A4IT
Services / AI consulting
01 / 08
AI consulting · Strategy, build and adoption

AI that does a job. Not a slide about one.

Most companies do not need an AI strategy of eighty pages. They need to know which three tasks are worth automating, what it costs to run, where their data goes, and who checks the result. A4IT answers those questions and then builds the answer into the systems you already use.

Plan an AI intake → What we deliver How an engagement runs

What you get, what stays yours, where it stops

Said before the first meeting
You get
  • A ranked list of use cases with effort, running cost and risk per case
  • A working pilot on your own data, not a demo on ours
  • Integration into your application, intranet, mailbox or workflow
  • Documentation, a handover session and training for the people who use it
Stays yours
  • Your data, your prompts, your evaluation sets and your source code
  • The accounts with model providers: opened in your name, billed to you
  • The choice of model, which can change without rebuilding the solution
  • The decision what a person must approve before anything is sent or booked
Boundaries
  • No promised accuracy figure before we have measured on your documents
  • No legal opinion: we map obligations, your counsel signs them off
  • No training of foundation models; we use and adapt existing ones
  • No automation of decisions about people without human review

What we deliver

Six building blocks · take one or all
01 · Start here

Readiness check and use-case selection

Two to three working sessions with the people who do the work, followed by a short written report. The result is a decision, not a vision.

Process walk-through
Where does time go? We follow real cases through mailbox, ERP, spreadsheets and paper, and note every step that is reading, typing over, searching or summarising.
Data inventory
Which information exists, where it lives, who may see it and how clean it is. Personal and confidential data are marked from the start.
Scoring
Every candidate is scored on value, feasibility, running cost and risk. You see why a case is first on the list and why another one is dropped.
Build, buy or wait
Sometimes a feature in software you already pay for is enough. Sometimes the right advice is to wait six months. We say so.
02 · Build

LLM integration in your own software

Language models as a component of an application, with the same engineering standards as the rest of the code: versioned, tested, logged and replaceable.

Text in, structure out
Classify incoming mail, extract fields from documents, summarise files, draft replies and translate, with output in a fixed schema your system can check.
Provider-neutral design
One internal interface in front of the model, so a hosted model can be swapped for another one, or for a local model, without touching the application.
Evaluation before release
A test set built from your own cases. Every prompt or model change is measured against it, so quality is a number and not an impression.
Cost and rate control
Token budgets, caching, batch processing where it fits, and a dashboard that shows what each function costs per month.
03 · Build

Answers from your own knowledge

Retrieval-augmented generation: the model answers from your manuals, contracts, tickets and procedures, and shows the passage it used.

Ingestion
Documents from file shares, SharePoint, wikis, databases and mail archives are converted, split and indexed, with a schedule that keeps the index current.
Access rights respected
A user only gets answers from sources that user may open. Rights are checked at query time, not copied once and forgotten.
Sources with every answer
Each answer links to the document and paragraph it rests on. No source, no answer: the assistant says it does not know.
Where people already work
As a search box on the intranet, a panel in your application, or a chat in the tools your team already has open.
04 · Build

Agents and workflow automation

A model that can use tools: look something up, fill in a form, create a ticket, prepare an order. Useful, and only safe with clear limits.

Scoped tools
An agent gets exactly the actions the task needs, each with its own permission. Reading is separated from writing.
Human approval points
Anything that sends, pays, deletes or commits waits for a person. The approval screen shows what will happen and why.
Audit trail
Every step an agent takes is logged with its input, its output and the tool it called, so an incident can be reconstructed.
Open protocols
Tools are exposed through standard interfaces such as REST and the Model Context Protocol, so they work with more than one assistant.
05 · Run

Local and self-hosted models

For data that may not leave the building, or volumes where a meter per token becomes expensive.

Sizing
Which model fits which task, and what hardware that takes. Often a small model on a modest server is enough for classification and extraction.
Deployment
Open-weight models on your own server or in a European cloud, behind your firewall, monitored like any other service.
Hybrid routing
Sensitive requests stay local, the rest may go to a hosted model. The rule is configuration, visible and auditable.
Honest comparison
We measure the local model against a hosted one on your test set and show the difference in quality, speed and cost.
06 · Govern and adopt

Governance, EU AI Act and training

The part that decides whether a pilot becomes daily practice.

AI policy for staff
A two-page policy people actually read: which tools are allowed, what may be pasted into them, and who to ask.
Register and risk mapping
A register of the AI systems in use, each mapped to the risk categories of the EU AI Act and to GDPR obligations, as input for your legal adviser.
Training by role
Hands-on sessions with your own cases: for management, for office staff and for developers. Half a day each, not a lecture.
Review rhythm
A quarterly check of quality, cost and usage, and a decision per function: keep, improve or switch off.

How an engagement runs

Small steps · a decision after each
  1. Intake
    A call of thirty minutes: your question, your systems, your constraints.
    30 minutes
  2. Readiness check
    Sessions with the team and a written shortlist of use cases.
    Report + decision
  3. Pilot
    One use case, real data, a test set and measured results.
    Working software
  4. Integration
    Into your application and processes, with approval points and logging.
    In production
  5. Adoption
    Training, policy, monitoring and a review every quarter.
    In daily use

Technology

Chosen per case · nothing is mandatory
ModelsHosted large language models from the main providers, and open-weight models run locally
TechniquesStructured extraction, classification, retrieval-augmented generation, tool use, machine learning and NLP
ApplicationJava and Spring Boot, .NET, JavaScript, REST APIs, microservices
DataPostgreSQL, SQL Server, vector search, ETL from existing systems
ProtocolsREST, OpenAPI, Model Context Protocol for agent tools
HostingYour own servers, Jelastic PaaS, Azure or a European cloud of your choice

Good fit, and not a fit

Saves both of us a meeting
A good fit when
  • People spend hours reading, retyping, searching or summarising
  • You have documents and data, but no time to turn them into answers
  • You want advice and build without a handoff in between
  • Data location and cost per month matter to you
Not a fit when
  • The goal is a chatbot because competitors have one
  • Nobody in the business can own the process being automated
  • You need research into new model architectures
  • A human may never look at the output again

Questions about AI consulting

All questions →
01Does our data end up in someone else's model?

Only if you decide so. We set up provider accounts in your name with the settings that exclude your data from training, and for sensitive material we run models locally. Which data goes where is written down before the pilot starts.

02How accurate will it be?

We do not know before we measure, and nobody else does either. A pilot starts with a test set from your own cases; the result on that set is the figure you decide on.

03What does it cost to run?

Running cost is part of the use-case score. For hosted models it depends on volume and model choice; for local models on hardware. You get an estimate per function per month, and a dashboard that shows the real figure afterwards.

04We have no developers. Can we still use this?

Yes. A4IT builds, hosts and maintains the solution if you want that, and trains your staff to use it. See software development and cloud & PaaS.

05Are you tied to one AI vendor?

No. We keep the model behind an interface so it can be replaced, and we have no reseller agreement that makes one provider more attractive to us than another.

06What about the EU AI Act?

We list your AI systems, map each one to the Act's risk categories and note the obligations that seem to apply. That is preparation for your legal adviser, not a replacement.

Bring one process that eats time. We will tell you if AI helps.

Reply within three business days.Plan an AI intake →How we work