In operation, with a public demo

    Axentia answers with your institution's knowledge and records how each answer was built

    Axentia is Datacoder's AI service assistant for RPPS pension institutes (Brazil's public-sector pension schemes), closed pension funds (EFPC) and municipalities. The AKS (Axentia Knowledge Server) defines which of the institution's rules and documents go into the answer, which language model writes it, which systems may be queried and what may be sent to the person being served, and it records each run.

    The demo is public, requires no sign-up and answers questions about Datacoder and Axentia. On deployment, each institution's content is registered in its own base, separate from the demo and from other clients.

    Axentia currently serves people in Brazilian Portuguese only. The AKS does not support other languages at this time, so the demo and the conversations take place in Portuguese.

    Anatomy of an answer

    Governed by the AKS (Axentia Knowledge Server)

    Question from an RPPS member

    Which documents do I need to gather to assess my retirement?

    Topic
    Institution rules
    Safety rules
    Model

    Guided answer

    Axentia organizes the guidance based on the institution's authorized sources and points to the way to reach the team when the case needs analysis.

    Run recorded: content used, passages of rules, queries, instruction to the model, answer and verification codes

    The problem

    The service the institution has to provide

    Members, participants and citizens contact the institution through several channels, with questions that repeat, and the team that answers is often the same one that analyzes the cases.

    The same question, different answers

    People ask by phone, by email and at the counter and receive guidance that does not match. Each staff member reads the rules their own way, and not always in the version in force.

    Repeated questions take up the day

    At a pension institute, documents for retirement and member re-registration. At a pension fund, withdrawal, portability and how the benefit is paid out. At a municipality, the documents and deadlines of a service. These are questions whose answer is already written in a rule, plan regulation or manual, and they take up people who should be analyzing cases.

    A small team and no record of what was said

    In an institution with a small team, the same people handle records, request analysis and service. When a piece of guidance is challenged, there is no record of what was said or of which rule was used.

    What the law requires of public service

    Plain language

    Lei 13.460/2017, the Brazilian law on the rights of public service users, such as people served by municipalities and pension institutes, lists among the guidelines for service plain and understandable language, without acronyms, jargon or foreign terms (art. 5, XIV), and the use of technology to simplify service (art. 5, XIII).

    Accurate information

    The same law gives users the right to accurate and easily accessible information, including on the internet, about the services provided and the status of the cases they are party to (art. 6, VI).

    Automated decisions and personal data

    The LGPD (Brazil's General Data Protection Law), which also applies to pension funds, gives data subjects the right to request a review of decisions made solely by automated processing that affect their interests (art. 20). It also requires public bodies to state how they process personal data and to appoint a data protection officer (art. 23, I and III).

    Sources: Lei nº 13.460/2017, art. 5, XIII and XIV, and art. 6, VI (opens in a new tab); Lei nº 13.709/2018 (LGPD), art. 20 and art. 23, I and III (opens in a new tab).

    Lei 13.460/2017 covers public services, such as those of municipalities and pension institutes. Pension funds are private entities; for them, the reference cited here is the LGPD.

    That is why Axentia gives guidance from the institution's own content, records each answer and refers to the team anything that involves a decision on a benefit or a request.

    Three ways a generic AI fails at an institution's service

    Connecting a generic language model to the institution's channel creates three risks the team does not control.

    It answers confidently and gets it wrong

    A generic model does not know your pension institute's ordinance, your plan's regulations or your municipality's decree. It completes the sentence with what seems likely and may cite a legal article that does not exist, in the same tone as a correct answer.

    It mixes content across clients

    Without separation by institution, content registered by one client can appear in the answer given to another. In pensions and public service, that leak reaches the personal data of the people being served.

    It leaves no record

    Months later, nobody can reconstruct why a given piece of guidance was given. Without a record of the run, there is no way to fix the cause or to show what was said to the person who asked.

    AKS

    The role of the AKS (Axentia Knowledge Server)

    A language model on its own does not identify who is asking, does not keep the history, does not control permissions and leaves no record. The AKS handles these functions before and after the model writes the answer.

    Assistant connected straight to the model

    Axentia + AKS

    The model alone decides what goes into the answer

    The AKS selects the content the model may use and checks the answer before sending it

    One knowledge base for all clients

    One content base per institution, in its own database, with credentials stored encrypted

    Instructions to the model written into the program

    Guidance, content and safety rules with version, history, reason for the change and a verification code

    A single model for any question

    Model chosen by topic, channel, the profile of the person asking and the product

    Unrestricted model access to systems

    A query to a system only runs if the tool is enabled for that institution and that channel

    The text of the question goes to an external provider

    By default, answers are generated on Datacoder's own servers

    What happens between the question and the answer

    Seven steps the AKS runs for every message, before and after the language model writes the answer.

    01

    Identity and institution

    Who is asking, and from which institution

    Access is validated and checked against the stated institution. The AKS opens only that institution's content base.

    02

    Question intent

    Topic identification

    The AKS identifies the intent of the question from the words used. The intent guides which content goes into the answer and which model handles it.

    03

    Institution content

    Rules and documents (RAG)

    Tone guidance, general rules and the content registered by the institution are brought in. A semantic similarity search finds the passages of rules and documents closest to the question.

    04

    Conversation memory

    Recent and summarized

    The latest messages go in full and older ones go in summarized. The person being served picks up the conversation without repeating what they already said.

    05

    Limits and model choice

    Safety rules (guardrails)

    Before generating the answer, the AKS applies the rules on what may and may not be answered. A routing rule chooses the model by intent, channel and the profile of the person asking.

    06

    System queries

    Authorized ones only

    When the answer depends on system data, such as registration status, the model requests the tool. The AKS checks whether it is enabled for the institution and the channel, runs it and returns the result to the model.

    07

    Check and record

    Before the answer goes out

    The answer is checked against the rules before it goes out. If it claims to have taken an action without calling the tool, it is blocked. The whole run is recorded with verification codes.

    The answer rests on the content registered by the institution, the versioned rules, the authorized queries and the record of the run. The language model writes this information as text the person being served can understand.

    What comes ready and what is defined in your project

    The first list shows what Axentia already does in operation; the second, what is set up with each institution.

    Available today, in operation

    Channels

    • Voice assistant that listens and speaks in Portuguese, with an animated avatar, in a widget that fits into any authorized website
    • Full-screen conversation
    • Conversation application programming interface (API) for other systems

    Website reception

    • A greeting specific to each page and resumption of the conversation
    • Contact details recorded only after an explicit "yes" from the person being served, with a signed notice to the team
    • Message taking outside service hours
    • Opening website pages for the visitor and continuing the conversation from one website to another
    • Visitor funnel and report

    In Calyps (RPPS)

    • Reading of data, status, statements, simulations, benefit awards, dependents, requests, documents and messages
    • At the request of the person being served: run a simulation, open and cancel a request, send a message to the institution, add and update a dependent and verify an email address
    • Registration change request, which becomes a request reviewed by the team
    • Link for document upload, without the file passing through the AI
    • Handing the conversation over to the team, with a summary

    Knowledge

    • PDF and DOCX converted into knowledge cards that a person approves before publication
    • Knowledge and rule packages with version, draft, publication, effective period and release by channel
    • Semantic search over registered rules and documents
    • Test rooms before publication

    Quality

    • Each statement in the answer checked against the source
    • Content gaps identified and turned into drafts for review
    • Answers flagged for later review by the team
    • Record of each run, with a verification code for the instruction to the model and for the answer

    Operation

    • By default, answers generated on Datacoder's own servers
    • Automatic failover between model servers, with health monitoring
    • Monthly usage cap per institution, limits per channel and a consumption dashboard
    • Model choice by intent, channel, profile and product

    Defined in each project

    • Registration of the institution's content: laws, plan regulations, ordinances and manuals
    • The institution's own safety rules, added to the central rules
    • Websites and systems authorized to host the assistant
    • Tools enabled in each channel
    • Profiles of the people asking, each with its own content, model and usage limit
    • Connection of other institution systems through the tool protocol
    • Service hours and where notices to the team are sent
    • Monthly usage cap and limits per channel

    “What did Axentia base that answer on?”

    This question comes up when someone reviews a conversation months later. Axentia records each run so that it has an answer.

    Original question and answer delivered
    Final instruction sent to the model (prompt) and its verification code (SHA-256 hash)
    Text generated by the model, before the check
    Content used, with the version and a copy of the text at that moment
    Passages of rules retrieved and the similarity score of each
    Systems queried, with what was requested and what came back
    Safety rules and checks applied
    Model used, run time and volume of text processed (tokens)

    Where Axentia works in the institution's service

    Initial guidance and organization of service, with referral to the team when the case needs analysis.

    First response to incoming requests
    Frequent questions about rules and procedures
    Guidance on required documents
    Help understanding preliminary simulations
    At a pension fund, rules on withdrawal, portability and how the benefit is paid out
    At a municipality, services, documents and deadlines for citizens
    Referral to analysts on the team
    The same guidance on the website and in the systems

    Axentia queries the systems the institution already uses

    In an RPPS, Axentia queries Calyps through a tool protocol, without asking the member for what the pension institute already has on file, and takes actions at their request, such as running a simulation or opening a request. Other authorized systems of the institution can be connected through the same protocol. Each tool is enabled per institution and per channel, and each query is recorded.

    Person being served

    Axentia

    AKS

    Axentia Knowledge Server

    Databases, rules and systems

    Traceable answer

    Governance

    Where the institution's data is kept

    By default, Axentia generates answers on Datacoder's own servers. Each institution has its own database for its content, and the control rules sit in a separate central database.

    Control rules in a central database; each institution with its own database
    Credentials and keys stored encrypted; API access keys stored as hashes
    The language model is not exposed to the internet, and no channel accesses the model directly
    Central safety rules cannot be turned off by a client
    Document ID and card numbers masked before the conversation is stored
    Deletion of a visitor's data on request
    Each answer generates an audit record, with controlled access
    Contact details of the person being served recorded only with their permission
    Deployment

    How Axentia is introduced at your institution

    Deployment consists mainly of turning the institution's rules, plan regulations, workflows and standard answers into versioned content. Datacoder does this work together with your team.

    01

    Service assessment

    Which questions come in, through which channels, from which profiles and, above all, what should not be answered by AI at your institution.

    02

    Rules become content

    Laws, plan regulations, ordinances and manuals, in PDF or DOCX, become knowledge cards that the team approves before publication, with version and effective period, in the institution's base.

    03

    Acceptance testing with your team

    In the test rooms, the team asks real questions and sees the whole path: content used, passages of rules, instruction to the model and answer. Adjustments are made in the content, without changing the system.

    04

    Go-live with records

    The channel goes live with a record of each run and defined usage limits. Gaps identified by the quality evaluation become drafts for the team to review.

    What Axentia does not decide

    Axentia supports the service. It does not replace the technical team, does not decide on benefits or requests, does not grant or deny applications and does not issue legal opinions. To reach the team, in Calyps Axentia hands the conversation over with a summary; on websites, it records the message and, with the permission of the person being served, their contact details for the team to get back to them. Answers can also be flagged for later review by the team.

    Bring a difficult question from your service

    Bring a real question from your institution to the presentation. We show how Axentia would handle it, what it already does in operation and what deployment would look like.

    Talk about Axentia