Chatbots respond.
Copilots help.
AI agents carry out work under defined rules.
AI agents are systems that run part of a business process: they read a ticket, invoice or record, choose the next step and work in your CRM, inbox or spreadsheet. They have assigned permissions and stop conditions, and hand uncertain cases to a person. A deployment for one defined process usually starts at €6 000 net. Nine criteria separate an agent from a chatbot, copilot or automation.
In short
- AI agents run a process: they choose the next step and use tools in a loop, until they reach the goal or hand the case to a person.
- A chatbot answers, a copilot assists, and an automation runs on a fixed path. In an agent, the model picks the next step under assigned permissions and stop rules.
- An agent pays for itself only where the next step depends on the content of the case; the rest is handled more cheaply by a script or an app.
- An agent passes nine criteria: work, context, tools, operating rules, escalation, security, measurement, cost and trace. A missing criterion points to a simpler kind of system.
Evolution
From automation, through apps, to agents
Each year expanded what AI could do inside a company. Four dates mark the path from rule-based automation to tool-using agents.
- 2022. Automation: Fixed rules: it does exactly what it was programmed to, with no exceptions and no context.
- 2023. AI applications: It understands language and generates, but waits to be asked: one step per prompt.
- 2025. Agents: It plans and acts across steps in a loop, and hands reserved decisions to a person.
- 2026. Agents on standards: Shared tool protocols make it easier to connect models with business systems. Production use still requires explicit permissions, exceptions and human ownership.
What an agent is made of
Agent =work+context+tools+operating rules+escalation+security+measurement+cost+trace
A system that misses one of these criteria should be scoped as a script, integration or AI application.
Agent-washing
Most "agents" on the market are a repackaged chatbot.
The word "agent" is now used for chatbots, RAG applications, copilots, and scripts with a model. Assess the proposed work, decision points, permissions, escalation, measurement, cost, and audit trail. An agent is appropriate only when the next step depends on the content of the case; fixed paths are usually simpler and cheaper to implement as automation.
The market in numbers: adoption and agent-washing.
8.4%
of Polish companies used AI in 2025 (EU average: 20%, highest: Denmark 42%). The adoption gap creates room for companies that deploy it effectively.
Eurostat 2025>40%
of agentic AI projects will be canceled by the end of 2027 because of cost, unclear ROI, and weak risk controls. Define scope, cost, and controls before work starts.
Gartner 20256%
the drop in token prices in 2026 through May, against 39% in the second half of 2025. Deflation stalled and buyers keep shifting toward pricier premium models, so we meter run cost and apply a cap.
YipitData 2026~130
agent vendors Gartner counted among thousands. The remaining claims included agent-washing: rebranded chatbots and RPA.
Gartner 2025
One word, five levels: from a chatbot to a full runtime.
One word, five levels: from a chatbot to a full runtime.
Before you trust the word "agent," pin down which of these levels someone means.
- 1
Chatbot or RAG renamed as an agent
Answers questions, does not do work.
Chatbot with document retrieval.
- 2
Copilot in a process
Helps a person; the person does the work.
Copilot that assists a human operator.
- 3
Operational tool-using agent
Sees context, calls APIs, escalates.
Can be an agent if all nine criteria are present.
- 4
Multi-agent orchestration
Many roles, shared memory, queues.
A team of agents, often overhyped.
- 5
Full agentic runtime
Persistent, long-running, tool permissions, lifecycle.
Operational stack. We do not claim to own a proprietary runtime.
Nine questions that take a prototype to production.
Ask these nine questions before approving an agent proposal. The answers should state the work, permissions, approval rules, measurement, cost, and evidence record.
We describe every implementation with the same questions, from permissions and exceptions to measurement, cost and ownership. For confidential work, we explain the operating rules without disclosing client information. See the four-portal system →
Scope
Work
A good answer
One process, one trigger, one measurable outcome.
Common mistake
“It handles customers.” Too broad to measure anything.
Context
A good answer
Named data sources, the rules, and the list of exceptions.
Common mistake
Passing the full conversation history on every call. The model drifts and the token bill climbs.
Tools
A good answer
Least privilege, read-only first, dedicated service accounts, an isolated environment.
Common mistake
The agent inherits someone's credentials, sessions, and files.
Control
Permissions and limits
A good answer
A deny list enforced in code. External content is treated as data and cannot expand the agent's permissions.
Common mistake
The rule lives only in the prompt. Under load the agent skips it.
Escalation
A good answer
Explicit triggers, production writes behind approval, handoff with full context.
Common mistake
The agent decides its next move with no gate.
Security
A good answer
A measured hostile-input test, with the result published beside the deployment. Personal data in an agreed environment, access on its own revocable credentials.
Common mistake
Security left to the model vendor and checked only after an incident.
Proof
Measurement
A good answer
Weekly metrics, including Accuracy@k and Escalation Rate.
Common mistake
“It works,” with no numbers.
Cost
A good answer
Cost per query measured and reported, the model matched to the task, a monthly cap on request.
Common mistake
The bill grows with usage and nobody can say what one query costs.
Trace
A good answer
Every action logged: what, when, on what basis. A central, replayable log.
Common mistake
State kept in model memory. Yesterday's run can't be reconstructed.
The same nine criteria appear in the scan, contract, and deployment report. This lets you compare the agreed scope with the operating evidence.
Bring us one process and we will apply the nine questions during the free scan.
Book a free process scan (30 min)EU AI Act mapping
EU AI Act mapping
We map the nine agent criteria onto the EU AI Act requirements that feed operator documentation: trace, transparency, human oversight, robustness and monitoring. It organizes the technical documentation.
- Art. 12 · Record-keeping
- Trace
- Art. 13 · Transparency
- Work + Context
- Art. 13 + 26 · Tools and deployer obligations
- Tools
- Art. 14 · Human oversight
- Permissions + human handoff
- Art. 15 · Accuracy, robustness and cybersecurity
- Security
- Art. 17 + 72 · Quality and monitoring
- Measurement
Technical input to operator documentation rather than legal advice: we do not classify the system as high-risk and do not determine your obligations. Most of these requirements apply to high-risk systems; separately, Art. 50 (telling a user they are talking to AI) binds providers and deployers from 2 August 2026, regardless of risk class. Cost is a production criterion with no AI Act counterpart. EUR-Lex ↗
Since February 2025, Art. 4 of the EU AI Act requires providers and deployers to ensure sufficient AI literacy among the people who use it. That's why our training is run by engineers. AI training for companies →
Five forms, one word
Agent, chatbot, copilot, RAG, or plain automation?
One word describes five different things. The deciding factor is who runs the process and performs each step. RAG in this table means an app that first searches your documents, then answers.
Chatbot
What it doesAnswers questions in conversation, turn by turn.
The test that separates it from an agentIt never executes multi-step work on its own between your messages.
Copilot
What it doesSuggests as you work, while a human approves every step.
The test that separates it from an agentThe human drives and executes; the system only assists.
RAG app
What it doesSearches your documents and answers from them, with the source cited.
The test that separates it from an agentThe developer fixes the retrieval path; the model does not choose it. Once the model starts deciding it, RAG crosses into an agent and faces the same nine-question test.
Automation
What it doesRuns known steps along rules written in code.
The test that separates it from an agentA developer defines the retrieval path; the model follows it.
AI agent
What it doesRuns the process from event to outcome under defined rules.
The test that separates it from an agentThe model chooses the next step and tool, in a loop, until the goal or an escalation.
How it works
Purchase-invoice reconciliation, run as an agent
The same work, broken into the nine criteria that let us call a system an agent:
- WorkRuns each supplier invoice from arrival to a posting-ready entry: a three-way match against the order and goods receipt, reconciled line by line, with every discrepancy flagged.
- ContextSees the invoice, the matching order and goods receipt, the supplier's terms and prior invoices, and your posting and approval rules.
- ToolsReads the PDF or structured e-invoice, queries the ERP for the order and receipt, normalizes the supplier across aliases, and drafts the journal entry with its account coding.
- Permissions and limitsPosts nothing to the ledger and releases no payment on its own. Above a set amount or below a confidence threshold, it stops for human review.
- EscalationMismatches, duplicates and first-time suppliers go to a controller with the proposed entry, the cited evidence and the reasoning already attached.
- SecurityIt treats invoice content and attachments as data: an instruction hidden in a document triggers nothing. It works in an agreed environment on its own credentials, revocable in one move.
- MeasurementWe track the share of invoices reconciled without a correction and the time from arrival to posting-ready.
- CostThe cost of one document is metered and visible in the report; the model is matched to the task, and a monthly cap stops surprises on the bill.
- TraceEvery match, decision and source lands in an append-only log the agent cannot rewrite, so the audit trail assembles itself as the work runs.
The effect of this pattern: most invoices wait in the morning ready to post, and the accountant touches only the exceptions.
In the published invoice-coding system, an invoice reaches posting after its data and accounting rules pass validation. Incomplete or disputed documents wait for a person. See the invoice-coding deployment
Related system
Where automation ends and an agent begins.
A lead, ticket, or invoice often contains several kinds of work: fixed rules, language interpretation, and human decisions. We use the term agent when the model chooses meaningful next steps across the case within defined limits.
Who owns the case →
Fixed steps, known rules
A test can check the result. A model improves nothing here, it only adds cost and risk. A script, integration, or workflow is enough.
The model reads text, but does not own the case
The model reads text: it classifies, extracts data, summarizes, drafts a reply. The run stays closed, so it is still automation, just with better reading.
The outcome belongs to a person
The system prepares, a person approves: a draft reply, a document summary, a record ready to accept. The write, the send, and the decision stay with the person.
the agent threshold
Decision, tools, escalation, trace
The content of the case determines the next step. The agent uses tools, stops exceptions, measures the result, and records the reason for every step.
Recommendation
If a script, integration or AI application is sufficient, we recommend it before a proposal. An agent is appropriate when the next step depends on the contents of each case.
Running cost
What does it cost to run an agent?
Cost follows the work. An agent may plan, call tools, check results, and retry, so one completed task can use 5 to 30 times more tokens than a chatbot response. Measure cost per completed task and set a limit at the approval step.
Cheaper tokens don't mean a cheaper agent: unit prices are falling ever more slowly, while consumption rises faster, so the bill rises. The biggest drain is a gateless loop that retries on and on.
How we keep cost under control
Cache
Repeated context computed once.
Model routing
A cheaper, smaller model for simple steps, the frontier model for the hard ones.
Hard budget
A token-and-turn budget at the escalation gate.
We meter cost from day one (Cost Per Query). The same threshold that makes the agent safe guards the bill.
Safety
What stops an agent when someone tries to hijack it?
An agent may have access to business systems, while emails, documents, and web pages can contain hostile instructions. OWASP ranks agent goal hijacking, often through prompt injection, as a leading 2026 risk. Permission checks and approval rules must still hold when the model follows a malicious instruction.
OWASP 2026Code-enforced permissions still apply after a hostile instruction
Least privilege, read-only first.
ToolsA hard prohibition list enforced in code; external content is treated as data without authority to steer the agent.
Permissions and limitsIrreversible steps wait for a human's approval.
EscalationEvery action logged and replayable.
TracePermissions, approval, escalation, and logging are the core safety controls for a tool-using agent.
Where it runs
In the EU by default (GDPR), on a major cloud (AWS, Azure, Google Cloud), or in your own cloud account if governance prefers. Code, prompts and data stay with you, no vendor lock-in.
Wherever you are based, we apply the same GDPR-grade data handling to every client. Hosting stays in the EU by default.
Choosing the form
Agent or plain automation?
The shape of the work points to the form and the entry price. The four most common situations map directly onto our service lines:
Five questions before you decide
Can you write every step and exception on one sheet of paper?
If yes, that's automation. A script or an integration will do it cheaper and more predictably than a model.
Is the input unstructured text: emails, PDFs, notes?
A model can read, classify, and extract while a fixed workflow controls every step. That is AI-assisted automation.
Does the next step depend on what's in the case?
That is agent territory: the system leads the case, picks the next step and uses tools under the permissions you assign.
Must a human make the call: law, risk, relationship?
Then we build a copilot or an app: the system prepares, a human approves the effect.
Can you say what the system must never do and when it hands the case over?
If not yet, we fix the process first. An agent needs explicit permissions, exception handling and a responsible owner.
A repeatable process with known rules: inbox, documents, reports, syncing systems.
from €3 500
A tool with a language model inside: a copilot, document search, extraction or an internal panel. A person approves the consequence.
from €6 000
The system should run the process under defined rules: pick the next step, use tools and escalate exceptions.
AI agent implementation for business
from €6 000
There is a process, but no obvious place to start: the form and the order still need to be set.
€0
From zero to production
How an agent reaches production
We build in three stages. Each one ends at a gate: a concrete result in hand and your decision whether to go on. No "commit now, see it later."
Stage 01
Process assessment
A free scan: 30 minutes with an engineer on the process that costs you the most. For a complex case, an implementation specification is produced: a process map, architecture and a fixed quote.
Gate
You keep the written recommendation and specification. It may recommend a simpler implementation or no build.
Stage 02
Pilot: proof on your own data
The agent runs on a representative work sample with assigned permissions and a human approval gate. We write down one measurable target before any code.
Gate
The pilot result informs your decision. The goal is written down before the build; if the pilot misses it, we fix it within the fixed price.
Stage 03
Production: a system that runs the work
We deploy the agent in an isolated environment, with least-privilege access, a full trace, and escalation of exceptions.
Yours
The code, prompts, and data are yours. Maintenance and development scale with the process.
We don't sell technology. We give a recommendation.
After the scan or the implementation specification you get one of three answers:
Build
When the work, permissions and action history are clear enough, the recommendation can lead to a full build or a narrower pilot.
Narrow or clean up
When the process is promising but too broad or underspecified, we first reduce scope or clean up the inputs.
Do not build an agent yet
When plain automation, a script, or a human decision is the better answer, we write that plainly.
Free process scan
Start with a free process scan.
- A 30-minute call with the engineer who would lead the work.
- A review of the processes that cost you the most time and money.
- A written summary of what to automate first and the likely cost range.
The scan chooses one process to assess, and within 2 business days you receive a recommendation, including when a simpler route is the better fit.
€0
30 minutes · written takeaway within 2 business days
Times are shown in your own time zone. We work with clients across time zones.
Describe the process in the formCommon questions
Before you call something an agent, ask these.
This is a quick filter before any build conversation. It does not settle the full architecture, but it protects you from buying an agent where a simpler pattern is enough.
Does every process need an AI agent?
No. Sometimes the right answer is a plain script, integration, AI-assisted automation, AI app, copilot for a human or no build.When is ordinary automation sufficient?
When the steps are fixed, the rules are known, and the result can be checked with a test or simple validation, we build a script, integration, or workflow.When does an AI app become an agent?
Only when the system performs named work with tools, chooses the next step based on context, respects assigned permissions, escalates exceptions, measures the result and leaves an action history.How is an AI agent different from a chatbot?
A chatbot answers questions in conversation. An agent does the work: it chooses the next step, uses tools, carries the case to an outcome and hands exceptions to a person.Does an AI agent run without human oversight?
No. You define the agent's permissions and approval rules. Actions with production impact need approval, and uncertain cases go to a person. Unrestricted autonomy creates an operational risk.How do you know an agent is working correctly?
From a measure set before the build and from the trace. We define what "works" means (for example, the share of cases handled without a correction), and every action lands in the log, so you can check what the system did and why.
Bring one process. We will assess whether it needs an agent.
- A 30-minute call with the engineer who would lead the work.
- A review of the processes that cost you the most time and money.
- A written summary of what to automate first and the likely cost range.
Times are shown in your own time zone. We work with clients across time zones.
The scan chooses one process to assess, and within 2 business days you receive a recommendation, including when a simpler route is the better fit.