Traços HQ: AI agents and Humans in harmony

Who it is for

Traços Interiores is a Portuguese SME that has represented premium design brands for more than 25 years. Its catalogue covers lighting, acoustics, outdoor furniture, workplace furniture and biophilic design. The clients are hotels, offices, architects and restaurants, and almost every project mixes several of those areas.

The team is small. Each segment has one person who knows its brands, collections and prices. That works well until a client asks for a lobby, a meeting room and a rooftop bar in the same week.

The problems we set out to fix

A combined proposal takes a day. When a salesperson visits a hotel, the client wants lighting, acoustics and furniture in one answer. Today that means asking three colleagues, searching hundreds of PDF catalogues and stitching the replies together by email.

The CRM falls behind. Visits, notes and new contacts reach Pipedrive late, or twice, because several people enter them by hand. Before a visit, nobody is sure the client history in the CRM is complete.

Prospecting happens when someone has time. New clients are found one Google search at a time. In practice we manage it maybe once a quarter, and the same company can end up in the CRM twice.

What Traços HQ is

Traços HQ is a workspace that looks like Slack. It has channels such as #clients and #leads, and people and AI agents talk in the same conversation. Some of the "colleagues" in the chat are agents:

• The Orchestrator manages the AI team. It reads the client's history in Pipedrive, decides which specialists to ask and returns one combined answer. It is the only agent allowed to read or write the CRM.
• Five product specialists, one per segment: Daniel for lighting, Tomás for acoustics, Maya for outdoor furniture, André for workplace and Inês for biophilia. Clients already talk to them as voice agents on the Traços website. In Traços HQ they answer the team, using only the brands' own catalogues.
• The Leads agent searches the web for companies that fit Traços, for example hotels that opened or were renovated recently, and explains why each one fits, with a link to the source.

Every action that would change the CRM goes to an approval queue first. A person clicks Approve, and only then is it written to Pipedrive.

The interface is in European Portuguese, the team's working language, and every agent replies in the language of the message.

What makes it different

Traços HQ is shaped around the way Traços sells. A hotel project crosses several segments, so the Orchestrator asks the right specialists in parallel and returns one answer organised by area of the building, the way a salesperson presents it to the client.

The CRM has a single door. Only the Orchestrator reads and writes Pipedrive, and every write waits for a person's approval. People and agents feed information to it, which keeps duplicates and half-filled records out.

The specialists work from the catalogues and price lists of the brands Traços represents. A first proposal therefore names real collections Traços can supply, with prices marked as indicative for the segment owner to confirm.

The role of AI

Grok does the reasoning and the web searches, and a person approves every change to the CRM.

• The Orchestrator is Grok with four tools: crm_search and crm_get_client to read Pipedrive, ask_specialist to consult a specialist, and propose_crm_write to place a note or activity in the approval queue. Grok reads the request, decides which tools it needs and combines the results into one reply.
• Each specialist is a Grok call whose only source is that segment's catalogue file. The prompt keeps it within those products and has it present a brand's price range as indicative, quoting a product price only where the catalogue gives one.
• The Leads agent uses Grok's live web search. It must search before answering, and it returns each company with the reason it fits and the link to the source.
• The approval queue is plain code, with no AI in it. Nothing reaches Pipedrive until a person clicks Approve.

How it works: one real request

This is the flow shown in the video.

1. On Monday, Paulo has a visit to Casa Maré, a boutique hotel in Lagos, on Thursday. He writes in #clients: "I'm visiting Casa Maré on Thursday. They have an echo problem in the lobby and want a rooftop bar. Suggest lighting, acoustics and outdoor furniture."
2. The Orchestrator opens Casa Maré in Pipedrive and sees that the hotel bought Kettal outdoor furniture from Traços for its pool in 2025, and that a lobby and rooftop project is open.
3. It asks Daniel, Tomás and Maya at the same time. The chat shows which specialist is working at each moment.
4. A single proposal comes back, organised by area (Lobby, Rooftop bar), with brands, collections and indicative prices. Each specialist answers from its own catalogue files, so the products are ones Traços actually sells.
5. The Orchestrator offers to log the visit in Pipedrive. The proposal appears as a card in the approval queue. Paulo clicks Approve, and the activity "Visit to Casa Maré Boutique Hotel" appears on the deal in Pipedrive.
6. In #leads, someone asks for hotels in the Algarve that renovated recently or are about to open. The Leads agent returns a table of real hotels, each with the reason it fits and a link to the source. Before anything is created, the app checks Pipedrive. If a hotel is already there, it says so, and no duplicate is created.

What changes for Traços Interiores

Proposals in minutes

A combined proposal for a client visit used to take about a day of emails between colleagues. With Traços HQ the first version is ready a few minutes after the request, and the segment owners only need to check it.

A CRM the team can trust

Only the Orchestrator writes to Pipedrive, and only after a person approves. Visits get logged when they are planned, and new leads are checked for duplicates before they are created. Before a visit, the client history in Pipedrive is the one place to look.

Prospecting every week

Looking for new hotels is now one message in #leads. The team reads the reasons and the sources, picks the hotels worth a visit and approves them into the CRM.

Every segment's knowledge, for everyone

A salesperson who knows outdoor furniture can still give a client a good first answer on acoustics, because Tomás answers from the acoustics catalogues. The segment owner keeps the final word on prices.

People stay in charge

Nothing enters the CRM without a click from a person, and prices are always marked as indicative. The specialists only see the catalogue files of the brands Traços represents, so their suggestions stay within what Traços can actually sell.

What is built and what comes next

Working today, and shown in the video:

• the team chat with channels and @mentions
• the five specialists answering from the brand catalogues
• the Orchestrator reading Pipedrive and combining the specialists' answers
• the approval queue and writing to Pipedrive after approval
• the Leads agent with real web search and the duplicate check

Next steps:

• a Finance agent that reads photos and PDFs of invoices and prepares the expense entry for accounting
• an email agent for each colleague that sorts priorities and drafts replies
• the website voice agents inside the app
• a pilot with the whole Traços team on the real Pipedrive account

Technologies

• Next.js 16 and React 19, with Tailwind CSS, built with Cursor during the two days of the lab
• Grok through the xAI API (model grok-4.7), called with the OpenAI-compatible SDK, for reasoning, tool calling and web search
• Pipedrive REST API v1 for reading clients and deals and for writing notes, activities and leads
• SQLite (better-sqlite3) for messages, approvals and leads
• Knowledge files per segment, built from 19 brand catalogues

The demo runs on a Pipedrive test account with fictional clients. No real client data and no personal data from the web were used. The Leads agent only collects public company information.

Links

• Source code: https://github.com/Wingheaded/tracos-hq
• Demo video (72 s, English voice-over with Portuguese captions): https://youtu.be/aHrjRdx1iGo