The job description
It defines the scope: what this function is entitled to answer on, and above all what does not concern it. That second half is what stops the agent from straying into a domain it does not master.
Your function is not among the 85? Create the 86th.
The platform ships with 85 specialised agents. When a company has a function nobody else has in quite the same form, which is common, it does not raise a ticket: it uploads two documents. A job description, and the function’s procedures. The agent is created, and it works alongside the others from the very first question.
What sets this builder apart from others
Creating an assistant from documents has become commonplace. What has not: the agent you create inherits the full mechanism. It cites its source with page and version, carries a score from 0 to 100, passes the anti-hallucination checks, and abstains when your documents say nothing.
An in-house agent that makes things up does not save time: it produces an answer that has to be withdrawn from a signed submission.
85
agents delivered: 72 function, 12 sector, 1 librarian
2
documents are enough: a job description, procedures
0
lines of code, and no IT project
∞
additional function agents, as many as your functions require
How an agent is born
Four stages, with human approval at the end.
The fourth stage is the one other agent builders forget: someone reviews the rules the machine has set for itself, and signs off. It is the same logic as the named approval of an answer.
The procedures
Internal policies, manuals, frameworks specific to this function. This is the material, and the only material: the agent will answer from these documents, never from general knowledge.
The generated rules
The platform reads the documents in full, extracts the themes and skills, then writes the agent’s rules: scope, abstention threshold, mandatory citations.
Review, then activation
A manager from the function reviews the rules produced and activates the agent. The abstention threshold is the setting to look at first: it decides when the agent would rather stay silent.
Interoperability
Your agent does not sit alongside the system. It is part of it.
When a questionnaire arrives and a question falls within your specific function, the orchestrator calls on your agent together with the agents that carry the relevant frameworks. Each handles its share, and the answers are merged into one coherent deliverable, not a series of paragraphs placed side by side.
Your agent benefits from the entire infrastructure: the five layers, including seven anti-hallucination checks, the explained confidence score, and the sovereignty of the deployment you have chosen.
The anatomy of a defensible answer →One question, several agents
The question asked
“Describe the control of privileged access to the systems that process our data, and the escalation procedure in the event of an incident.”
Information security agent Continuity agent Your Risk Manager agent
The first two answer from your policies. Yours answers from your own in-house escalation procedure, the one no public framework describes. The final answer cites all three sources, with their page and version.
Three agents that clients have created
The function nobody else has in quite the same form.
Risk manager in banking
Job description, risk management policy, escalation procedures. The agent works alongside those that carry operational resilience and anti-money laundering.
What it brings: the way this bank escalates, which no framework describes.
M&A lawyer
Job description, collection of standard clauses, due diligence checklist. The agent tests every commitment requested against the group’s written standards.
What it brings: detection of a commitment the group has ruled out making.
Pre-sales at an IT services firm
Job description, project references, qualification grids. The agent knows the company’s actual catalogue, not the one in the brochure.
What it brings: no more losing a bid because the right reference was not found in time.
What happens next
Adoption comes through value, not through an internal memo.
Legal creates its agent because it has an urgent file. Procurement sees the result and uploads its procedures. Compliance follows, then security. No adoption campaign, no mandatory training: one team copies the team next door when it sees what there is to gain.
And every agent added makes orchestration better on questionnaires that span several domains, which is to say almost all of them.
The limit, stated plainly
An agent only knows what your documents say. If the escalation procedure exists only by word of mouth, the agent created for that function will abstain, and it will be right to. You will then be exactly in the case covered by drafting assistance: the gap is declared, an expert is named, the document is co-written, and the agent becomes useful on the day the document exists.
When a document is missing →Frequently asked questions
01
Do we need to know how to code?
No. Two documents are enough: a job description and the function’s procedures. From these, the platform derives the agent’s scope, its abstention threshold and its citation obligations. No code, no manual configuration. The skill required is not technical but functional: you need to know which procedures are authoritative in your organisation.
02
Is an agent we create as reliable as a native agent?
It inherits the same mechanism: sources cited with page and version, a score from 0 to 100, anti-hallucination checks, named approval. Its reliability, however, depends on the quality of what you upload. The native agents carry years of ingested regulatory expertise; yours carries your procedures. Both can be verified in the same way, which is what matters when an answer is challenged.
03
Are our procedures used to train a global model?
No. Uploaded documents stay in your instance and are not used to train the global models of Optivalue.ai or those of its technology partners. This is a written commitment, in Article 7.3 of our service terms, not a statement of intent.
04
Who owns the agent we have created?
You own the training data, your job descriptions, your procedures, unconditionally. The agent’s technical architecture remains the property of Optivalue.ai. At the end of the contract, there is an option to acquire the trained instance, with its terms and price written into your contract; if the option is not exercised, the agent is destroyed along with your instance. We would rather you read these terms before signing than after.
05
How many agents can we create?
As many as your functions require. The practical limit is not the number of agents but the volume of documents indexed in your knowledge space, as priced in the contract. The right reflex is to create an agent when a function has its own procedures, not when it has its own job title.
Bring the job description and the procedures. We build the agent together, in the session.
Then we ask it a real question from your last questionnaire, and you judge the answer with its source and score in front of you.