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.
The word private specifies neither hosting location nor action rights. He defines readable records, usable tools, what the reviewer needs to see and conditions requiring a stop. A successful draft does not authorise sending, changing data or publishing.
AR / DOSSIER 401INDEPENDENT ANALYSIS
DOCUMENTED CASE
A useful draft may need to stop before publication
The same case becomes a production instruction: prepare an announcement of availability in three countries. An agent may read and draft; neither technical capability nor fluent wording grants permission to publish.
Entirely fictional teaching example: documents, answers and logs are invented. No client result or real platform test.
The claim under review
The service is already open in three countries.
Read
Consult only the six authorised records.
Qualify
Identify the gap between announcement, pilot and general availability.
Prepare
Draft a summary with supporting records and qualifications.
Seek approval
Submit the wording and stop the external action.
Control decision
Proposed controls for this example; this is not an operating agent.
Read D01 to D06
Allowed within the documentary scope.
Prepare a summary
Draft allowed; retain the qualifications.
Write “open in three countries”
Review required: general availability is not established.
Publish or send
Stop: no external action is authorised in this case.
Open the case documents
| Record | Declared origin | Fictional excerpt | Derivation |
|---|---|---|---|
| D01Launch announcement | Organisation documentation | A pilot is planned in three countries, subject to completion of testing. | Direct document |
| D02Editorial repost | Professional blog | The new service is open in three countries. | Copy of D01 |
| D03Aggregated repost | Aggregator | The service is available in three countries. | Copy of D02 |
| D04Register entry | Public register | The organisation is registered and its status is active. | Direct document |
| D05Pilot documentation | Organisation documentation | Access is restricted to invited participants. | Direct document |
| D06Service status | Organisation documentation | One pilot region is enabled; the others are still being prepared. | Direct document |
Read the reasoning
The check occurs before action: general availability goes beyond what the records establish. Calling the draft ready to send would hide that gap.
Useful approval presents the proposed text, sources, recipient and exact action. Permission to summarise is not permission to publish.
The expected result here is a qualified draft and a request for a decision. Nothing is sent and no real tool is connected to this demonstration.
Method references : NIST: AI Risk Management Framework · Content expanded on .
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.
Distinguish tasks from attempts
Ten tasks produce 14 attempts in total. Six tasks succeed immediately, two after retry; two remain unfinished.
| Observed unit | Fictional data |
|---|---|
| Tasks / attempts | 10 / 14 |
| Initial successes + after retry | 6 + 2 = 8 |
| Completed tasks / programme | 8 / 10 = 80 % |
Which rate answers “how many tasks are complete”?
8/10 = 80% describes completed tasks. 8/14 uses a different unit, the attempt. Retain task identifiers, retries, cost and observed effects to avoid counting the same success twice.
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. Record your observations and decisions as checks proceed.
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.
What should a human approval request show?
The exact action, recipient or target, proposed content, sources and qualifications. The reviewer must be able to approve, amend or refuse before any external effect.
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
- Authorised corpus, users, representative questions and actions requiring human approval.
- Suggested process
- Test a small batch, record failures and check recovery and output controls.
- Delivery to agree
- Bounded prototype, source-linked answers and test-case results.
- 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
- Trigger, input examples, transformation rules, destination and retry policy.
- Suggested process
- Trace an input to its destination, simulate a failure, then check recovery and duplicates.
- Delivery to agree
- Controlled workflow, error handling and tested recovery without unintended duplicates.
- Review for this service
- Check a successful batch, a failure and recovery without unwanted duplication.
- Scope boundaries
- Agreed volumes, quotas and external services bound the workflow; further integrations need 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
- Authorised sources, editorial instructions, reviewer roles and approval rule.
- Suggested process
- Trace content from source to draft, check facts and verify approval before publication.
- Delivery to agree
- Draft-to-approval workflow with source checks and explicit responsibilities.
- Review for this service
- Check review, sources and the approval step before any planned publication.
- Scope boundaries
- Generation does not replace verification; automatic publication and additional source use require explicit agreement.
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.
Acceptance check to prepare
Check an imported cell selected by categories, not only its value. Recompute its index, check status and preserve missingness: a correctly sized matrix can still be misaligned.
See scope and pricingDEEPER READING / PRACTICE
Define an agent’s autonomy before building it
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.
Updated on .
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 dataset accompanies the developments to control the expected behavior and possible errors.
Separate reading, proposing and executing
Grant permissions by operation and scope. An agent reading a catalogue needs different access from one changing a database; check proposals before enabling controlled writing.
Test ambiguity, misleading documents and unavailable services. Define acceptable refusal, human handover and an effective interruption before expanding tools.
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.