Skip to content
Loïc Chossière AI Consultant

Starting with the work to automate

10 standard services cover most of my engagements, as much in a laboratory or an engineering office as at an independent consultant, a manufacturer or a services SME. Each is built the same way: a lead agent that steers, specialised sub-agents, a memory holding your rules and decisions, and an improvement loop fed by your experts’ corrections.

  • Lead agent
  • Specialised sub-agents
  • Continuous memory
  • Domain skills
  • Sourced RAG
  • Tool-equipped agents
  • MCP on your tools
  • Human validation

  1. Technical documentation

    Specifications and technical dossiers

    Technical writing eats expert hours in document research before the first sentence. The system does the research, proposes a first version in the right format, and leaves the decision to the expert.

    Lead agent

    Specification writer

    Steers the production of a complete document, from the stated need to a draft ready for review.

    Sub-agents

    • Document researcher Finds standards and datasheets, cites every reference
    • Data extractor Pulls the values, compound by compound
    • Checker Verifies every claim against its source
    • Formatter Applies your template and your headings
    Continuous memory
    Your templates, your approved phrasings, the compounds already qualified and the past technical calls.
    Improvement loop
    Every correction by the expert becomes a writing rule. The next draft starts higher.
    • Skill de rédaction
    • RAG sur votre documentation
    • MCP SharePoint
    • MCP Google Drive
    • OCR de PDF techniques
    • Gabarits Word
    • Citations obligatoires

  2. Quotes and tenders

    Pricing, quotes and tender responses

    A client specification comes in, a reasoned quote comes out. The system reads the request, finds comparable jobs, proposes the line items and the outline of the response.

    Lead agent

    Response manager

    Orchestrates reading the tender, pricing it and assembling the response.

    Sub-agents

    • Tender analyst Requirements, constraints, traps, missing documents
    • Estimator Brings back comparable jobs and the prices actually charged
    • Technical memory Assembles approved references and appendices
    • Presentation designer Builds the pitch deck to your brand
    Continuous memory
    Past quotes, prices actually charged, observed gaps, approved technical memoranda, win rate by type of job.
    Improvement loop
    The outcome of every job, won or lost, and the real margin flow back into memory and correct the next estimates.
    • Agent de chiffrage
    • RAG sur l’historique de devis
    • MCP ERP
    • MCP Excel
    • Sorties structurées
    • MCP Google Slides
    • MCP PowerPoint
    • Charte de présentation
    • Validation humaine obligatoire

  3. Compliance

    Regulatory watch and compliance checks

    The rule changes, your documents do not know it. The system watches the official sources, flags what actually concerns you and points at the documents to revisit.

    Lead agent

    Regulatory watcher

    Watches the official sources and turns changes into actions on your documents.

    Sub-agents

    • Source scanner Follows official publications by topic
    • Relevance filter Discards what falls outside your scope
    • Document controller Compliant, to check, non-compliant
    • Alert writer A short note, with the reference
    Continuous memory
    Your regulatory scope, the history of rule versions and the documents already checked.
    Improvement loop
    Every alert judged off-topic sharpens the filter. The noise drops week after week.
    • Agent de veille planifié
    • RAG réglementaire
    • Sources officielles citées
    • MCP Slack
    • MCP Teams
    • Journal d’audit

  4. Incoming requests

    Qualifying and routing what arrives

    Sales enquiries, tenders, complaints, applications. Everything is read, qualified against your criteria and sent to the right person with the reasoning attached.

    Lead agent

    Request router

    Reads everything that comes in, qualifies it against your criteria and sends it to the right person.

    Sub-agents

    • Request reader Understands the request and extracts the facts
    • Scorer Applies your grid, not an opaque score
    • Enricher Adds the company context and the history
    • Acknowledgement writer A first reply, ready to send
    Continuous memory
    The history of requests, the routing decisions taken, and what was actually signed after.
    Improvement loop
    The grid recalibrates on the deals actually won, not on the initial hunch.
    • Skill de qualification
    • Sorties structurées
    • MCP HubSpot
    • MCP Salesforce
    • MCP Gmail
    • Règles de routage versionnées

  5. Producing deliverables

    From field notes to the client deliverable

    Interviews, measurements, client data and past analyses become a structured, reasoned and sourced report. This is the heart of a consultancy or an engineering office: the system prepares, the expert decides.

    Lead agent

    Deliverable production assistant

    Assembles a complete deliverable from raw material: interviews, data, analyses, references.

    Sub-agents

    • Interview synthesiser Transcripts and notes reduced to facts and quotes
    • Analyst Cross-references client data with your frameworks
    • Recommendation writer Findings, options, a reasoned recommendation
    • Readout designer The presentation deck, to your brand
    Continuous memory
    Your analysis methods, your standard deliverable outlines, past engagements and the positions you stand by.
    Improvement loop
    Every reviewed deliverable enriches the standard outline and the reference phrasings. The next engagement starts higher.
    • Skill de plan type
    • RAG sur vos missions passées
    • Transcription d’entretiens
    • MCP Google Docs
    • MCP Google Slides
    • MCP Notion
    • Sorties structurées

  6. Research and state of the art

    Background research that keeps itself current

    Scientific publications, patents, standards, market players, calls for projects. One scanner per topic, a synthesis ranked by importance and sourced, instead of a shared folder that ages.

    Lead agent

    Research lead

    Keeps the state of the art current and turns it into material your experts can use.

    Sub-agents

    • One scanner per topic Publications, patents, standards, players, funding
    • Deduplicator Compares against past weeks
    • Synthesis writer A sourced note, ranked by importance
    • Distributor The right note to the right person, at the right time
    Continuous memory
    Every past finding, the versions of each synthesis and what the team has already read.
    Improvement loop
    What your experts judge useful rises to the top, the rest leaves the synthesis.
    • Agents de scan parallèles
    • Sources scientifiques et brevets
    • Extraction de pages web
    • Déduplication sémantique
    • MCP Notion
    • MCP Slack
    • Versionnage des synthèses

  7. Internal knowledge

    Customer support and the company memory

    Procedures, solved cases and past decisions stop living in three people’s heads. Knowledge becomes a queryable layer that answers with its sources.

    Lead agent

    Support assistant

    Answers from your internal knowledge, and knows when to hand over.

    Sub-agents

    • Knowledge base searcher Hybrid search across procedures and tickets
    • Reply writer In your tone, with its sources
    • Confidence judge Chooses to escalate rather than invent
    • Procedure writer Turns a new case into a procedure
    Continuous memory
    Solved tickets, approved procedures, escalation decisions and the corrections made by the team.
    Improvement loop
    Every corrected reply becomes the reference, every unprecedented case becomes a proposed procedure.
    • Base de connaissance interne
    • Recherche hybride
    • Reranking
    • MCP Zendesk
    • MCP Notion
    • Escalade humaine

  8. Published expertise

    Your expertise written in your voice, without losing your weekends

    Whether you publish expert notes, technical articles or talks, the work on the voice is done once, in a maintained skill. Everything else inherits it, with the author always validating.

    Lead agent

    Brand editor

    Produces in your voice across every channel, with the author always validating.

    Sub-agents

    • Angle researcher Feeds the drafts with real material
    • Writer in your voice A maintained voice skill, not a throwaway prompt
    • Fact checker Checks figures, claims and the legal frame
    • Per-channel adaptation Site, talks, campaigns
    Continuous memory
    The voice skill, your approved content and past performance by angle and by format.
    Improvement loop
    The angles that perform rise, the rejected phrasings become explicit prohibitions.
    • Skills éditoriaux par rôle
    • RAG sur vos contenus validés
    • MCP Notion
    • MCP WordPress
    • Vérification des affirmations
    • Validation de l’auteur

  9. Measurement and steering

    The layer above: measurement, reporting, decision

    The one built last, because it rests on everything else. It is also where I come from: without clean measurement, no gain is demonstrable.

    Lead agent

    Chief of staff

    The layer above the other agents: measurement, consolidation, decision.

    Sub-agents

    • Data collector Server-side and first-party tracking, consent respected
    • Source reconciler One single truth per indicator
    • Morning briefing writer One screen, with the reasoning
    • Gain calculator Before and after, per process
    Continuous memory
    The before-and-after series per process, the costs, the volumes and the decisions taken in committee.
    Improvement loop
    The gap between forecast gain and observed gain corrects the priority of the next builds.
    • Tracking serveur et first-party
    • Couche de données unifiée
    • Attribution multi-sources
    • Tableaux de bord décisionnels
    • MCP Slack
    • Suivi des coûts et des tokens

  10. Industrialised practice

    The consultant who industrialises their own expertise

    Take a consultant running the same engagement for several clients: one knowledge base per client, read before every task, then one skill per recurring gesture of their trade. Onboarding a new client takes minutes instead of starting over.

    Lead agent

    Account conductor

    Chains the skills into a workflow, with explicit human validation at each key step and a log of every run.

    Sub-agents

    • Client base Context, constraints, tone, prohibitions, quantified goals
    • Situation analysis Account data, market comparison, gaps to explain
    • Deliverable production Recommendations, content and reports in the client’s format
    • Steering and trade-offs Decision rules transposable from one client to the next
    Continuous memory
    One reusable base per client, and the decision rules that worked elsewhere, transposable from one account to another.
    Improvement loop
    Every new client enriches the base template, and validated rules harden instead of being reinvented.
    • Socle de connaissance par client
    • Un skill par geste métier
    • Orchestration en workflow
    • Validation humaine explicite
    • Journal des exécutions
    • MCP Markifact
    • MCP sur les outils du client
    • MCP sources de données métier
    • Passation et documentation d’usage

And at your place

None of these examples beats your real process.

Bring the one that takes you the most time. In twenty minutes we know whether it belongs to one of these families, and what it is worth.

20-minute diagnosis  The method