rules-based software?Until the dedicated comparison is outWhen should a SaaS use AI?
The architecture
The parts of an AI product
- Model and provider choice
- Compare on your own examples: quality, latency, cost, context size, where data is processed and kept. Keep the code independent of a single provider so you can switch.
- Instructions (prompt)
- A clear instruction: the role, the task, the expected format, examples, and what to do when unsure. Version it like code.
- Tool calling
- The model asks to run a function you defined; your code runs it, after checking the arguments and the rights.
- Structured output
- The model answers according to a schema (typed fields) your server validates. An invalid answer is rejected or retried, never used as is.
- Retrieval (RAG)
- Find the relevant passages in your content and give them to the model with the question. The most common way to answer from your data without retraining a model.
- Context management
- The model only sees what you send: pick the passages, summarise the history, and never send more personal data than the task requires.
Quality
Evaluate before claiming it works
- An evaluation set: a few dozen real examples, with the expected answer or the criteria of a good one, written before building.
- Regression evaluations: rerun that set on every change of prompt, model or retrieval, and compare scores before deploying.
- Hard cases and trap cases: off-topic questions, missing documents, attempts to hijack the assistant.
- Observability: trace every call (input, retrieved context, output, tools called, duration, cost) without keeping more personal data than needed.
- User feedback: a simple way to flag a wrong answer, connected to the evaluation set.
Cost
Keeping cost under control
- Cost per task
- Measure the average cost of a complete task (not of one call), on real examples, before setting a price.
- Caching
- Reuse identical results, and the provider’s prompt caching when it offers one.
- Rate limiting
- Cap calls per person and per minute: an open AI form is an open invoice.
- Per-user quotas
- A number of calls or credits per account and per month, visible to the user, with clear behaviour once reached.
- The right model at the right step
- A smaller model to classify or extract, a more capable one only where the task needs it.
Security and data
Protect the data, limit the actions
- Personal data: know what you send to the provider, where it is processed, how long it is kept, and whether it is used to train their models. In Switzerland, the revised Federal Act on Data Protection (nFADP) has been in force since September 2023; the Federal Data Protection and Information Commissioner (FDPIC) publishes guidance. This is not legal advice.
- Prompt injection: malicious text in a document, a web page or an email can give the model instructions. It is the first risk on the OWASP list for LLM applications. Treat any external content as data, never as instructions.
- Least-privilege tools: each tool reaches only what the task needs, for the signed-in person, and nothing else.
- Human approval: before sending, paying, deleting or publishing, a person confirms.
- Errors and fallback: timeout, invalid answer, provider down — the product shows a clear state and offers an alternative, it does not hang.
- Authorised access only: an AI product only connects to systems and data you have explicit authorisation for.
In production
Deploy, monitor, change models
- Deployment
- API keys stay on the server, per environment. Long calls go through streamed responses or background jobs.
- Monitoring
- Error rate, latency, cost per day, user reports: someone looks at these numbers every week.
- Model changes
- Providers update or retire their models: pin the version you use, and move to a new one only after rerunning the evaluation set.
The path
Eleven stages, from the problem to an AI feature in production
- 01
Write the problem down
Who, which task, how much time or how many errors it costs today.
AvailableWhen should a SaaS use AI?
- 02
Check that it needs AI
Rules, search, a form: the simplest solution first.
AvailableFree diagnostic: build, buy or delegate (in French)
- 03
Gather examples
Real cases, with the right answer: your future evaluation set.
- 04
Choose the model
Compare two or three models on your examples: quality, cost, latency, data handling.
- 05
Build the first call
On the server, with structured output validated against a schema.
AvailableAI SDK documentationVercel
- 06
Add context
Retrieval over your content, with access rights applied.
AvailableOpenAI platform documentationOpenAI
- 07
Limit the tools
Least privilege, human approval before irreversible actions.
- 08
Evaluate
Rerun the evaluation set, fix, repeat — before every deployment.
AvailableThe limits of AI in web development
- 09
- 10
Check data protection
What goes to the provider, for how long, on what basis.
AvailableFDPIC — Swiss Federal Data Protection and Information CommissionerSwiss Confederation
- 11
Go live and monitor
Errors, latency, cost, reports, a pinned model version.
Build it yourself with AI, or delegate?Until the dedicated comparison is outFree diagnostic: build, buy or delegate (in French)
Frequently asked questions
How do you build an AI agent?
Define a precise task, give the model a few least-privilege tools, cap the number of steps, log every call and require human approval before any irreversible action. Often start with a fixed workflow, where your code decides the order of steps: it is more predictable than a free-roaming agent.
How do you evaluate an AI product?
With an evaluation set written before building: real examples and the criteria of a good answer. Rerun it on every change of prompt, model or retrieval, add hard cases and errors users report, and do not deploy a version that lowers the score.
RAG or fine-tuning?
To answer from your documents, retrieval (RAG) is usually the first choice: content stays current, sources can be cited, and access rights apply to the search. Fine-tuning is better suited to learning a consistent format or style; it is not a good way to teach facts that change.
AI automation or conventional software?
If the rules are known and stable, conventional software is more reliable, faster and cheaper. AI is useful for the step that requires understanding free text or varied documents, with checks and human review around it.
How much does an AI product cost?
Two separate costs: building it (depends on scope, like any software) and using it (every model call is paid). Measure the cost of a complete task on your examples before setting a price. No generic figure replaces that measurement; the SaaS cost calculator helps with the build side.
Is my data safe with an AI model?
It depends on the provider and its configuration: check where data is processed, how long it is kept and whether it is used for training. Send only what the task requires, and apply access rights before retrieval. In Switzerland, the nFADP applies; this is not legal advice.
Can an AI agent replace a person?
No. An agent can prepare, sort, draft and suggest; the decision that commits — send, pay, refuse, publish — stays with a person who sees what the agent did and why.
Sources
- NIST AI Risk Management FrameworkNISTChecked on 23 Oct 2026
- Claude developer documentationAnthropicChecked on 23 Oct 2026
- AI SDK documentationVercelChecked on 23 Oct 2026
- OpenAI platform documentationOpenAIChecked on 23 Oct 2026
- OWASP Top 10 for LLM applicationsOWASPChecked on 23 Oct 2026
- FDPIC — Swiss Federal Data Protection and Information CommissionerSwiss ConfederationChecked on 23 Oct 2026