Deliverable 01
Indexed knowledge base
Your docs, help-centre articles, FAQs, past tickets, and any internal wiki you point us at, ingested and chunked into a vector store with source-of-truth links preserved.
AI Support Agent.
An AI support agent is a retrieval-augmented chatbot grounded in your own docs, tickets, and product knowledge. We build, evaluate, and deploy yours: it answers customer questions in your tone, deflects the repetitive tier-1 traffic that’s eating your team, and escalates cleanly when it’s out of its depth.
Time to ship
~3 weeks*
Starting from
€4,500*
Channels
Web + API
Languages
Multilingual
Indicative timeline and starting price. Final scope and quote agreed after the intro call.
Who it’s for
Support ticket volumes are growing roughly 14–20% year-on-year across mid-market SaaS while support-team headcount stays flat (Zendesk CX Trends, 2025). The teams below feel it first.
- SaaS, e-commerce, and service teams whose support volume is growing faster than the team.
- Founders watching the same five questions burn ten hours a week of customer-facing time.
- Agencies and operators who want a deflection layer in front of human support, not a replacement for it.
- Teams that have tried generic chatbots, hated the “I can’t help with that” loop, and want something grounded in their actual docs.
What you get
Deliverable 02
Answer agent
A grounded chat agent built on Claude or frontier OpenAI models. Answers cite their sources, refuse to hallucinate when the docs don’t cover something, and escalate to a human with full context.
Deliverable 03
Channel deployment
Embedded chat widget on your site plus an API endpoint your existing helpdesk (Intercom, Zendesk, Help Scout, Front, custom) can call. Slack and WhatsApp on request.
Deliverable 04
Eval & monitoring
A test set built from your real tickets, automated quality checks on every prompt change, and a small dashboard showing deflection rate, escalation rate, and the questions you’re still missing answers for. Grounded RAG agents typically resolve 40–60% of tier-1 inquiries end-to-end (Intercom State of AI in Customer Service, 2025); the dashboard tracks where you actually land.
How a build runs
Week 1 / D 1–3
Kickoff & content audit
One working session with your support and product leads. We pull a sample of recent tickets, map the question categories, and identify the doc gaps that will limit the agent’s ceiling before any code is written.
Week 1 / D 4–7
Ingest & first agent
We stand up the vector store, ingest your sources, and ship a first version of the agent against a real account. You can chat with it by end of week one.
Week 2
Eval-driven tuning
Build a test set from your real tickets, run it against the agent, and tune retrieval, system prompt, and refusal behaviour until quality clears the bar you set in week one.
Week 3
Channel deployment & handover
Embed the widget, wire the API into your helpdesk, ship the dashboard, and walk your team through the runbook. You leave week three with a live agent and a quote for any add-ons.
Indicative timeline. Larger knowledge bases or non-trivial helpdesk integrations can stretch this; we confirm dates after the kickoff session.
Fixed scope. EUR pricing.
Starting from €4,500*, payable 50% on kickoff and 50% on handover. Scope and final price are agreed in writing after the intro call — no surprise add-ons.
Optional retainer from €750/month covers monitoring, content updates, and prompt tuning as your product changes. Cancel any time.
Industry benchmarks put the loaded cost of a human-handled tier-1 ticket between €8 and €25 (Forrester); a working agent typically pays back the build inside the first quarter.
Start a projectStarting price reflects a focused single-product knowledge base and a standard helpdesk integration. Multi-product or custom-stack builds quoted after scoping.
FAQ
Will the agent hallucinate?
The whole architecture is built to prevent it. Answers are grounded in retrieved chunks from your sources, the system prompt instructs the model to refuse when context is missing, and the eval suite blocks prompt changes that regress refusal behaviour. It’s never zero risk, but it’s a different category from a generic chatbot.
Which model do you use?
Default is Claude (Anthropic) or frontier OpenAI, picked per project based on your data residency, latency, and cost requirements. You own the model choice and can swap providers post-handover.
Where does the data live?
In infrastructure you control. Vector store and orchestration deploy in your cloud account (AWS, GCP, Azure) by default; we can also run on managed providers like Pinecone or Supabase if you prefer.
How do you measure success?
Three numbers: deflection rate (tickets the agent fully resolved), escalation quality (does the human get useful context?), and refusal accuracy (does it know when to say “I don’t know”?). Baseline is captured in week one; we report against it on handover.
What about API costs?
Anything under €50 in model spend during the build is on us. Production usage is billed by the provider directly to your account — no markup, no reseller margin.
What happens after the build?
Three options: (1) take the repo and run it internally, (2) keep us on retainer for monitoring and content updates, (3) scope a follow-on build (voice, additional channels, multilingual expansion). No pressure to continue.
Ready to deflect the easy half of your inbox?
Tell us about your product, your support volume, and where the docs live. We’ll come back within one business day with a proposed scope and a quote.
Open the contact form