← Home / Service 02 / 03

Agents that sound like the company.

Custom AI agents with your brand voice, in your languages, trained on your knowledge. Built in weeks, owned by you. The same studio that designs the site writes the prompt.

Delivery1–6 wk
From€2,500
LanguagesEN · ES · PT
01 — Three routes

The same conversation, three conditions.

[ 01 ]

The off-the-shelf widget

Generic chatbot, the same script as the next ten thousand sites. Doesn't know the company, routes everyone to a form, brand voice set to SaaS default. The prompt and the data aren't really yours.

[ 02 ]

The AI consultancy

Six months of discovery before anything ships. Strategy decks before working software. Built by a team that won't be there to maintain it. Lock-in by design, dressed up as enterprise rigor.

[ 03 ]

The studio route

System prompt with your brand voice, retrieval over your real documents, multilingual from day one. Shipped in weeks. Same hands from prompt to launch. Full ownership of prompt and data — no lock-in.

↳ This is us
02 — Agent types

Four formats.
Same craft.

[ 01 ]

Site agent

One agent, one job, done well. A single-purpose agent embedded on your site — answers questions, qualifies leads, routes when it should. The cleanest way to put a real conversation on a landing page.

From€2,500
Timeline1 wk
Single-purposeEmbeddedBilingual
[ 02 ]

Knowledge agent

Trained on what your company actually knows. Retrieval over your real documents — PDFs, internal pages, product specs. Answers from source, not from a generic LLM guess. The format most corporate clients land on.

From€5,500
Timeline2–3 wk
RAGFrom sourceMultilingual
[ 03 ]

Multi-channel agent

One voice across web, WhatsApp and email. Same agent, multiple surfaces — the conversation that starts on your site continues on WhatsApp without losing context. For teams that already live where their clients write.

From€8,500
Timeline3–4 wk
WhatsAppEmailShared context
[ 04 ]

Custom workflow agent

An agent that does the work, not just answers. Intake, triage, scheduling, structured handoff — automated end to end. For teams where the bottleneck is operational, not informational.

From€12,000
Timeline5–6 wk
Tool useCRMEnd-to-end
03 — Running costs

The agent lives after launch.

The real costs of keeping it running — hosting, model inference, monthly tuning. Declared up front, no surprises.

Light

From €500/mo
  • Hosting and model inference
  • Up to ~5,000 conversations / month
  • Monthly prompt tuning from real logs
  • Email support, 48h response

Standard

From €900/mo
  • Hosting and model inference
  • Up to ~20,000 conversations / month
  • Biweekly tuning, performance reports
  • Priority response, 24h

Heavy

Custom
  • High-volume, multi-channel deployments
  • Dedicated tuning and monitoring
  • SLA on response and uptime
  • Direct line to the studio
04 — What's in the build

Clear about what costs extra.

Always included

  • System prompt with brand voice
  • EN / ES bilingual by default
  • Site integration, <2s latency
  • Conversation logs and analytics
  • Full ownership of the prompt
  • Vendor-agnostic model choice

Often included

  • RAG over client documents
  • Calendar or CRM integration
  • Human handoff with context
  • Structured form capture
  • Lead qualification flows

On request

  • WhatsApp / email / Slack
  • Voice mode
  • Fine-tuning on private data
  • Additional languages (PT, FR, IT)
  • Internal system integration
  • Custom evals and monitoring
05 — Process

From brief to first conversation.

W1

Listen + Frame

What does the agent need to solve. What it shouldn't touch. Sources of knowledge — documents, site, internal pages. System prompt v1 drafted by the end of the week.

W2

Make

Build, internal testing, retrieval setup if applicable. The agent starts holding conversations with us before it holds them with anyone else.

W3

Rehearsal

Real-case testing with the client. Tone calibration, edge cases, refusals. The agent learns what to say, and what to stop saying.

W4

Ship

Deploy on site, handoff of prompts, logs and credentials. The client team learns to read conversations and where to ask for tuning.

Tuning

Prompts get sharper with real use. Monthly or biweekly tuning, depending on tier. The agent gets better at sounding like the company over time.

06 — Selected work

Three slots. One real, two soon.

[ 001 / 003 ]

ARC site agent

Type / Site agent Industry / Studio (self) Year / 2026 Status / Live

The agent on this site. Built with the same prompt, the same voice and the same retrieval that we ship to clients. Live and learning. Talk to it →

[ 002 / 003 ]

Knowledge agent — In production

Type / Knowledge agent Industry / Fitness studio Status / In production Languages / EN · ES

Bilingual agent for trial bookings, schedule and pricing questions, complex requests routed to the team.

[ 003 / 003 ]

Custom workflow — In production

Type / Custom workflow agent Industry / Cultural institution Status / Discovery Languages / EN · ES · PT

Visitor-facing agent for programming, ticketing and accessibility questions. Multilingual from day one.

07 — Honest answers

Questions worth asking.

Why not just use ChatGPT or Claude directly?

Because no. They have to be trained, tamed. A custom agent knows your business, carries your voice and your way of communicating — you notice it fast.

What happens with our data? GDPR?

We sign a DPA before kickoff and everything is stored in a secure place.

Who owns the prompt?

Yours, and we hand everything over at the end. If you later want to work with someone else, you take it with you and you've got a head start.

What model do you use?

Depends on what for. Anthropic, OpenAI, open models — chosen by cost, latency, and how well it handles the language or the case. There's no best one; there's a best one for this.

What if the agent gets it wrong?

It can happen, but it gets resolved — that's what training is for. Logs, refusal rules for sensitive topics, and human handoff when it can't answer. Monthly tuning at minimum, that's what the retainer is for.

What about post-launch?

Details up above.

Want to see it before commissioning?

The fastest way to know if this fits is to talk to the agent on this site. It was built with the same prompt and the same voice we ship to clients.

Next opening: Q3 2026 · 3 founding spots.