Everything is called an agent.
From a FAQ chatbot to a system that executes financial operations. Without clear definitions there is no basis for assessing risk or ROI.
Service · Autonomous systems
Auditable autonomous systems. Not chatbots.
We do not sell agents. We sell engineering that, when the problem calls for it, composes autonomous systems on top of LLMs with a bounded role, controlled access and an auditable log. If your case does not need an agent, we tell you what it does need. We do not automate for the sake of automating.
Confusion in the market.
Dangerous opacity.
Vendors sell "agents" whose behaviour cannot be explained or audited. In production, that is an incident waiting to happen.
Inflated expectations.
Salespeople promise agents that do "everything". What they deliver are chained prompts that break at the first real edge case.
Designing a role, not a chat.
Before writing any code we define what the agent does, which tools it can invoke, what data it sees and which decisions escalate to a human. All in writing, signed before we start.
-
01
Role design
Workshop with your team: tasks, limits, human fallback, success criteria. Output: a signed agent spec.
-
02
Bounded build
Agent implemented on the Claude API or another model depending on the case. Tools limited to what was agreed. Adversarial tests.
-
03
Supervised piloting
2–4 weeks with a human reviewing 100% of the outputs. We tune prompts, add guardrails, measure accuracy.
-
04
Gradual operation
A progressive move to autonomy. Real-time dashboard of cost, accuracy and human interventions. Kill switch available.
Agent, playbook and control.
One or several agents
Each with a specific role (classify tickets, process invoices, answer level-1 queries). Not one single agent that does everything.
Operational playbook
A manual of what the agent does, how to ask it for things, when a human should step in, how to escalate problems.
Control dashboard
Live visibility of runs, cost per operation, error rate, human interventions. Full transparency.
Auditable log
Every agent action is recorded: input, decision, output, cost. Exportable for internal or external audit.
Kill switch
An operational button to shut the agent down live. A mandatory feature: if something breaks, you stop the flow in seconds.
Updates
When better models come out or the regulations change, we update the agent without rewriting from scratch.
- 100% Auditable actions
- <1s Kill switch available
- 2–4 wks Supervised pilot before autonomy
Our own products, not slides.
To decide with judgement.
-
What an AI agent is and when your company needs one
An AI agent is not a chatbot. It is an autonomous system with a bounded role, controlled access to tools and an auditable log. This is the honest definition, no hype.
-
RPA vs AI agents: the differences and when to use each
RPA follows the rules you gave it; an AI agent decides how to meet an objective within a bounded perimeter. They are not rivals: they are different layers. We explain when each is enough and how much they cost to maintain.
-
What auditable AI (AI without a black box) is and why your company needs it
Everyone says "we use AI". Almost nobody explains how it decides. That difference, between a black box that spits out answers and a system you can audit, is what separates a reliable tool from a legal risk with good marketing. This is what auditable AI really means.
What people ask us.
-
01 What is the difference between an AI agent and a chatbot?
A chatbot replies with text. An agent executes actions: it calls APIs, modifies databases, sends emails, invokes tools. It has objectives, authorised tools and the ability to decide which steps to take to complete a task.
-
02 Can I control what the agent does?
Yes, completely. We define contractually which tools it has available, what data it sees, which decisions it is authorised to take and what it must escalate to a human. The agent cannot step outside those technical limits.
-
03 What happens if the agent makes a bad decision?
Every decision is recorded with its input and output. If we detect an error we can shut the agent down (kill switch), analyse the log and correct it. In supervised autonomy phases, a human reviews every action before it executes.
-
04 What is the running cost of an agent in production?
Three components: model cost per operation (it varies by task and provider), infrastructure (it depends on volume and the latency you require) and evolutionary maintenance. The breakdown with the real rates and volume for your case comes out of the diagnosis, before committing to anything.
-
05 Does it comply with the AI Act and the GDPR?
Yes, by design: impact assessment, data minimisation, decision logging, the right to human review in high-risk categories. Documentation ready for the data protection authority if needed.
-
06 Do you use Claude Code or your own agents?
Both, depending on the case. Claude Code (Anthropic) for engineering work; our own agents built on the Claude API or other models for business cases. We choose by technical fit, not commercial preference.
Let's start
Let's talk about your operation.
Tell us what costs more than it should. We come back with a concrete plan, realistic timelines and a clear yes or no.
contacto@plantekia.com