Skip to content
Loïc Chossière AI Consultant

From the real process to the measured gain, with no detour through the hype.

Only one question matters: where do you spend time that a machine would handle just as well? The other six steps are just engineering.

  • How much time does this process actually take us?
  • Where do we start with AI?
  • Can our data leave the company?
  • Who validates what?
  • Are our experts going to be replaced?
  • How do we prove the gain with hard numbers?
  • We tried already, it did not stick.
  • How long before the first results?
  • Do we have to change our tools?
  • Who maintains all this after you?
  • Will it pass compliance?
  • What if the model gets it wrong?
Typical duration
2 to 12 weeks
Entry point
A process, not a tool
Stopping condition
ROI not demonstrated
Exit
Agents in production, memory included

The seven steps

  1. Process discovery

    Understanding how the work is actually done.

    I sit down with the operational people. Who decides what, on which information, in which tool, with which exceptions, how often. Only the real gesture counts, including the famous unofficial spreadsheet everyone uses.

    Deliverable
    A step-by-step map of the workflow, exceptions included
    Prerequisites
    • 3 to 5 one-hour interviews
    • read access to your tools

  2. Process economics

    Putting a price on every step.

    Human time x frequency, plus the time lost to errors, rework and delay. This is the step nobody takes, and it is the one that makes the decision obvious: I almost always find two or three steps that cost more than the system replacing them would.

    • € / transaction The real unit cost
    • Minutes / step Measured, not estimated
    • Error rate And the cost of rework
    • Cycle time The delay as the client sees it

  3. Mapping the AI opportunities

    The right technology, step by step.

    Every step gets a label, and the most frequent one stays “human only”. Putting a language model where a rule would do costs as much as the other way round.

    • Human only
    • Deterministic automation
    • LLM
    • RAG
    • Agent
    • Skill
    • Computer use
    • API
    • MCP
    • Dedicated sub-agent
    • Shared memory
    • Human validation

  4. Prioritising by ROI

    A priority queue, not a wish list.

    Financial impact×frequency×automatability÷complexity & risk

    I build first what pays quickly and breaks little. The first system has to fund the next ones; it is the only way the rest gets accepted internally. It is also what separates an AI budget that falls into the trough from one that reaches the plateau.

  5. Architecture

    Decisions written down.

    Which model for which task, at what cost and what latency. Where the data lives and must not leave. Which human validation point is mandatory. What the system does when it does not know, the most important case of all.

    • Cost, latency, reasoning depth, context size
    • Data sovereignty, GDPR, retention
    • Evaluation, traceability, behaviour under doubt

  6. Going to production

    It runs at your place, and you know why.

    I build, I connect to your systems, I ship to production with your teams. Execution logs, explicit human validation at the critical points, documentation, handover, and one person on your side able to evolve the system without calling me. No black box, no organised dependency.

  7. Measurement

    The proof, before and after the setup

    This is where I come from: tagging plans, server-side collection, attribution, consolidated multi-source reporting. Applied here to automation, it gives one simple table: cost per transaction, minutes per process, capacity returned, error rate, cycle, conversion, measured before and measured after.

    • Server-side & first-party tracking, a clean tagging plan
    • Consolidated reporting: a single decision dashboard
    • This step is also available as a standalone engagement

What it looks like once in production.

One trigger, a queryable memory, a lead agent handing work to its sub-agents, and outputs one of which always goes through a human.

Agent pipeline, live

1 lead agent, 3 sub-agents, 0 errors

  1. Trigger

    New case file

    Arrives by email or from the document store

  2. Memory

    Sourced search

    Your queryable base, every extract cited

  3. Lead agent

    Writing

    Hands the work to three sub-agents

  4. Output

    Extracted values

    42 values, each one sourced

  5. Output

    Specification draft

    9 pages, 4 points flagged

  6. Output

    Human validation

    The expert decides, the system learns

Received: new case file to qualify, reference 2026-118

Files processed
1,248
Sources cited
11,400
Average turnaround
3 min 48

The anatomy of a system

4 elements

  1. The lead agent

    The one steering the process end to end, handing out the work and reporting on what it did.

  2. Its sub-agents

    One gesture, one specialist: search, extract, check, draft, format.

  3. Continuous memory

    Your rules, your templates, your past decisions, reused by every system.

  4. The improvement loop

    What your experts correct becomes a rule. The system starts higher next time, indefinitely.

  • Skill
  • RAG
  • Agent
  • Validation
See this anatomy on ten real cases 

The proof, before and after.

Tracking, attribution, consolidated reporting: where I started. Every system I deliver is measured against the before, in your own analytics tools, not in a separate spreadsheet.

  • Google Tag Manager
  • TagCommander
  • Google Analytics
  • Piano Analytics
  • Ad platforms
  • Looker Studio

Step zero

Bring the process you want to stop putting up with.

20-minute diagnosis  The ten use cases See the pricing