MSV Labs Enterprise

AI agents inside your systems. Under your control.

For banks, distribution and wholesale, we build AI agents that take over reading, checking and preparation, while decisions stay with your people. Delivered in your infrastructure, with evidence your risk and compliance teams can assess.

On-premise or private cloud in the EUAudit trail for every caseData security and agent permissions
35+years of experience

An agent is only as good as its integration. For more than 35 years MSV sustavi has built ERP, reporting, sales and production systems and implemented them in large organisations. We know what real data, exceptions and permissions look like.

  • Integrations with external systems
  • Implementations in large organisations
  • ERP and custom reporting
  • Sales and CRM systems
  • Production and dispatch
  • Cloud solutions

Use cases

The agent does the preparation. Your people keep the decision.

We don’t build an “agent for the whole company”. We pick a few high-volume processes with clear rules, where an agent measurably shortens the path to a decision.

We automate workflows and process data as standalone projects or combine the two when needed.

Completeness check of credit applicationsThe agent checks the file against internal rules and prepares a list of gaps for the analyst.
Case summary for the approverFacts from the documents, each with its source, on a single page.
Sorting client requestsIdentifying the request type and required documents, and drafting a reply.
The decision stays with a person

Credit decisions, creditworthiness assessment, limit changes and payment orders.

What a project looks like

From the first meeting to an agent in production. Step by step.

We contract an enterprise project as a whole, with agreed scope, timeline and budget. Within it are checkpoints where we review the results together and adjust the scope if needed.

  1. 1–2 meetingsIntro meeting

    Usually at your office. We go through candidate processes, volume, systems and who approves budget and suppliers.

    Your team
    sponsor, process owner, IT
    Result
    project proposal with a budget range
  2. 3–5 weeksDiscovery

    Analysis and specification: workshops, observing the work, 30–50 real cases, architecture and success criteria.

    Your team
    process owner, IT, security, risk
    Result
    specification, plan and a fixed price for the rest
    Checkpoint: signed scope and quality threshold
  3. 4–8 weeksBuild

    We build the agent in your test environment and test it on historical cases. Weekly demo for the process owner.

    Your team
    process owner, IT
    Result
    eval report: test results per category of work
    Checkpoint: threshold met, security review passed
  4. 4–8 weeksPilot

    Shadow mode: the agent proposes on new cases while your people work as before. Comparing the two, we measure accuracy and savings.

    Your team
    users, process owner, risk
    Result
    measured accuracy and savings per category
    Checkpoint: production approval per category
  5. 2–4 weeksRollout

    Release through your change management, monitoring, kill switch and a manual fallback drill, with extra support in the first weeks.

    Your team
    IT, users
    Result
    agent in daily use and handover
  6. monthlyMonitoring

    We measure quality on new cases, tune the agent and report. Every quarter we review the results with you.

    Your team
    process owner, risk
    Result
    monthly report, incident response
Alongside the project: vendor onboardingThe process your organisation uses to assess and approve a supplier: NDA, security questionnaire, data processing agreement and, if you are a financial entity, the clauses DORA requires. We start right after the intro meeting, because we only see real data once it’s complete.
Not sure which process yet?We can contract discovery on its own. You get a specification and an assessment you can use with or without us.
Typically 4–6 months from the intro meeting to the first agent in production.

Agent autonomy

The agent starts by proposing. It does more only once it proves itself.

Autonomy levels define what the agent may do without a person. They’re about permissions, not the model, and they’re set separately for each category of work. A category moves up a level only once results on your cases pass the agreed threshold and your risk function approves it in writing.

Level 0ProposesDrafts, extracts, proposals. Sends nothing and writes nothing to your systems. Every agent starts here.
Level 1PreparesPrepares an order or message in your system, and a person approves it before anything goes further.
Level 2Executes narrowlyPerforms only reversible, low-risk actions on its own, within agreed limits and with after-the-fact review.
NeverThe decision stays with a personPayments, trading orders, limit changes and decisions about clients.

Example: with incoming invoices, the agent can sort invoices by type on its own (level 2), while the posting proposal still waits for the accountant (level 1).

How we build agents

Six rules that apply to every project.

What makes an agent reliable isn’t the model, but the system built around it.

Workflow before agentIf the steps are known upfront, we build a workflow. An agent only where the sequence truly depends on the content.
Control flow lives in our codeThe model proposes the next step, the code decides whether it may run. No black box.
Small, focused agentsOne agent, one business capability, a small set of tools. Easier to test and to audit.
Evals on your dataWe test the agent on your real cases with known outcomes, per language and category of work, and classify errors so we know what to fix. The results go into a report for your risk function.
Swappable modelsEvery call goes through a model gateway, a single point for all model calls. A new model passes the same test before release, without rebuilding the system.
Review is measured tooWe track how often people change proposals and whether they catch errors, because approval has to mean something.

Delivery

The agent runs where your IT already runs.

We choose the architecture during selection, together with your IT and security teams. For most organisations that’s the cloud they already use. The agent’s code is the same in every option; only where the model runs changes.

Most common choiceIn your cloud tenantThe model runs through your existing subscription and agreement with your cloud provider, in an EU region.e.g. Azure OpenAI (EU Data Zone), AWS Bedrock (Frankfurt), Google Vertex AI
For the strictest requirementsOn-premise with an open modelThe agent and model run on your servers. This needs GPU infrastructure, and we test the model’s quality in Croatian before you decide.e.g. Mistral, Llama or Qwen on your GPU servers
When your rules allow itAn AI provider’s APIThe fastest start: direct model calls under a data processing agreement, EU processing and no training on your data.e.g. Anthropic, OpenAI or Mistral

Security and compliance

We build the documentation your control functions need alongside the system.

Guardrails, the built-in limits on what the agent may do, live in code, not in the model’s instructions. Every case has an audit trail, a record of every step with the model and rule versions. The agent can be stopped with a single kill switch, and work continues manually under an agreed procedure. When we use an AI model, we send it only the data it needs for the task. We omit or mask any sensitive data it doesn’t need.

System and data flow descriptionWhat the agent reads, where it sends data and who has access.
AI Act applicability assessmentDone with your compliance team, for the specific use case.
DPIA supportYour DPO leads the impact assessment; we provide the technical input.
Eval reportsQuality per category of work and per language, before and after rollout.
Runbook and manual procedureWhat happens when the agent stops and who takes over the work.
Exit planHow you take over the configuration, prompts, eval set and code.

Regulations we check with you: GDPR · EU AI Act · DORA · NIS2. For your supplier assessment we answer your security questionnaire and adapt the contract to your requirements.

Glossary

Terms we use.

These are the terms your IT, security and compliance teams use. Here is what each of them means in short.

Discovery
The analysis and specification phase: what we build, how we measure success and what it costs.
Shadow mode
The agent works on real cases but only proposes, while people work as before.
Eval
Systematic testing of the agent on examples with a known correct outcome.
Audit trail
A permanent record of every step and decision, with the model and rule versions.
Guardrails
Built-in limits that stop the agent from doing what it must not do.
Model gateway
A single point all model calls go through; it makes the model swappable.
Kill switch
Stopping the agent in one move, after which people take over the work.
Runbook
Instructions for operation, failures and continuing the work manually.
Vendor onboarding
The process your organisation uses to assess and approve a supplier.
DPA and DPIA
Data processing agreement and data protection impact assessment.
DORA
The EU regulation on digital operational resilience in finance; it also governs contracts with ICT suppliers.
PoC
Proof of concept: a short feasibility proof on controlled data.

Questions

What your CIO, CISO and compliance will ask.

How long does it take to reach production?

Typically four to six months from the intro meeting to the first agent in production. It depends most on the integrations and on how long vendor onboarding takes on your side.

Do you work with banks?

Yes. We adapt the project to the bank’s rules: vendor onboarding, DORA requirements, the internal AI policy and, where the bank is part of a group, group approvals. For a first project we choose a process that doesn’t touch credit decisions.

What do we need to provide?

A sponsor who makes decisions, a process owner, 30–50 real cases, read access to the data, and contacts in IT, security and risk.

Can everything run inside our infrastructure?

Yes. Most often we use your existing cloud tenant in an EU region, and for the strictest requirements a local model on your servers. An external AI provider is used only with approval from your legal and security teams.

Who is responsible for a decision the agent proposed?

Your employee makes and approves the decision. We implement and test permission controls to specification and provide the audit trail as agreed. Responsibilities and incident handling are defined in the contract.

How do we know the agent is reliable enough?

We agree and sign off the quality threshold before the pilot. We measure it on your real cases, per category of work and per language. If the threshold isn’t met, the agent stays at proposals.

What if we want to change the model or end the engagement?

The model is swappable through the gateway. Prompts, specification, eval set and the code written for you are yours, and the exit is described in the contract upfront.

First step

Let’s pick the process where an agent proves its value.

A conversation with you and the process owner. If an agent isn’t the right answer, we’ll tell you before we quote.

or call +385 1 3369 800