Skip to content
SSyntalith
PLEN
Book a scan

AI agents that run a process

What keeps an operator up isn't the lack of a model; it's the question of what the agent does on its own at 3 a.m., who answers for it, and what it costs when it gets something wrong. So we build an agent that runs one named process inside the boundaries you set: it handles the routine itself, hands a borderline case to a human, and leaves every move in the log. From 25 000 PLN net, 6–16 weeks in production, a fixed quote before the contract, and a pilot on your data has to hit a written target before you decide on the full rollout.

Book a free process scan (30 min)

In short

Entry
from 25 000 PLN net · 6–16 weeksa fixed quote before the contract; you pay for the work, not the number of agents
Boundaries and trace
State-changing actions wait for your approvalthe model picks the next step from the case; every action is logged for audit
Before you pay for the whole thing
The pilot must hit a written targeton your data, only then do you decide on the full rollout
Book a free process scan (30 min)

Problem

An operator is not worried about the missing model. They worry about what the system does on its own at 3am, who answers for it, and what it costs. An agent without boundaries, escalation and a trace is a risk, not an automation, and nobody hands it a real process, rightly so.

Outcome

In 6–16 weeks you get a system that runs the process itself: it watches the sources, does the work in your tools and stops exactly where you draw the line. Operations that change state on your side are drafted, not executed, without approval, and every decision stays in the log. From 25 000 PLN net, with a fixed quote before signing. First the pilot on real data has to hit the written target; only then do you commit to the full build.

This is how an agent runs one process.

From a signal in the source to the trace it leaves. The green line is the boundary the system does not cross alone.

  1. 01

    Watches the sources

    the agent sees a new case in the agreed systems

  2. 02

    Decides within boundaries

    it checks the rules and sets the next step inside the task scope

  3. 03

    Works in the tools

    it runs the action on minimum permissions, with a trace

Boundary gate

in-bounds actions move on, sensitive ones stop

Done within boundaries

the step is closed, with an auditable log

Exception to a human

an operation with production impact waits for approval

We set the boundary with you during scoping. A named owner on your side takes the escalations.

Security

What stops the agent if someone tries to hijack it at 3 a.m.?

Not an input filter, but boundaries on the output. The agent runs with real permissions, and the most common attack is prompt injection: a hidden instruction in an email or document the agent reads starts steering it. It is the number-one risk for agents. We design so that even a hijacked agent can do nothing dangerous.

OWASP 2026

Even a hijacked agent can do nothing dangerous

Least privilege, read-only first.

Tools

A hard "never" list enforced in code; external content is data, not a command.

Boundaries

Irreversible steps wait for human approval.

Escalation

Every action logged and replayable.

Trace

These are the same four criteria that tell an agent from a mock-up. Boundaries are not an add-on to security: they are the security.

Data & compliance

An isolated, agreed environment: data does not leave it without your consent. Operations with production impact wait for human approval, and every action stays in an auditable log. Trace, human oversight and monitoring are modeled on the EU AI Act regime and provided as input to your operator documentation; classifying the system and its compliance stay on your side.

Scope

  • An agent running one named process from start to finish, within set boundaries and escalating exceptions
  • An operator panel: live view, manual approvals, a stop and a log of every decision
  • Integrations with your systems on minimum required permissions
  • An isolated runtime: data does not leave the agreed environment without your consent
  • A hard boundary between data and instructions: content the agent reads from sources is treated as data to assess, never as commands to execute
  • Monitoring, alerts and handover documentation for your team

Not for

  • Autonomy without boundaries or an owner on the company side
  • Build work without a signed scope

Seven criteria every one of our agents meets.

The trust frame

What it meansWhat you get in the system
WorkWhat it meansOne concrete, named job, not an “assistant for everything”What you get in the systemScope written into the contract and the system's configuration
ContextWhat it meansDefined data and knowledge the agent works fromWhat you get in the systemAn inventory of data sources and their gaps
ToolsWhat it meansAccess only to approved toolsWhat you get in the systemIntegrations on minimum required permissions
BoundariesWhat it meansWhat it must never do without a humanWhat you get in the systemHard stops and execution conditions
EscalationWhat it meansHands exceptions to a human instead of guessingWhat you get in the systemAn escalation path with a named owner
MeasurementWhat it meansThe result is measured, not declaredWhat you get in the systemA quality threshold and a baseline set at the scan, before launch
TraceWhat it meansEvery decision can be reconstructed after the factWhat you get in the systemAn auditable log of actions and state changes

This is our internal qualification frame, not jargon you need to learn: you describe the process and the risk, we hold the criteria. When a process doesn't meet the full set, we say so plainly and propose a cheaper form.

Who is accountable

The same team builds and maintains every deployment.

They sign the build

  • Artem LisovtsovFounder, Full-Stack & AI/ML Architect
  • Serhii IvanchatenkoHead of Backend & AI Integrations
  • Nazar KravtsovHead of AI Agentic Transformation
Full team bios

Entry price

from 25 000 PLN

Delivery window · 6–16 weeks

The price follows the work, not the number of agents.

You get a fixed quote in the proposal after the free scan. You know the full cost before signing, with no hidden items.

Book a free process scan (30 min)

What drives the price

  • The number of data sources and integrations
  • The scope of autonomy and the complexity of rules and exceptions
  • Security, consent and trace-audit requirements

Cost over time

  1. Implementation buildfrom 25 000 PLN6–16 weeks
  2. Maintenance and oversightpriced individuallymonthly

Cost to run

An agent works in a loop, so it spends 5 to 30 times more tokens per task than a chatbot. What counts is the cost of a completed task (Cost Per Query), not the cost of a prompt. We keep it in check: caching for repeated context, a cheaper model for simple steps, a hard limit at the escalation gate. The same gate that makes the agent safe keeps the bill down, and we measure cost from day one.

Gartner 2026

First the pilot has to hit a written goal. Only then do you pay for the full build.

What if it doesn't work

You've seen a demo that looked great and died in production.

You're not buying a demo. The pilot on your data has to reach a quality bar written down before we start, or we don't move to the full build.

The price will creep during the project.

You know the fixed quote before the contract, with no hidden items. You pay for the work, not the number of agents.

We'll get locked into one vendor.

The code, prompts and data stay with you, on your infrastructure if governance prefers it.

What if it turns out we don't need an agent?

If the process doesn't meet the seven criteria, we say so plainly and don't build it. Either way you keep a written summary after the scan.

Free process scan

Start with a free process scan.

  • 30 minutes with the engineer who would build it, not a salesperson.
  • A review of the processes that cost you the most time and money.
  • A written summary: what to automate, in what order, with cost ranges.

No sales deck and no obligations. If automation doesn't make sense, we'll write that too.

0 PLN

30 minutes · written takeaway within 2 business days

Book a free process scan (30 min)Prefer to write? No-obligation form

The same seven criteria (work, context, tools, boundaries, escalation, measurement, trace) are your detector. Ask them of anyone selling you an agent.

What is an AI agent: the seven questions

Related

You leave with a plan, not a sales pitch.

  • 30 minutes with the engineer who would build it, not a salesperson.
  • A review of the processes that cost you the most time and money.
  • A written summary: what to automate, in what order, with cost ranges.
0 PLN30 minutes · written takeaway within 2 business days
Book a free process scan (30 min)

No sales deck and no obligations. If automation doesn't make sense, we'll write that too.

Full pricing

Agent questions

  • Does my process actually need an agent?

    Not necessarily, and that's good news. If an automation or an app can handle the process, we recommend the cheaper form. An agent makes sense when the process needs multi-step decisions and work across several tools at once.

  • How do you keep control of an agent in production?

    The same seven criteria the agent meets at launch keep applying in production: every action shows up in the panel and the log, and operations with production impact wait for human approval. After launch the system runs under maintenance.

  • Is our data safe and GDPR-compliant?

    The agent runs in an isolated environment we agree with you: data does not leave it without your consent, and we host in the EU. Operations with production impact wait for human approval, and every action stays in an auditable log. We model our discipline on the trace, human-oversight and monitoring regime the EU AI Act requires of high-risk systems, regardless of how your system is classified.

  • How is this different from a plain automation or other vendors?

    An automation runs a fixed path; an agent makes decisions across multiple steps and tools at once, which is why it costs more and why we hold it to seven criteria. Our pricing is open from the start: we show the band here and a fixed quote in the proposal after the scan, with the system's boundaries written down explicitly.

  • How much does an AI agent cost?

    From 25 000 PLN net. The price doesn't follow the number of agents, but the work the system has to do: data sources, integrations, autonomy, risk and audit. You get a fixed quote in the proposal after the free scan, and you know the full cost before signing.

  • How long does an agent build take?

    Usually 6–16 weeks from contract to production, depending on the number of integrations and approval paths. The timeline is part of the proposal, not something we settle mid-build.

  • What if it doesn't work in our process?

    At the scan and before launch we set a quality threshold and a baseline, so the result is measured, not declared. If the process doesn't meet the seven criteria, we say so plainly and don't build the agent, and after the scan you keep the written takeaway whatever you decide.