Useful AI, built into the work.

Last Word designs and ships agents, automations, scraping systems and websites that work inside real teams and real tools.

Start with one workflow

Drawing: many lines enter from your inbox, drive, CRM and the web, wander, and pass through one narrow gate into a single calm line. One line does not fit the gate and stops at a red mark, where a person decides.

The valuable part of AI is rarely the chat window. It is the workflow around it.

Teams do not need another spectacular prompt. They need fewer repeated checks, fewer copy-pastes, cleaner handoffs, better monitoring, and tools that say when they are unsure.

We take one recurring workflow, understand how it really behaves, then build the smallest reliable system that can absorb it without creating a new mess elsewhere.

How to read the line

  • copper: the system at work
  • pencil: first pass, raw work
  • red: a person decides

What we build

Six kinds of system. Each one leaves a different line.

  1. AI agents for real workflows

    Branches out, comes back, hands the hard case to a person.

    Agents that qualify, retrieve, draft, update tools, prepare decisions and escalate when judgement matters.

  2. Process automation

    The same rhythm, every time.

    Connected workflows for recurring work: reminders, reports, CRM updates, documents, exports and internal handoffs.

  3. Web scraping and market watch

    Goes to the edge and samples it.

    Scrapers, monitors and dashboards with traceable sources, explicit missing data and alerts that are useful before they are pretty.

  4. Local LLM and data-sensitive systems

    Stays inside a closed boundary.

    Private or controlled AI architectures for teams that need to use sensitive documents without losing governance.

  5. Creative websites

    The line turns into letterforms.

    Fast editorial sites, landing pages and product pages with strong copy, sober motion and a real design point of view.

  6. Custom prototypes

    Drawn as a draft before anyone commits.

    When the shape is not obvious yet: dashboards, internal tools, agentic workflows and proof-of-value experiments.

All services

First the sources. Then the system. Then a clear decision on what to build, stop or improve.

  1. Map the work , in pencil

    We start with the actual sources, exceptions, decisions and handoffs. The shape of the system comes from the work, not from the model.

  2. Build a small proof , in ink, once

    A first version must run on real cases, expose its limits and make the next decision easier. No staged demos, no fake data.

  3. Make it observable , instrumented

    Logs, review points, failure states and human override are part of the product. If nobody can supervise it, it is not ready.

We started with our own website.

No client data and no borrowed success story. We ran a fixed set of French and English pages through the same checks: 2 domains, last-word.fr and last-word.ai, at 2 viewports, desktop 1440 × 900 and mobile. Every page left a timestamped record. Here is one of them, replayed at its real duration.

target
https://last-word.fr/ fr, home
viewport
desktop, 1440 × 900
One record, replayed at its true length: started 11:29:24.284Z, completed 11:29:27.923Z, 3.639 s. Status 200. 0 console messages, 0 page errors, 0 failed requests.

The record, verbatim

{  "target": {    "locale": "fr",    "kind": "home",    "url": "https://last-word.fr/"  },  "viewport": {    "id": "desktop",    "width": 1440,    "height": 900  },  "startedAt": "2026-08-25T11:29:24.284Z",  "completedAt": "2026-08-25T11:29:27.923Z",  "status": 200,  "finalUrl": "https://last-word.fr/",  "navigationError": null,  "consoleMessages": [],  "pageErrors": [],  "failedRequests": []}
  1. Lines 2 to 6Our own French site. No client data.
  2. Lines 12 and 133.639 s. One record, one point in time.
  3. Line 14HTTP 200
  4. Lines 17 to 190 console messages, 0 page errors, 0 failed requests.
  5. Whole runForms not submitted.

What this record does not prove

  • It is one record, one point in time.
  • We did not submit the forms.
  • It says nothing about conversion, long-term reliability or pages outside the sample.
  • The full bundle stays internal. The hash identifies it; it does not make it downloadable.

A separate reviewer checks the findings, separates evidence from assumptions and can stop the release.

Bundle SHA-256, verified 25 August 2026:
64d1319ebda7b3016c3ed33971f31f1eda3d92882e6237feb63d05285a9a28fa

Read the full verification note

Bring the awkward process. We will make it testable.

A vague idea, a painful routine, a monitoring need, a site that needs a point of view: that is enough to start.

Start with one workflow