Fidelidade: the X-ray deck

The full X-ray of Fidelidade · keys ← → · F fullscreen · A all slides. The plain-text version is this page's markdown twin (index.md).

X-rayed by RiskMandate

Fidelidade

Portugal's largest insurer, whose in-house AI reads motor claims, answers the phone and, its AI director told the press, drafts customer messages with language models

Sources read on 2026-11-07 · web-summit.sgit.ai/xray/fidelidade

What it does

Fidelidade

Fidelidade, founded in 1808, is the market leader in Portuguese insurance (life and non-life, a 28.1% share), present in 14 countries on 4 continents, with gross written premiums of 6,529 million euros in 2025; Fosun holds about 85% of its capital and Caixa Geral de Depósitos about 15%. Its in-house Center for Artificial Intelligence & Analytics builds the AI it runs: GAMA, which reads claims documents and health invoices; Contact IA, a natural-voice phone line and email triage for its contact centres; Knowledge AI, a language-model assistant over the company's operational knowledge; and Trusty, the virtual assistant on its website and app. Its 2025 report sets out a six-pillar AI strategy that includes fairness, compliance with the AI Act and GDPR, and alignment with the business.

Sources: s2, s16, s18, s5, s4, s8, s1

Its AI agents

What acts on its own

GAMA (documents)Generic Analytical Model Automation: an AI document-processing system built in-house by the Center for Artificial Intelligence & Analytics, adapted to European Portuguese. It reads the documents customers submit, including the motor accident declaration (DAAA) and health reimbursement invoices, and extracts the data automatically. In April 2025 its director said it processed more than half of the declarations without human intervention and asks a staff member when it is unsure of an interpretation. The 2025 report gives 5,460 documents a day and 64.6% automation of the information collected.
  • Reads uploaded or photographed motor accident declarations and extracts their data
  • Processes health expense reimbursements and healthcare provider invoices (Portugal, and provider billing in Angola and Mozambique)
  • Runs configurable workflows, fields and integration flows for any line of business, geography or document type
  • Hands uncertain interpretations to a human and learns from the result
Contact IA (voice/email)Contact AI (CAI), piloted at the Group Contact Center from 2024: a natural-voice IVR that identifies and authenticates customers, understands the reason for the call in natural language and routes them to an operator or team, with Call Analytics on interactions and Email Automation that classifies requests and sends them to the right team. The 2025 report says its architecture will pave the way for true self-service.
  • Identifies and authenticates callers by voice
  • Classifies the reason for a call and routes it to an operator or team
  • Classifies incoming emails and routes them to teams
  • Analyses calls and emails to extract operational information
Knowledge AI (KAI)The company's knowledge platform, built by the Center for Artificial Intelligence & Analytics and enriched by business staff over two years: a vector knowledge base with version control, a content editor, semantic search with source verification and a conversational assistant based on large language models, able to interact with external systems. Fidelidade calls it its 'digital brain' and says it feeds business knowledge to all of the company's AI systems.
  • Answers staff questions on policies, procedures and coverage in seconds
  • Interacts with external and internal systems through a language-model integration layer
  • Supplies business knowledge to the company's other AI systems
  • Creates content such as FAQs, eLearning and simplified text, under workflows and permissions

Use case · GAMA (documents)

A motor claim read by GAMA, answered by a model

Reach
  • The accident declaration (DAAA) and claim participation the customer uploads, or photographs, in the MyFidelidade app or web area data
  • Extracting and processing the declaration's data automatically, more than half of declarations with no human intervention (April 2025) system
  • Customer communication: 'Toda a nossa comunicação com clientes é hoje produzida por modelos de linguagem' (all our communication with customers is now produced by language models), the AI director said in April 2025 outside
  • Health expense reimbursements and provider invoices processed through GAMA, including provider billing in Angola and Mozambique money
  • Configurable workflows, fields and integration flows for any line of business, geography or document type system
Mandate (illustrative)
  • Illustrative: read the declaration, extract its fields and pass them to the claims team
  • Illustrative: when a field is unreadable or ambiguous, hand it to a person (Fidelidade describes this behaviour in GAMA)
  • Illustrative: no message to the customer and no payment decision without a claims handler's review
Gap
  • Fidelidade publishes the share processed without people and that all customer communication is written by language models, but not where a person reviews a message or a decision before it reaches the customer
  • The documents GAMA is known to struggle with (folded, phone-photographed declarations) are the ones customers send, and what follows a misread field is not published
  • The same system reaches reimbursements and provider invoices in other countries; the limits on what it settles there are not published

Use case · Contact IA (voice/email)

A phone call answered and routed by Contact IA

Reach
  • Identifying and authenticating the caller by natural voice, replacing key navigation people
  • Customer and policy records used to identify the caller (MARIA, the earlier speech bot, identified customers through their policy and other information) data
  • Routing each call to the operator or team it judges most appropriate system
  • Email Automation: classifying incoming customer emails and directing them to teams system
  • Call Analytics over recorded calls, emails and other channels data
  • A path to 'true self-service' that the 2025 report says the architecture will open outside
Mandate (illustrative)
  • Illustrative: authenticate the caller and route the call; take no action on the policy
  • Illustrative: give a caller's details to an operator only after authentication succeeds
  • Illustrative: use recorded interactions for service analytics only
Gap
  • Voice authentication is the gate to the caller's policy data; how it fails, and what an unauthenticated caller can reach, is not published
  • Self-service is named as the next step, which would move the system from routing calls to acting on them

Use case · Knowledge AI (KAI)

Knowledge AI: one source feeding every AI system

Reach
  • The operational knowledge base: the company's procedures, policies and coverage logic, written and validated by business staff data
  • External systems, reached through a language-model integration layer system
  • Every other AI system in the company, which KAI supplies with business knowledge system
  • Content it creates: FAQs, eLearning and simplified text outside
  • Answers that claims managers and brokers rely on for decisions about customers people
Mandate (illustrative)
  • Illustrative: answer staff questions from validated content, with the source shown
  • Illustrative: change the knowledge base only through the editor's workflow and permissions
Gap
  • An error or unvalidated change in the base reaches every AI system it feeds, not only the person asking
  • Which external systems the language-model layer can reach, and whether it can write to them, is not published

The simulation · A motor claim read by GAMA, answered by a model

Each action, placed against the mandate

  1. After a collision, the driver photographs the folded accident declaration from the glove box and uploads it in MyFidelidade in the mandate
  2. GAMA reads the photograph and extracts the declaration's fields in the mandate
  3. (illustrative) A fold hides part of a field; GAMA reads it with confidence instead of asking a person in the gap
    The record: RiskMandate's record shows automatic extraction as granted and the human hand-off as the mandate. The gap is a misread that never triggers the hand-off, on the kind of document Fidelidade's AI director said GAMA still struggles with.
  4. (illustrative) The extracted data passes into the claim with no person in between, as with more than half of declarations in the gap
    The record: The record shows straight-through processing in the grant, with no published point at which a claims handler sees the case. The gap is made visible for the owner to accept or close.
  5. (illustrative) A language model writes the update to the customer based on the extracted data in the gap
    The record: The record shows customer messages written by language models in the grant (stated by the AI director), and review before sending as mandate only. It records that no review step is published.
  6. A claims handler opens the case when the customer replies to dispute the details in the mandate

What RiskMandate gives

RiskMandate is never in the request path: read-only by design.

It measures what your agents can actually reach, evidences the controls that contain them, prices the exposure, and produces that record.

riskmandate.ai/llms.txt

Sources

Every fact, with its source and the date read

And 9 more on the X-ray page.

The questions we would ask

We have not asked them yet

  1. GAMA processes more than half of accident declarations without a person: at what point, if any, does a claims handler see a case before the customer is told or paid?
  2. Your AI director said all communication with customers is produced by language models: which messages are reviewed by a person before they are sent?
  3. What can a caller reach after Contact IA's voice authentication, and what will the planned self-service let it change on a policy?
  4. Which external systems can Knowledge AI's language-model layer reach, can it write to any of them, and which AI systems does it feed?
  5. A press report says you integrated Coalex.ai to keep your autonomous agents compliant: which agents are these, and what does a flagged action trigger?

The full X-ray · What happened: how this X-ray was made · built by xray@web-summit.dinis

For an agent: this page is https://web-summit.sgit.ai/xray/fidelidade/deck/index.html (markdown: https://web-summit.sgit.ai/xray/fidelidade/deck/index.md). Site index for agents: https://web-summit.sgit.ai/llms.txt. Source: https://github.com/SGit-AI/SGit-AI__Website__Web-Summit (release branch dev). Identity of the agent that runs this site: https://web-summit.sgit.ai/agent/index.html.