Skip to main content

Kodex

Technology solutions

LLM & AI Assistants

We design augmented workstations: the assistant understands your domain, cites sources, and acts on real systems — with identity, guardrails, and ISO 27001.

See how it works Book a consultation

What it is and why it matters

From chatbot to operational collaborator

A bare LLM answers. A Kodex assistant collaborates: memory (RAG), tools (MCP), identity policies, and UX designed for adoption — from the floor to the C-suite.

  1. 01

    We define the role

    What the assistant must do, for whom, with which permissions, and which adoption KPIs.

  2. 02

    We anchor the truth

    RAG with citations and abstention when there is no evidence — less hallucination, more trust.

  3. 03

    We enable action

    MCP and guardrails: it can query and execute only what is allowed, with an auditable trail.

Generic chatbot

  • Generic answers without your jargon
  • No access to live data or systems
  • Hallucination risk without citations
  • Low adoption: ‘just another chat’

Kodex assistant

  • Your business terminology and roles
  • Document memory + MCP tools
  • Guardrails, identity, and evals
  • UX that reduces cognitive load

Live flow

Del mensaje al puesto aumentado — con memoria y gobierno

Two lanes: the augmented-seat interface and the runtime with memory, tools, and guardrails.

Flujo del asistente: usuario, modelo, guardrails, RAG, MCP y respuesta con evidencia

01 · Interfaz del puesto

  1. user

    Usuario / rol

    sso · canal

  2. assistant

    Asistente

    llm · sesión

  3. guardrails

    Guardrails

    policy · PII

02 · Runtime con evidencia

  1. memory

    RAG

    corpus · citas

  2. tools

    MCP tools

    read · write

  3. policy

    HITL / scopes

    allow · approve

  4. output

    Respuesta

    cita · acción · audit

Interfaz del puesto · luego runtime con RAG, MCP y respuesta gobernada.

Depth

Choose your assistant level

Three depths. Start with an evidence-backed copilot and grow to multi-role with governed execution.

  1. 01
    Copilot Answers with citations
  2. 02
    Operational Acts with tools
  3. 03
    Multi-role Several seats
  4. 04
    Governance Evals · identity
  1. Assistant with cited answers

    01

    Nivel 01

    Evidence-backed copilot

    Answers well — and says from where

    Ideal for support, legal, or internal technical teams: RAG, mandatory citations, and abstention policy.

    Qué incluye

    • 1 assistant role
    • Bounded RAG corpus
    • Citations + abstention
    • Basic observability

    Ideal para · Teams that need truth before autonomy.

    Resultado · Fewer repetitive tickets · more trust in the answer.

  2. Assistant with governed tools

    02

    Nivel 02

    Operational assistant

    Reads, writes, and escalates to a human

    The assistant uses MCP tools (CRM, helpdesk, ERP) with scopes and HITL on critical actions.

    Qué incluye

    • Copilot + MCP tools
    • Identity and roles
    • Approval gates
    • Quality evals

    Ideal para · Ops and CX that want to close the loop in the same thread.

    Resultado · Fewer clicks · more cases resolved with context.

  3. Augmented workstations at scale

    03

    Nivel 03

    Recomendado

    Industrial augmented seat

    Multi-role · multi-model · perimeter

    Multiple profiles (legal, procurement, plant), cloud or on-prem, observability, and ISO governance.

    Qué incluye

    • Everything operational
    • Multi-role / multi-channel
    • Cloud or on-prem
    • Adoption program

    Ideal para · Product and IT industrializing AI in day-to-day work.

    Resultado · A system of assistants — not N orphaned pilots.

The model is the engine; RAG, MCP, and guardrails are the chassis. Without them, you only have a chat.

Book assistant design Ver flujo del asistente

Corporate assistant

How AI collaborates with your team

From user message to action: identity, model, memory, tools, and policy. Business sees adoption; engineering sees a governed runtime.

Engine
LLM
Memory
RAG
Tools
MCP
Adoption
UX

What the client gains

We do not sell ‘a chatbot’: we sell operational capacity with control.

01

Less recurring work

The assistant resolves the repeatable and escalates the critical with full context.

02

Defensible answers

Citations, abstention, and logs: useful for legal, quality, and ISO.

03

Action on the real stack

Updates CRM, opens tickets, or queries ERP — via MCP, not copy/paste.

04

Measurable adoption

UX and usage KPIs: if no one opens it, there is no ROI — we design so they do.

Dual reading

Same interface, two languages.

  • Negocio

    It speaks like our team

    Ingeniería

    System prompts · glossary · sector few-shots

  • Negocio

    It does not invent when it does not know

    Ingeniería

    RAG + citations · refusal policy · grounded evals

  • Negocio

    It can do what is allowed

    Ingeniería

    MCP tools · scopes · HITL on writes

  • Negocio

    We know what happened

    Ingeniería

    Traces · feedback loops · quality per case

From forgotten pilot to adopted seat

Illustration: how value redistributes when the assistant stops being a demo and becomes real work.

  • Day-to-day queries 100%
  • Resolved with evidence ~65%
  • With system action ~35%
  • Escalated to human ~18%
  • No trail / no policy ~2%
AI assistant adoption in the workplace
Assistant value is measured in cases resolved and trust — not tokens generated.

Casos de uso

Cloud or on-premise assistant deployment

01

Local or cloud inference

Business: you choose sovereignty. Technical: OpenAI, Anthropic, Llama — same product contract.

Multi-role assistants

02

Multi-role agents

Business: one system, several seats. Technical: profiles, prompts, and tools per role.

Assistant executing governed actions

03

Action execution

Business: reports, CRM, alerts. Technical: MCP + guardrails + audit trail.

Relacionado

Why the assistant protects margin and adoption

A chat without memory or tools is noise. An augmented seat with governance is scalable capacity.

Business

Capacity without linear headcount

The team focuses on exceptions; the assistant closes the recurring work with evidence.

Engineering

Runtime + product

Model, RAG, MCP, evals, and UX: a versionable system — not a prompt in a wiki.

Illustration · Operational trust

+38 pts

illustrative trust vs generic chatbot without RAG or guardrails

Assist DS
Reported hallucinations −55%

↓ 55 % vs unanchored LLM

Cases with evidence +38 pts

↑ 38 pts vs unanchored LLM

Answer confidence index (0–100)

Illustrative example: trust rises when moving from a bare LLM to an assistant with RAG, citations, and tools.

Impact on capacity and risk

Each lever brings the assistant closer to real work — with less risk and more adoption.

  1. 01 −70% wait

    Time-to-answer

    Anchored answers in seconds, not searching across 4 systems.

  2. 02 Grounded

    Evidence

    Citations and abstention when the corpus does not cover the case.

  3. 03 Scopes

    Safe action

    MCP tools with scopes and HITL on writes.

  4. 04 UX

    Adoption

    Clear UX and roles: the assistant enters the workflow.

Cycle in days

20d → 6d

illustrative time from pilot to first seat in production

  • Role, KPIs, and permissions 4d → 1.5d

    Antes

    Con Kodex

  • Corpus / tools / prompts 8d → 2.5d

    Antes

    Con Kodex

  • Guardrails and evals 5d → 1.2d

    Antes

    Con Kodex

  • UX and go-live 3d → 0.8d

    Antes

    Con Kodex

Illustrative values. In project we measure your case backlog and the post-assistant delta.

How we deliver

Methodology: from pilot to augmented seat

From the business role to the assistant in production — with clear deliverables for product and engineering.

  1. 1–2 wk

    Fase 1

    Seat consulting

    Who, what, and with which permission

    We define roles, channels, adoption KPIs, and the human/assistant boundary.

    • Role map
    • Permission matrix
    • Copilot / operational scope
  2. 2–3 wk

    Fase 2

    Risk audit

    Data, hallucination, and PII

    We assess corpus, guardrails, hallucination risk, and identity integration.

    • Data gap
    • Abstention policy
    • Evals plan
  3. By scope

    Fase 3

    Assistant implementation

    LLM + RAG + MCP + UX

    We deploy the runtime, connect memory and tools, and tune the adoption interface.

    • Assistant in prod
    • Traces and feedback
    • Case playbook
  4. Ongoing

    Fase 4

    Continuous improvement

    The assistant learns from real use

    Evals, prompts, corpus coverage, and new roles based on adoption.

    • Quality dashboard
    • Role roadmap
    • Periodic retuning

Ver metodología completa →

Trust perimeter

Identity, guardrails, and evidence

The assistant inherits your security policy — it is not an anonymous widget on the intranet.

  • Identity

    SSO and roles

    Every session knows who the user is and what they can request or execute.

    IAM

  • Risk

    Guardrails and HITL

    PII filters, abstention, and human approval on critical actions.

    Policy

  • Evidence

    Citations and traces

    Where the answer came from and which tool was invoked — retained and auditable.

    Audit trail

FAQ — product and operations

Can it replace a team?

No. It amplifies capacity: it resolves the recurring work and escalates the critical to humans with full context.

How do we avoid invented answers?

With RAG, mandatory citations, abstention policies, and groundedness evals in the pipeline.

Integration with internal tools?

Yes, via MCP and governed connectors — the assistant operates on your real stack, not only PDFs.

Cloud or on-premise?

You choose sovereignty: cloud for elasticity or on-prem / VPC for sensitive data — same product design.

Filtro de oro

What level of autonomy do you need?

Option A · Corporate assistant (DFY)

Design, deployment, and operation of the assistant with RAG, MCP, and guardrails.

Book assistant design

Option B · Explore models (DIY)

Review guides and resources to inform the stack decision.

View resources

Más soluciones

Turn chat into an augmented seat

Kodex assistants: LLM with memory, tools, and governance — so your team does more with less risk.

Book a consultation View solutions catalog

Let's collaborate

Tell us your goal

A short triage to qualify your request. In a few minutes we reach the right scope.

1

How would you like to collaborate with Kodex?

We open the right path — no unnecessary questions.

How would you like to collaborate with Kodex?

Tap a card to continue