Agent Task Force · Lyon, France

AI systems for workflows that don’t fit in a chatbot.

We design and build AI systems around your real files, email, tools and operating procedures — with a human in control where the decision matters. Bring one workflow that eats hours every week. In 30 days we show you what it looks like as an AI-native system, running on your data.

Book a workflow call →Read the BrickNote case

24 artifacts · 0 through the gate

Scroll — watch the system run

What do you build?

Focused systems around one workflow. Not a chatbot with your logo on it.

Your team already has the information. It is scattered across documents, inboxes, spreadsheets, meetings and the heads of two experienced people. We build the system that gathers the right context, does the repetitive part, and hands the decision back to a person.

One workflow as an AI-native systemSix inputs — meeting audio, inbox threads, a PDF report, an Excel tracker, site photos and one person’s knowledge — feed one system made of context, tools and reasoning. Its output passes through a human gate before becoming a drafted report, an updated register, replies to approve and an audit trail.MEETING AUDIOINBOX THREADSREPORT_V7.PDFTRACKER.XLSXSITE PHOTOSONE PERSON’S HEADTHE SYSTEMCONTEXTproject memory, not everythingTOOLSextract · compare · draftREASONINGcloud or private modelEU-RESIDENT · TENANT-ISOLATEDHUMANGATEAPPROVEREPORT · DRAFTEDREGISTER · UPDATEDREPLY · TO APPROVEAUDIT TRAIL
Fig. 1 — One workflow as an AI-native system (example). Inputs on a bus, one system, one human gate, validated outputs.

  • Understands the right context

    Project memory built from your files, email threads and past outputs — not a generic index of everything you own.

  • Uses the tools you already run

    Outlook or Gmail, SharePoint or Drive, the ERP export, that Excel tracker. We integrate with what is there; we do not ask you to migrate.

  • Prepares or executes the repetitive work

    Drafts the report, updates the register, prepares the reply. Deterministic tools first; model reasoning only where judgement is needed.

  • Runs on cloud or private models

    Frontier models when quality wins, private or EU-hosted endpoints when the data says so. The system is built so the model can change without a rewrite.

  • Keeps humans in control

    Approval is a feature, not a fallback. The system prepares the action; a person validates it in seconds, and every step is on the record.

For whom?

Teams whose week goes into moving information between places it already exists.

We do not target “companies that want AI”. We look for one pattern.

  • A knowledge-heavy operational workflow
  • Lots of files, email and messages feeding it
  • Repetitive coordination: chasing, consolidating, re-typing
  • Expensive people doing low-leverage admin
  • A fragmented software stack nobody will replace this year
  • Enough volume that automation matters
  • Enough complexity that generic ChatGPT is not enough

Tick what is true of your team.

Where the pattern shows up
  • Construction & project coordination
  • Engineering firms
  • Industrial operations
  • Compliance
  • Procurement
  • Field service
  • Insurance operations
  • Legal operations
  • Logistics
  • Quality management
  • Property management

Where we have seen it — not a list of verticals we are entering. We recognise the pattern; you know your domain.

Which workflows?

The ones that take a Friday afternoon and produce a document nobody fully trusts.

WorkflowWhat the system doesWhat stays human
The weekly progress reportReads what changed across files, email and meeting notes since last week; drafts the report in your template, with sources attached.Reads it, edits the two paragraphs that need judgement, sends.
Meeting → minutes → action registerTranscribes, extracts decisions and actions, and matches each one to the open items from last time so nothing falls through.Validates every item before anything is sent to anyone.
Commitments buried in the inboxDetects promises, deadlines and requests across threads; proposes the register update and a reply.Approves the update. Sends the reply — or not.
Comparing revisionsCompares the new version of a document, plan or contract with the last one; lists what moved and what it affects.Decides what matters.
The person who knows everythingTurns one expert’s tacit process into project memory the rest of the team can query and reuse.Stays the expert. Stops being the bottleneck.

Each row is a workflow, not a feature. We scope one, then build the system around it.

Why credible?

We built one of these for a profession we could reach — and measured it.

BrickNote

Construction coordination (OPC) · France
The problem

A French OPC coordinator runs the weekly site meeting, then spends hours turning audio, photos, emails and last week’s report into a compte-rendu that dozens of contractors act on. Actions fall through between meetings. Everything lives in five places.

The system

A desktop app on the coordinator’s own computer: the chantier folder read in place — reports, plans, schedules, contracts, emails, photos, scans — worksheets that answer from those files with their sources, a réunion recording turned into the compte-rendu, propositions that reach the perpetual action register only on validation, relances and the report sent from the coordinator’s own Microsoft 365 mailbox.

Where it stands

MVP built in September 2026, local-first, with the model of the cabinet’s choosing — Azure OpenAI France Central, Mistral, OpenAI, Anthropic, or a local model. Designed with a practising OPC coordinator; it opens as a free public beta in autumn 2026. No customer claims yet — we publish what we measure.

Read the case study →Watch the 37-second film →bricknote.ai (opens in a new tab)
What we measured
15 / 16
worksheet questions answered on the live model, every citation resolving to a real file
Zero invented pieces. The one miss is covered by the deterministic draft.
0.33
entity recall of the default speech provider on construction jargon
Our own bench caught it before a customer did.
5
model providers behind one transport, a local model among them
The cabinet picks the endpoint; the bench ranks what comes back.
How we build — the four rules that transfer

  • Model-independent

    A provider seam and a scoring bench, so the system moves when the models do — without a rewrite.

  • Tools before autonomy

    Deterministic search, extract, compare and draft. Model reasoning where it earns its place. No vague agent loops.

  • Approval as a feature

    The system prepares; a person validates. Adoption goes up, risk goes down, and there is an audit trail.

  • Private by default

    The data plane is a decision, not a default: on the customer’s own machine where that is the answer, EU-hosted endpoints where it is not. Inbound email is treated as an untrusted boundary.

Agent Task Force is the commercial brand of Apithings, a one-person AI-native studio in Lyon: one senior builder, agents doing the building, judgement kept human.

What can you buy?

The 30-Day AI Workflow Prototype

One workflow. Thirty days. A working system on your data, and a clear answer to one question: is this worth deploying?

Fixed price
€10–30k excl. VAT

€7–15k excl. VAT for smaller companies and narrower workflows. A price band is indicative: the fixed price is the one written in the signed proposal.

The month
  1. W1

    Map and decide

    We map the current process with the people who run it, and decide what AI should — and should not — automate.

  2. W2

    Build against real data

    A working prototype on your actual files, threads and exports. No sample data.

  3. W3

    Connect and test

    The essential tools connected; tested on real scenarios with the people who will use it.

  4. W4

    Harden, demo, roadmap

    Evaluation criteria, security and deployment assumptions, a demo to your team, and the production roadmap.

You leave with
  • A working prototype you can run
  • The essential integrations
  • A test workflow and its evaluation criteria
  • A recorded demo for the people who were not in the room
  • Security and deployment assumptions, written down
  • A roadmap to production — with a price
How it is priced

Set against the value of the workflow — the people it touches, the hours it burns, the integrations it needs. Never against developer-days.

What it is not
  • A six-month transformation programme
  • Slideware-first AI strategy
  • A daily rate, or staff augmentation
Before

Not sure which workflow? A 2–5 day Workflow Discovery comes first: workflow map, feasibility, data and connector inventory, a prototype concept and an ROI hypothesis.

After

Worth deploying? Productionisation — permissions, private deployment, monitoring, evaluations, training — is scoped and priced separately, and only once the prototype has proven value.

One offer, on purpose. A second one appears when this one has been sold twice.Book a workflow call →
How do you start?

A 30-minute call about one workflow. No deck required.

Come with the workflow that annoys you most. We will tell you honestly whether it is a fit, and roughly what a prototype would cost.

Book a workflow call →

One prototype at a time · remote-first from France · EU hours

Bring answers to five questions
  1. What does your team repeat every week that feels unnecessarily manual?
  2. Where do people copy information from one tool into another?
  3. Which report is painful to produce?
  4. What depends on one experienced person knowing everything?
  5. Which task would you never let a system do without someone’s approval?