Closed scope
The agent performs a declared set of tasks. Outside it, it routes instead of improvising.
Solution
AI Agents
AI that acts, under supervision.
An agent is not a chatbot: it does not only answer, it acts — queries a system, fills in a record, routes a case. We build agents with a closed scope, explicit tools and supervision where the decision matters.
01 The problem
Between the chatbot that only talks and the employee who does everything
What usually happens
- The current chatbot answers questions but resolves nothing — and the customer ends up with a human agent anyway.
- Triage, classification and routing eat hours of qualified people's time.
- The information that decides the case is scattered across systems the agent must open one by one.
- Experiments with generative AI impressed in the demo and did not survive the real case.
- Nobody can say how much each answer the AI gives costs per month.
Signs this is your case
- There is a written procedure a person follows step by step, several times a day.
- The decision depends on checking two or three systems and applying a known rule.
- The volume of simple cases hides the ones that genuinely need people.
- There is already enough historical record to say what a good answer looks like.
- The company accepts human supervision on the higher-impact decisions.
02 What we do
The practices in this solution.
Tools
Every possible action is an explicit tool, with permission and limits — not free text.
Context
Controlled access to the internal sources the task requires, and to those only.
Supervision
Sensitive decisions go through human approval before becoming action.
Audit log
Every action is recorded: input, tool used, result and who approved it.
Evaluation
A set of test cases measuring quality before and after every change.
Cost control
Usage limits per run and per period, with alerts — AI cost is variable by nature.
Handover
A clear escalation path for when the agent lacks enough confidence.
03 Scope
What you get — and what you do not.
The second column is the one almost nobody publishes. It exists because a scope with a declared limit is the only honest way to agree on price and deadline.
Included
- Scope definition listing the tasks the agent performs and the ones it refuses.
- Agent deployed and integrated with the required sources and systems.
- Tools with permissions bounded per task.
- Human supervision flow for the decisions defined as sensitive.
- Auditable record of every run.
- Evaluation set with real cases and expected results.
- Cost and volume dashboard per period.
- Operating documentation and training for whoever supervises.
Not included
- Model API costs, which are metered by usage and billed by the provider.
- Any guarantee of absolute accuracy: language models make mistakes, and the design assumes it.
- Organizing or cleaning the knowledge base the agent will consult.
- Replacing the team — the design presumes human supervision.
- An open-scope agent that decides on its own to act outside what was agreed.
- Ongoing operation and evolution without a specific contract.
04 When not to hire it
When not to use an agent
If the task fits a deterministic rule, ordinary automation solves it better, cheaper and without variation in the answer. If there is no reliable knowledge base, the agent will answer with conviction about wrong information. And if the operation accepts no margin of error and no supervision, the problem is not a candidate for AI — it is a candidate for a system.
05 Timeline
How long it takes — and what makes it vary.
A proof of concept with real cases usually stands up in a few weeks. What decides the total schedule is access to the internal sources and the time to build the evaluation set — without it, there is no way to state the agent improved.
How many tasks are in scope and how clearly they are defined.
Access to internal sources and the state of that information.
How many systems the agent must call.
Level of supervision the operation requires.
Expected volume of runs per month.
Record-keeping and compliance requirements of the sector.
06 Services involved
Where this solution comes from.
A solution is not a separate box: it is the same services applied to one specific problem.
- Artificial Intelligence
- AI agents, chatbots and automation that speed up decisions and results. See the service →
- Enterprise Systems
- ERP, CRM, integrations and system modernization for your operation. See the service →
07 How a solution starts
Before the proposal, the assessment.
A solution is not bought from a catalogue. It starts with an assessment of the real process, and only then becomes scope, schedule and proposal.
- Assessment
- A free initial conversation with the people who run the process.
- Mapping
- The process as it happens today, with the points of loss marked.
- Recommendation
- What to automate, what to integrate and what to leave alone.
- Proposal
- Scope, stages and schedule — after the assessment, never before.
08 Frequently asked questions
The questions people
actually ask.
01What is the difference between an agent and a chatbot?+
A chatbot answers; an agent acts. The agent has tools — look up an order, open a ticket, update a record — and uses those tools within a declared limit. It is the difference between a conversation and work carried out.
02What if the agent gets it wrong?+
It will get some proportion wrong, and the design assumes that from the start: sensitive decisions go through human approval, every action is recorded, and there is an escalation path when confidence is low. Anyone promising absolute accuracy is not describing the technology.
03How much does it cost to run?+
The cost is variable, metered by model usage, which is why limits per run and per period come with alerts and a dashboard. We estimate the range during the assessment, based on the real expected volume.
04Will our data train the provider's model?+
It depends on the provider and the plan contracted. That is decided before the first line is written, with services and contracts compatible with the confidentiality of your operation.
05Can we start small?+
That is what we recommend. One task, one evaluation set, a period of supervised operation. Widening the scope later is cheap; fixing an agent that did too much too early is not.
09 Other solutions
What else we solve.
01A
Process Automation
Repetitive work leaves the hands of the people who should be deciding.
SolutionSee →
02ISystems Integration
The systems the company already has finally start talking.
SolutionSee →
03MLegacy Modernization
Changing the engine with the car running — which is how it actually happens.
SolutionSee →
AI Agents
Shall we talk about
your case?
The first conversation and the preliminary assessment cost nothing. Only then come the scope and the proposal.