Document agent
Query an approved knowledge base, synthesise and cite useful passages.
DOSSIER 401
Private AI agents
Specialised assistants operating within a defined perimeter, using known sources and explicit human control.
Operational briefPRIVATEclosed perimeter
HITLhuman control
LOGGEDtraceable actions
Operational brief
Automate without surrendering command.
A useful agent is not a spectacular demo. It is a restrained system that knows what it may read, what it may produce, when it must stop and who validates its action.
AR / DOSSIER 401INDEPENDENT ANALYSIS
MEASUREMENT WORKSHOP
An agent with 80% success may pass only 64% of its programme
The denominator changes the interpretation. Of 50 planned trials, 40 finish and 32 pass: success among completed trials is 80%, but success across the whole programme is 64%. Keep the ten unfinished trials visible.
AI-generated illustration
Fictional example · no measurement of this site or a client
| Indicator | Calculation or threshold | What the result establishes |
|---|---|---|
| Programme completion | 40 / 50 × 100 = 80 % | 40 / 50 = 80%: ten trials still lack a completed outcome. |
| Success across all planned trials | 32 / 50 × 100 = 64 % | 32 / 50 = 64%: an unfinished trial is not a success. |
| Success among completed trials | 32 / 40 × 100 = 80 % | 32 / 40 = 80%: this conditional rate excludes ten unfinished trials. |
From observation to decision
Define success before testing
Write the expected outcome, allowed access, time limit and critical errors.
Keep every trial
Separate success, completed failure and non-completion. Critical failures are a subset of completed failures.
Review critical cases
Examine each incident and recovery path. A favourable average cannot authorize deployment.
CONTROLLED SYSTEM
Automation remains a responsibility.
Sources, permissions, evaluation, validation, logs and stopping conditions are designed with the system. Performance matters only while data, decisions and reversibility remain governable.
Illustrative scene.
02 / Intervention areas
Intervention areas
Monitoring agent
Collect, deduplicate, qualify and report only meaningful changes.
Editorial agent
Prepare structures, variants, metadata and checks without publishing alone.
Audit agent
Check pages, data, rules and anomalies against an explicit framework.
Business orchestrator
Coordinate tools or agents with escalation rules and centralised validation.
Decision copilot
Prepare options, sources, contradictions and consequences without replacing the decision-maker.
03 / FIELD NOTES
High-value automation
An agent is relevant when a process is frequent, documentable and important enough to deserve controls.
01DOCUMENTATION
Query a corpus
Retrieve a rule, compare versions and produce a synthesis with citations.
02MONITORING
Reduce noise
Absorb source volume and report only changes matching precise criteria.
03PRODUCTION
Industrialise with control
Prepare variants, checks, metadata and files without publishing alone.
04COMPLIANCE
Verify systematically
Apply a stable framework to pages, files or data and explain each anomaly.
04 / OUTPUTS
System components
The agent is only part of the system. Rules, tests, logs and interfaces matter as much as the model.
Mission specification
Objective, users, limits, measurable success and prohibited actions.
Source register
Authorised data, freshness, ownership, confidentiality and quality.
Tool matrix
Available capabilities, permissions, costs and invocation conditions.
Validation policy
Free, supervised and prohibited actions.
Evaluation set
Normal cases, traps, expected failures and success criteria.
Operational log
Inputs, decisions, tools, outputs and human interventions.
WORKING RECORDS
Make the dossier verifiable
Three practical records connecting diagnosis to a traceable decision.
Authority record
Minimum content : Define permitted tasks, accessible data, tools, spending limits and actions needing human approval.
Check : Can the agent do more than its task requires?
Evaluation set
Minimum content : Include normal, refusal, hostile instruction, tool failure and partial-action cases with expected outcomes.
Check : Does the system stop and explain failure without hiding the real state?
Recovery log
Minimum content : Record state before and after action, human decision, repair and release condition.
Check : Can a person resume without guessing what ran?
05 / Method
Security by design
Least privilege, bounded sources, logging, validation and stop strategies are designed before automation.
- 01Measurable business objective
- 02Authorised sources and data
- 03Strictly necessary tools
- 04Human validation
- 05Decision log
- 06Tests and known limits
- 07Design a safe stop
- 08Monitor cost and drift
DECISION CHECKS
Decisions to document
Two checks that change the next decision.
An agent requests more permissions
Tie each permission to an authorised action and an owner. Avoid broad access when the task needs only a limited dataset or operation.
An action may exceed scope
Test deceptive instructions, tool failures and out-of-scope requests. Define where the agent stops, retains a record and asks a person to decide.
Two approaches to distinguish
Two approaches to distinguish
The right method depends on the question and the decision it must inform.
Deterministic automation
For a stable sequence of known rules, an explicit workflow is easier to test, audit and resume. It exposes exceptions and required permissions.
Action-selecting agent
Useful when variable input must be interpreted and bounded tools selected. Define an evaluation set, action log, limits and human takeover.
Proposed work plan
Adapt this outline to the agreed scope. It describes a proposed method, not a completed mission, measured result or guaranteed schedule. JSON tracking fields are empty.
AI-generated illustration
Prepare before starting
A bounded task, authorized corpus, required permissions and acceptance cases. Identify actions needing human approval.
Scope the work
Define sources, permitted operations and expected outputs. Test access refusal before adding documents or tools.
Apply the method
Build a first task with reading and action traces. Separate generated answer, executed operation and final-state verification.
Verify and hand over
Repeat permitted, forbidden and incomplete cases. Check stopping, resuming and transferring traces before expanding use.
Pause or revision conditions
Block an operation without established permission or effect verification. A plausible answer is insufficient to authorize an action.
06 / FAQ
Frequently asked questions
Clear answers on scope, limits, methods and engagement conditions.
Can an agent publish alone?
Technically yes, but it is not always desirable. Sensitive external actions remain validated.
Must a model be trained?
Rarely at first. Good context and tool architecture is often sufficient.
Can it run internally?
Local, private or hybrid architecture can be considered according to constraints.
How is this different from a chatbot?
An agent pursues a bounded mission, uses tools, keeps useful state and follows an action policy.
How are hallucinations limited?
Bounded sources, citations, validation, tests, explicit refusal and separation between generation and action.
Which data should not be provided?
Any data whose use, hosting, retention or recipient is not controlled.
How is return on investment calculated?
Actual time saved, errors prevented, processing speed, achieved quality and total operating cost.
Methodological references
Official documents explaining methods, review criteria and their limits. These references do not certify the service or guarantee search rankings.
08 / CONTACT
Open a private channel
State the context and objective. You will receive a clear initial reading.
Turn a need into a defined assignment
A few interventions relevant to this topic. Each amount covers the stated reference scope; the quote confirms priorities and overlapping tasks.
Private document agent prototype
Bounded sources, sourced answers and functional tests.
Corpus to scope : 1 · 24 h
From 1,920 € excl. VAT / 2,304 € incl. VAT
Preparation and delivery checks
- What to prepare
- Prepare an anonymised sample, expected output and data-processing rules.
- Suggested process
- Test a small batch, record failures and check recovery and output controls.
- Delivery to agree
- Agree the prototype, test data, review rules and operating limits.
- Review for this service
- Repeat test cases and check situations requiring human approval.
- Scope boundaries
- A prototype does not establish production readiness; quotas, costs and privacy require separate scoping.
Indicative working scope, to confirm in the quotation.
Workflow automation
Trigger, processing, checks and error log.
Workflow : 1 · 12 h
From 960 € excl. VAT / 1,152 € incl. VAT
Preparation and delivery checks
- What to prepare
- Prepare an anonymised sample, expected output and data-processing rules.
- Suggested process
- Test a small batch, record failures and check recovery and output controls.
- Delivery to agree
- Agree the prototype, test data, review rules and operating limits.
- Review for this service
- Check a successful batch, a failure and recovery without unwanted duplication.
- Scope boundaries
- A prototype does not establish production readiness; quotas, costs and privacy require separate scoping.
Indicative working scope, to confirm in the quotation.
AI-assisted editorial workflow
Drafts, metadata, human review and traceability.
Workflow : 1 · 16 h
From 1,280 € excl. VAT / 1,536 € incl. VAT
Preparation and delivery checks
- What to prepare
- Prepare an anonymised sample, expected output and data-processing rules.
- Suggested process
- Test a small batch, record failures and check recovery and output controls.
- Delivery to agree
- Agree the prototype, test data, review rules and operating limits.
- Review for this service
- Check review, sources and the approval step before any planned publication.
- Scope boundaries
- A prototype does not establish production readiness; quotas, costs and privacy require separate scoping.
Indicative working scope, to confirm in the quotation.
Prepare and review a service
Use these notes to scope the request and review delivery. The quotation defines the agreed work, volumes and schedule.
AI, automation & data
AI-generated illustration
Build a useful workflow with defined data, errors and human review.
- What to prepare
- Prepare an anonymised sample, expected output and data-processing rules.
- Suggested process
- Test a small batch, record failures and check recovery and output controls.
- Delivery to agree
- Agree the prototype, test data, review rules and operating limits.
- Scope boundaries
- A prototype does not establish production readiness; quotas, costs and privacy require separate scoping.
DEEPER READING / PRACTICE
Define autonomy before building an agent
An AI use case should be defined by its task, scope and the means of checking its outputs. Response quality and control over actions need to be assessed together.
Editorial supplement added on .
Starting situation
A agent must seek, write or execute a professional action. The definition of need specifies what it can consult, what it can change and the decisions that require human validation.
Priority checks
List entries, tools, permissions and exits. Test normal cases, missing information and requests that exceed the permitted scope.
Interpretation pitfalls
A good outcome on a demonstration does not demonstrate reliability in operation. An ambiguous instruction or a contradictory source can lead to an inappropriate action.
Working record to retain
The agent sheet describes the task, limits, validation rules and the arrest procedure. A evaluation game accompanies the developments to control the expected behavior and possible errors.
Apply the approach to an initial case
Start with a limited task whose entries and expected results can be examined. Provide incomplete information, a contradiction and an out-of-scope request. These cases show how the system deals with the unknown and when a person must resume the analysis or authorize an action.
Present the result and its limitations
Documenting the errors with the entrance and exit concerned. Before a new use, check if the correction concerns the order, data, permissions or human control. An evolution must be accompanied by a new observation of the affected cases so that it does not confuse confirmed adjustment and improvement.
EXAMPLE / READING THE NUMBERS
Separate success from permission compliance
Fictional data illustrating the method; not client results or measurements of this website.
| Observation | Count | Interpretation |
|---|---|---|
| Tasks tested | 100 | Test scope. |
| Successful tasks | 94 | 94 / 100 = 94% success. |
| Unauthorised actions | 1 | 1 critical incident assessed separately. |
Show the interpretation limit
What these numbers do not prove
An average does not make a forbidden action acceptable. Set stopping criteria before testing, based on consequences. The last two rows are not categories to add together.
PRACTICAL CASE / WORKED EXAMPLE
An agent receives a request outside his scope.
This pedagogical case examines AI use through its entries, permissions and controls. It does not describe a guaranteed performance; it helps to build an assessment adapted to the possible consequences of the task.
Example context
A agent designed to write a summary receives a request to amend a shared document. Its tools technically enable action, but this autonomy has not been defined.
Suggested approach
Examine tasks, permissions and validation rules. Testing the demand and the ambiguous variants. Reduce unnecessary access and predict human recovery when action exceeds the mission.
Conclusion and next step
The expected behavior is documented and evaluated. The success of a writing remains different from the permission to act on an external resource.
Checklist for applying this example
- Relate permissions to the necessary actions to the authorized task.
- Test missing information, contradictions and requests out of scope.
- Define human times and results deemed acceptable.
- Keep the errors, versions and controls after modification.
Questions for deeper practice
How do you decide what an agent can change?
Relate each permission to an authorized task, to an acceptable consequence and to a control. Unnecessary capacities increase the scope to be controlled without automatically improving the service provided.
How to decide if an AI system is ready for wider use?
Examine representative cases, major errors and conditions of recovery. The result must show what has been tested, with what criteria and what limits. An extension of task, data or permissions requires a review of risk and controls, even when the initial behavior was satisfying.
Why do I use the evaluation cases and control rules?
Inputs, tools and instructions can change. Keeping versions can explain why a behavior has changed and compare compatible executions. The cases added as a result of an error also document the situations now monitored, without making the boundaries of the previous evaluation game disappear.
DETAILED WORKSHOP / APPLICATION
Align permissions with the actions that are really necessary
This workshop offers a five-step sequence to apply the page labels. The entries, observations and conclusions are kept together in order to facilitate the verification and recovery of the work.
01 / Prepare the necessary information
Assemble tasks, tools, data and expected outputs. List the consultation and modification actions as well as situations requiring human validation.
02 / Carry out the approach
Test a normal demand, a lacking information and an outperimeter action. Examine the tools used and the possible consequences. Reduce permissions without the necessary role in the task.
03 / Examine an alternative situation
If the agent can technically perform an unauthorized action, correct the access and control rule. A single order should not hide an inexpensive capacity.
04 / Decide on the next step
The system is controlled when expected behaviors and repetitions are assessed. The success of an example does not automatically represent all possible uses.
05 / Hand over the verification records
Transmit task sheet, permissions and assessment cases. Data or tool changes are linked to the controls they make necessary before a new use.
ANNOTATED WORKING DOCUMENT / REVIEW GRID
A list of tasks with permissions and repetitions
The file connects purpose, tools, data, actions and controls. It describes the limits of autonomy and situations in which a person takes the analysis again or authorizes the follow-up.
On a small screen, scroll the grid horizontally to read all three columns.
| Observation | Supported interpretation | Next step |
|---|---|---|
| The tools used correspond to the authorized task. | The behavior is consistent on the cases and permissions examined. | Keep the tests and check the important variants before expansion. |
| A data is missing and the agent requires accuracy. | The recovery can be a correct answer as needed. | Check that the question is useful and does not trigger a premature action. |
| The agent changes a resource outside the scope. | The capacity exceeds the defined framework. | Revise permissions and validation and then review the ambiguous requests. |
Reviewing the working document
The final check examines actions and outcomes together. Good writing does not automatically validate permissions; any capacity without the necessary contribution must be justified or removed from the system.
REPORTING / EXAMPLE WORDING
Presentation of results: Defining autonomy before building an agent
Point to clarify
The quality of a output and the conformity of the actions must be examined together.
Before handing over
Compare the tools used, permissions and outputs before qualifying the result.
Example of a qualified conclusion
The agent performed the task in the authorized scope on the cases examined. Ambiguous or out-of-scope requests have documented behaviors and repetitions.