AI automation

AI that solves a real problem.
Not AI for the press release.

We integrate practical AI features into the websites and workflows we already build. Selective projects, honest framing, real measurable outcomes. Quoted per project from €350.

What this service is

WebSilm AI process automation is the practical work of integrating AI features into websites and small-business workflows. We focus on the use cases where AI saves a measurable amount of time, removes a real bottleneck, or improves a metric — and we decline projects where AI is decorative or where the use case has no defensible business value. We are not an ML lab, we don't train custom models from scratch, and we don't sell autonomous agent swarms in production.

Every AI feature we ship is wrapped in three things you can see: a feature flag so you can disable it instantly, a hard usage cap so a runaway bug can't generate a €5,000 API bill, and grounding constraints so the model can only quote from your actual content. The site has to keep working when the AI doesn't.

Who it's for

You're the right fit if …

I.

You have a measurable bottleneck

Repetitive customer questions, slow document handling, content backlog. Something quantifiable — hours per week, queries per day, dollars per task. We'll tell you whether AI is the right tool, and how much you'd save.

II.

You're considering an AI feature

A chatbot, smart form, or AI-assisted search on your site — but you don't know if it's worth it. We start with a free 30-minute scoping call where we'll either green-light, downsize, or honestly recommend not doing it.

III.

You want someone who'll say no

Most agencies will sell you whatever you ask for. We turn down roughly half the AI projects we're approached with — usually because the use case doesn't justify the ongoing API cost or the risk of hallucinations. If we say yes, it's because we've thought about it.

What we actually build

Specific AI features, not vague capabilities.

01
Site-embedded chat assistant
A chatbot that answers questions using only your actual content (FAQ, docs, knowledge base). Retrieval-augmented generation grounded in your data, with explicit fallback to a human contact. Typical cost: €1,500–€3,500 build, €5–30/month API.
02
AI-assisted contact & intake forms
Smart routing (lead goes to the right person), draft replies generated for your review, and structured data extraction from free-text fields. Useful when you handle 20+ enquiries per week. Typical cost: €350–€1,200 build, < €5/month API.
03
Semantic search on your content
Visitors search by meaning, not just keywords. Vector embeddings of your articles / products / docs, served from Supabase Vector or Pinecone. Especially useful for content libraries above 50 pages. Typical cost: €1,200–€3,000 build, €0–10/month.
04
Document processing pipeline
OCR + summarisation + structured data extraction from invoices, contracts, forms. Drops processed output into your CRM, spreadsheet, or storage. ROI obvious when your team handles 50+ documents per week. Typical cost: €800–€3,500.
05
Content generation pipeline (with human review)
Drafts of blog posts, product descriptions, or social copy generated from a brief, queued for human review before publishing. We never ship 'fully automated' content — Google penalises bulk AI content without oversight. Typical cost: €600–€2,000.
06
Workflow automation with AI steps
Make.com, n8n, or Inngest workflows where one or two steps use an LLM (e.g. classify incoming emails, route by sentiment, summarise into a Slack message). The AI is one node in a larger automation, not the centrepiece. Typical cost: €400–€1,500.
07
AI-generated marketing visuals
Brand-consistent images for blog posts, social, hero sections — generated via Sora 2, Nano Banana, Midjourney, or open-source Stable Diffusion, with a style guide so output stays on-brand. Quoted per project; usually €350–€1,200.
08
An honest scoping conversation
The most underrated deliverable. 30 minutes, free, no sales pitch — we'll either tell you AI is the right tool and quote a project, or tell you a non-AI solution would work better and recommend that. The latter happens about half the time.
What we say yes & no to

The projects we take — and skip.

Yes, this is what we do

Projects with a measurable problem, a defensible business case, and a path to maintenance.

  • AI features inside web projects we already build Smart forms, FAQ chat, semantic search — bundled into a normal web build. Cleanest integration point.
  • Workflow automation with measurable time savings Hours saved per week, with a clear before / after metric we can measure.
  • Document Q&A grounded in your knowledge base Internal staff tool or external customer support — both work, both have clear ROI when document volume is high.
  • Content pipelines with explicit human review Drafting, summarising, classifying — always with a human approval step before anything ships publicly.
  • Semantic search across content libraries above 50 pages Below that threshold a normal full-text search is fine. Embeddings only earn their keep when the corpus is meaningful.
No, we'll politely decline

Projects where AI is buzzword-marketing, technically risky, or where a non-AI solution is obviously better.

  • Training custom LLMs from scratch We're a web studio, not an ML lab. For that work we'll point you to specialists.
  • Autonomous agent swarms in production Too risky for small business. The blast radius of an agent making a wrong call is hard to bound.
  • AI-generated bulk content with no human review Will get penalised by Google in months. Hurts your domain long-term. We'll never build this for you.
  • AI as the only differentiator from a non-AI version If your product is 'X but with AI' and X already exists at half the price, the math doesn't work.
  • Replacing your sales team with chatbots For B2B small business, qualified human conversations close. A bot that handles inbound is a polite gatekeeper, not a closer.
The stack

What we build with — and what we avoid.

Models & APIs
OpenAI (GPT-4o, GPT-5) Anthropic (Claude Sonnet, Haiku) Google AI (Gemini) Mistral Together AI Replicate Sora 2 Nano Banana 2

OpenAI and Anthropic for highest quality on language tasks; Mistral or open-source via Together / Replicate when cost or data residency matters; Sora 2 and Nano Banana for image / video generation. We pick per use case and tell you the expected monthly API cost upfront.

Tooling & orchestration
Vercel AI SDK LangChain Inngest Pinecone Supabase Vector n8n Make.com Anthropic MCP

Vercel AI SDK for streaming responses on the frontend, Inngest for durable async workflows, Supabase Vector or Pinecone for embeddings storage, n8n / Make.com for the no-code automation glue. We avoid agent-orchestration frameworks in production — too brittle, too many moving parts.

Custom AI vs alternatives

Where a WebSilm AI build fits.

Concern WebSilm AI integration ChatGPT subscription alone DIY no-code (Make / Zapier) Hiring an AI consultant
One-off cost €350–€5,000 quoted upfront €20/month subscription €0–€100 to set up yourself €5,000–€20,000 for slide decks
Ships to your customers Yes — embedded in your site No — only you use it Internal workflows only No — strategy doc only
Hallucination safeguards RAG + grounding + fallback None — raw model output Up to you to add Recommended in slides
Maintainable next year Yes — code in your repo Yes — provider-managed Brittle at scale No — consultant gone
API cost transparency Hard cap + monthly cost estimate Flat €20/month Per-operation pricing — surprises possible Recommended in slides — you implement
Best for Customer-facing AI features inside a real product Personal productivity, drafts, brainstorming Internal automations < 1,000 runs / month Enterprise transformation projects
AI questions

What people ask before adding AI to their site.

Honest answers — including the cases where the honest answer is 'don't add AI yet'.

What kind of AI work does WebSilm actually do?
Practical AI features built into the websites and small-business workflows we already make: AI-assisted contact forms, on-site chat trained on your documents, content generation pipelines with human review, semantic search using embeddings, document processing (OCR, summarisation, extraction), and workflow automation with services like Make.com or n8n that include an AI step. We integrate AI as a feature inside a normal web project — we are not an ML lab, we don't train custom models from scratch, and we don't sell autonomous agent swarms.
Should I add AI to my website? Honest answer.
Probably not yet, unless one of three things is true: (1) you have a measurable problem AI could solve — repetitive customer questions, slow document handling, content bottleneck — (2) you have enough usage that the API cost is justified, or (3) the AI feature is the actual product, not decoration. If your site gets 50 visitors a month and you're considering an AI chatbot to engage them, the chatbot is the wrong investment. Fix the underlying problem first.
How much does an AI integration cost?
Project work is quoted from €350 — a small AI feature added to an existing site, like a contact form with smart routing or a basic chatbot on your FAQ. More complex builds (semantic search across your content, document Q&A trained on your knowledge base, multi-step workflow automation) typically run €1,500–€5,000. The variation depends mostly on how clean your existing data is. We always quote a fixed price upfront after a 30-minute scoping call.
Will AI features make my site rank better in Google AI Overviews?
Adding an AI chatbot to your site does not, by itself, make Google or ChatGPT cite you more. What does help: structured content (FAQ schema, comparison tables in HTML, clear entity definitions), explicit AI-crawler permissions in robots.txt (GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, Google-Extended), and Person author bios with sameAs links. We do this work as part of our SEO service — see /services/seo/ for that side. The two services are complementary but distinct.
Do I need to use OpenAI or Anthropic, or can we use a free model?
Both are options. OpenAI (GPT-4o, GPT-5) and Anthropic (Claude Sonnet, Claude Haiku) deliver the highest quality and are paid per token. Free or self-hosted models (Llama 3.1, Mistral, Qwen via Together AI or Replicate) cost less but generally produce weaker output for complex tasks. For most small-business use cases — chat, classification, summarisation — a small paid model at the cheapest tier (€0.10–€2 per million tokens) costs less than €5/month even with steady traffic. We pick the model per use case and tell you the expected monthly cost upfront.
Who pays for the AI API usage after launch?
You do, directly to the API provider (OpenAI, Anthropic, Google AI). We set up the account in your name, configure usage limits and a hard monthly cap so a runaway bug can't generate a €5,000 bill, and document how to monitor usage. We never charge a markup on AI API costs. For a typical small-business chatbot or smart-form integration, expected monthly API cost is €5–50; we tell you the realistic range upfront based on your traffic estimates.
What about hallucinations — what if the AI makes things up?
This is the right concern. We mitigate it three ways: (1) ground all answers in your actual content using retrieval-augmented generation, so the model can only quote from documents you uploaded, (2) include explicit refusal instructions in the system prompt for anything outside that scope, (3) add a confidence threshold below which the AI hands the conversation to a real person or to a 'we don't know — please contact us' fallback. Hallucinations cannot be reduced to zero, but for an FAQ chatbot trained on your own documents, the rate is typically under 1% and easy to catch in monitoring.
Can the AI feature be turned off if it's not working?
Yes, by design. Every AI integration we ship includes a feature flag so you can disable the AI part with a single toggle without touching the rest of the site. If a chatbot starts misbehaving, hide it. If a smart-routing form is sending leads to the wrong place, fall back to direct email. We never build AI features that are load-bearing for the site — the site has to work without them.
Got an AI idea you're not sure about? Start with a 30-minute honest scoping call.
Talk to us — free 30-minute call