Skip to content
Giallo

A customer service chatbot that knows when to stop and hand over to your team.

What should a customer service chatbot be able to do?

A good customer service chatbot answers only with the company's rules and data: warranties, service visits, invoices, user instructions. Giallo Studio builds it as part of a team of AI agents with Claude, GPT and Gemini that triage tickets and emails, prepare complaints and keep answers up to date; refunds and exceptions are decided by the team.

Frequently asked questions, straight answers.

Yes, in most cases. The models we use read through typos, abbreviations and messages in multiple languages; when a request remains ambiguous, the chatbot asks a clarifying question or hands over to a person.

No. It can open a warranty claim that falls within the rules already written, but refunds, exceptions and compensation are decided by the team.

If the helpdesk offers APIs, yes. At the start we check what it allows us to read and write and build the connector with minimum permissions; for closed tools we start from the email inbox.

It says so and hands the conversation over to a person, with a summary and the data already collected. It doesn't improvise: it answers only from the articles approved by the company.

It depends on where your customers already write. The same agent can work on both channels with the same rules; WhatsApp has its own rules for messages the business sends first, and we take them into account in the project.

Yes. The agent replies in the customer's language, but the source articles must be approved at least in the main language; for languages the team cannot check, the reply still goes to a person in sensitive cases.

By limiting its sources. The chatbot answers only from approved articles and CRM data, has an explicit instruction not to promise anything that isn't written there, and every reply is logged for review.

The AI agents we design, one by one.

The complaints agent pieces the customer's history together like a constellation: contract, service visit, ticket, chat and email connect up to the complaint; out comes a timeline of what happened, what was promised and what wasn't done, with the possible solutions left to the manager, who decides and signs.

Triage agent

Labels each message by type, urgency, language and customer, and routes it to the right helpdesk queue.

As each email, chat or ticket arrives.

Connects to: Helpdesk such as Zendesk or Freshdesk · Gmail or Outlook shared inbox · Customer management system or CRM

  1. Reads the message and any attachments, such as photos of a damaged product.
  2. Recognizes the request type among those defined by the team, and the customer's language.
  3. Links the message to the contract, serial number or invoice, if the customer included a reference.
  4. Assigns priority and labels and moves it to the correct queue, flagging messages with a tone of strong dissatisfaction.
Claude Haiku
Labels messages in multiple languages, even with photos attached, and handles the volume of a shared inbox at a low cost.

Who approves. The team defines categories and priority rules; messages the agent cannot classify stay in a queue to be read by hand.

Warranty and service request agent

Handles faults, contract changes and invoice questions, collects the data needed and opens the right request, or hands over to a person.

Anytime on the website and on WhatsApp, even outside business hours.

Connects to: Website chat · WhatsApp Business Platform · CRM or contract management system · Knowledge base

  1. Introduces itself as an automated assistant and says when a person will reply.
  2. Verifies the customer's identity with a customer code or serial number and the registered email, before showing any data.
  3. For a fault, it asks for photos, the serial number and proof of purchase and checks coverage against the warranty terms written by the company.
  4. For a change of address or account holder, or an invoice line item, it prepares the request with the documents already attached.
  5. Hands over to a support rep if the customer asks, if they are angry or if the question has no approved answer.
Claude Sonnet
Keeps the brand's tone in long conversations and follows precise rules on what not to promise, even when the customer insists.

Who approves. The manager approves the articles the agent answers from and the coverage rules; out-of-warranty repairs, invoice corrections and exceptions go to a person.

Complaints agent

Rebuilds the customer's history and prepares a draft reply with the possible solutions allowed by company policy.

When a formal complaint arrives or a customer comes back for the third time about the same problem.

Connects to: Helpdesk · CRM or contract management system · Company complaints policy

  1. Gathers the same customer's contracts, service visits, tickets, chats and previous emails.
  2. Writes a concise timeline: what happened, what was promised, what wasn't done.
  3. Compares the case with the complaints policy and lists the possible solutions, without choosing one.
  4. Prepares the draft reply for the manager, who decides and signs.
Claude Opus
Complaints are few but long and tangled: you need the model that best handles many conversations at once without losing details.

Who approves. Refunds, replacements and compensation are always decided by a person; the agent prepares the file and the draft.

Knowledge base agent

Takes the support rep's correction, finds the article the wrong answer came from and immediately proposes an edit to that article.

Every time a support rep rewrites a chatbot reply by hand.

Connects to: Helpdesk · Website help center · Custom connector

  1. Compares the chatbot's reply with the one rewritten by the support rep.
  2. Traces back to the knowledge base article used and pinpoints the outdated or missing sentence.
  3. Writes the new version of that single article, along with the conversation that made it necessary.
  4. Publishes to the helpdesk only after the manager's OK; until then, the chatbot hands that topic over to a person.
GPT
Compares two short texts and returns, in a structured format, the article involved and the proposed edit, easy to check.
Codex
Writes the helpdesk connection to read tickets and publish approved articles; it does not touch customer data in day-to-day work.

Who approves. No article changes in the help center without the OK of the customer service manager.

Channels, helpdesk and the heaviest queue

Tell us which channels requests come from and which helpdesk the team works in: we'll read a week of tickets with you and choose which queue to start the agents on.

Prefer the full form? Go to the contact page.

No spam or sales pitches. Response within 1 business day.

We use this data to contact you and, if we work together, to manage the project. We store it in Giallo's internal management system, on servers in the European Union, where only our team can see it; by email we only receive a notice, without your data. If you do not become a client, we delete it 3 months after our last contact. Details and your rights: privacy policy.

Do you prefer to write or call us?

Where time gets lost today.

These are the points we hear about most often. Each one is a possible first agent.

The shared inbox is one big pile

Warranty faults, pre-sales questions, invoices to correct and complaints all land in the same inbox. Every morning someone reads them all to work out who handles what.

Every message arrives already labeled by type, language, urgency and linked contract, in the right person's queue.

Warranties, contract changes and invoices: always the same checks

For a fault, the support rep looks up the serial number and purchase date; for a change of address or account holder, they open the customer record; for an invoice that doesn't add up, they compare line items and rates. The same routine repeats all day.

The chatbot collects the serial number, photos and receipt, checks coverage against the written rules and prepares the service or change request. Staff are left with the out-of-policy cases and genuine disputes.

A complaint means rebuilding the whole history

To reply well, you have to dig up contracts, service visits, previous emails, chats and internal notes, scattered across different tools.

An agent pieces the customer's history together in a single record and proposes a reply that the manager reviews and signs.

Standard replies go stale

The helpdesk macros were written years ago, warranty terms and rates have changed, and everyone fixes them by hand in their own way.

When a support rep corrects a chatbot reply, the article it came from goes back to the manager with the change already proposed.

Which AI model we use, and for which task.

We don't tie you to a single provider: we work with Claude by Anthropic, GPT and Codex by OpenAI, and Gemini by Google, and we choose the version during testing, on your own cases.

Claude Haiku

  • Triage agent. Labels messages in multiple languages, even with photos attached, and handles the volume of a shared inbox at a low cost.

Claude Sonnet

  • Warranty and service request agent. Keeps the brand's tone in long conversations and follows precise rules on what not to promise, even when the customer insists.

Claude Opus

  • Complaints agent. Complaints are few but long and tangled: you need the model that best handles many conversations at once without losing details.

GPT

  • Knowledge base agent. Compares two short texts and returns, in a structured format, the article involved and the proposed edit, easy to check.

Codex

  • Knowledge base agent. Writes the helpdesk connection to read tickets and publish approved articles; it does not touch customer data in day-to-day work.

How we choose between models: the best AI model for agents.

How we work with you.

  1. A week of tickets, read together

    We take a week of real, anonymized messages and sort them with the team: what can be closed automatically, what always goes to a person.

  2. Approved answers before the chatbot

    Before writing a single line of instructions, we fix the articles the chatbot answers from: warranty terms, rates, service procedures, user guides.

  3. Connections to helpdesk and CRM

    We connect the helpdesk, CRM or contract management system and channels with read-only permissions where that's enough, and write access only where needed.

  4. Shadow mode before going live

    For a while the agent writes replies without sending them: the team compares them with their own and decides when to open the channel to it.

AI agents for customer service and customer support

Which queue would you like to find already in order tomorrow morning?

Tell us which channels requests come from and which helpdesk the team works in: we'll read a week of tickets with you and choose which queue to start the agents on.