Consulting

Artificial Intelligence consulting for software companies

This is not about adopting AI because everyone else is. It is about deciding where it produces results in your product and your team — and where it only adds cost and risk.

Every software company is under pressure to “use AI”. The pressure comes from the market, from clients, from investors and, frequently, from the team itself.

The problem is that most of that pressure does not come with a useful question attached. “How do we use AI?” is a bad question. “What in our product or our process measurably improves with AI?” is a question that can be answered — and whose honest answer is sometimes “nothing, for now”.

This consulting exists to reach that answer methodically, and then to execute it.

Quatro frentes de aplicação

A consultoria atua onde a IA produz resultado mensurável — não onde ela apenas impressiona.

Software novo

desde a primeira linha

  • Definir onde a IA entra no ciclo de desenvolvimento da equipe
  • Prototipagem e geração assistida com revisão humana obrigatória
  • Testes, documentação e revisão de código apoiados por modelos

Software legado

conforme linguagem e ambiente

  • Leitura e mapeamento de bases sem documentação
  • Novas funcionalidades sem reescrever o que funciona
  • Modernização por partes, com risco controlado

Agentes de BI

dentro do próprio sistema

  • Pergunta em linguagem natural, resposta em número e gráfico
  • O agente lê estrutura, gera SQL e nunca vê o dado bruto
  • Custo previsível e limitado por cliente

Chatbots

atendimento e operação interna

  • Base de conhecimento própria, não respostas genéricas
  • Integração com o sistema que a empresa já usa
  • Escalonamento para humano quando o bot não deve decidir

The four tracks

New software: AI from the first line

Building a product today with the process of five years ago leaves productivity on the table. But adopting assisted generation without criteria produces something else: a codebase nobody understands, written fast and maintained slowly.

The work here is defining where AI enters the team’s cycle, and with which safeguards:

  • Prototyping and exploring technical alternatives before locking the architecture
  • Assisted code generation, with mandatory human review and clear rules on what is never generated unsupervised
  • Automated tests, documentation and code review supported by models
  • Project conventions that keep the code readable by people and by models alike

The real gain is not writing faster. It is spending more time on the decisions that matter and less on the repetitive work around them.

Legacy software: understand before touching

This is, in practice, the most common scenario — and the one the market serves worst.

Systems in production for ten, fifteen, twenty years, that work, that sustain real clients’ operations, and that nobody wants to touch because the documentation does not exist and whoever wrote it has left the company.

Language models are remarkably good at one specific task: reading code nobody understands and explaining what it does. That changes the economics of legacy maintenance.

  • Mapping undocumented codebases: flows, dependencies, side effects
  • Rebuilding documentation from the code itself
  • Impact analysis before changing anything: what else depends on this?
  • New features attached to what exists, without rewriting what works
  • Modernisation in parts, with controlled risk and a way back

An honest caveat: results depend on the language and the environment. Codebases in widely represented languages — C#, Java, Python, JavaScript, PHP, SQL — do very well. Old or niche languages, closed proprietary environments and tool-generated code do less well. We assess that before promising anything, and we say when it is not worth it.

BI agents inside the product

This is the track with the greatest potential for competitive differentiation — and the one most frequently done wrong.

The promise is simple and powerful: the user of your system asks “who were my ten best clients this quarter and how does that compare with last year?” and gets the answer as a number and a chart, without opening a report, without asking IT, without exporting a spreadsheet.

The naive implementation — wiring a model straight to the database — is a disaster waiting to happen: sensitive data leaving the infrastructure, unpredictable cost, SQL generated without control, and write access at risk.

Como funciona um agente de BI seguro

O modelo de linguagem enxerga a estrutura do banco. Nunca o conteúdo.

dentro da sua infraestrutura Usuário pergunta em linguagem natural Agente (LLM) traduz a pergunta em consulta Catálogo schema, views e exemplos SQL gerado Views de negócio usuário somente leitura Resultado agregado → tabela, gráfico ou resposta resposta O que o modelo NÃO recebe ✕ tabelas brutas ✕ dados de clientes ✕ acesso de escrita ✕ o banco inteiro
  1. 1
    Usuário pergunta em linguagem natural
  2. 2
    Agente (LLM) traduz a pergunta em consulta
  3. 3
    Catálogo schema, views e exemplos — a única coisa que o modelo lê
  4. 4
    SQL gerado consulta as views de negócio
  5. 5
    Views de negócio usuário de banco somente leitura
  6. 6
    Resultado agregado tabela, gráfico ou resposta em texto
  7. ✕
    O que o modelo nunca recebe tabelas brutas · dados de clientes · acesso de escrita · o banco inteiro
The model reads structure, not content
It receives the schema, the business views and example questions with their queries. It never receives raw client data.
Business views, not tables
The agent sees a curated catalogue organised by subject — revenue by period, sales by region, overdue accounts, occupancy — not the internal structure of the database. That improves answer quality and drastically reduces the risk surface.
Read-only database user
Restricted to the views schema. It is not a matter of trusting the model; it is a matter of the architecture making writes impossible.
Only the aggregated result comes back
The model receives consolidated numbers, never sets of records.
Predictable cost
Limits per client and per period, with consumption metering. A transparent cost model, with no surprise at the end of the month.
No training a model of our own
The approach is text-to-SQL with well-built context. Training a model from scratch for this is expensive, slow and almost always worse.

Chatbots that actually resolve

The market is saturated with chatbots that answer nothing. They fail for two reasons: they do not know the business and they have no access to the system.

A useful assistant needs three things:

  • Its own knowledge base — the company’s documentation, procedures and rules, not the model’s generic knowledge
  • Integration with the system — look up an order, check a balance, open a ticket, not merely talk about it
  • Explicit limits — knowing what it must not answer, and escalating to a person when the decision has consequences

We apply this both in customer service and internally, where the return is often higher and the risk lower: first-line support, process questions, technical documentation lookup.

How the consulting runs

  1. Assessment

    Understanding the product, the team, the codebase and where the company stands.

  2. Opportunity map

    What makes sense, what does not, and in what order. With effort and return estimates.

  3. Proof of concept

    One real, small, measurable application before any larger commitment.

  4. Implementation and training

    Built alongside your team, not instead of it. The goal is for them to carry on alone.

Want to assess where AI makes sense in your product?

Book a call