AI that hands you backentire working days.

1. You dictate

   22 s of speech

   A voice note from the site, in your own words, with no form to fill in.

   Voice note — 00:22

   “Perrin job, in Annemasse. 42 m² of oak flooring,
   strip out the old covering, two days of laying
   with two people. They want to start in March.”

2. The system understands

   3.1 s

   Transcription, then structured extraction: client, line items, quantities, timing.

   Extracted fields

   client        : "Perrin"  → existing record #418
   site          : "Annemasse (74)"
   item_1        : "strip out old covering" · 42 m²
   item_2        : "oak flooring supplied & laid" · 42 m²
   labour        : 2 days × 2 people
   target_start  : "March 2026"

3. Pricing is applied

   0.8 s

   YOUR rate card, your margins, your terms. No invented prices.

   Pricing · 2026 rate card

   Strip out covering    42 m² × €14.00   →     €588.00
   Oak flooring laid     42 m² × €68.00   →   €2,856.00
   Labour                 4 d × €320.00   →   €1,280.00
   Waste removal              flat rate   →     €180.00
   ─────────────────────────────────────────────────────
   Total excl. VAT €4,904.00   ·   incl. VAT €5,884.80

4. The document is generated

   1.4 s

   Your template, your logo, your legal terms — in French, English or German.

   QUO-2026-0418.pdf

   Quote no. QUO-2026-0418 · Valid 30 days
   Terms: 30 % on order, balance on completion
   Electronic signature built in
   Ready to send — waiting for your approval

5. You approve, it goes out

   30 s of your time

   Sent, filed, and a follow-up scheduled if the client goes quiet for 8 days.

   Sent · 14:38

   Quote sent to s.perrin@…  ·  read receipt on
   Filed under client record #418
   Automatic follow-up scheduled for 14 March
   Alert if the quote is opened twice

You read the quote before it goes out. That is the only step left, and it is deliberate: we never let a price reach a client without a human having seen it.

1. The message arrives

   instant

   Email, web form, WhatsApp or chat: everything comes in through one pipe.

   contact@ · new message

   From: [email protected]
   Subject: Re: proposal — urgent question
   
   “Hello, following our March discussion we need to
   equip two more sites before year end.
   Could you call us back quickly?”

2. It gets read and classified

   1.2 s

   Request type, urgency, language, buying intent — in a single pass.

   Qualification

   type        : "contract extension"
   urgency     : high  (deadline stated)
   intent      : commercial — buying signal
   language    : en
   est_value   : high (2 sites × existing contract)
   spam        : no

3. Context is rebuilt

   0.9 s

   Client history, live contracts, last exchanges: nobody starts from scratch again.

   Client context recovered

   Mercier Group · client since 2023 · 2 active contracts
   Last contact: 14 March, proposal sent, no reply
   Contact: purchasing@ — decision maker identified
   Outstanding balance: €0

4. The reply is ready

   2.4 s

   Written in your voice, with the right facts, never sent without you.

   Draft reply

   “Hello Ms Chevalier, thank you for getting back to us.
   We priced a 2-site extension back in March — I can offer
   Thursday 11:00 or Friday 09:30 to lock the schedule
   before year end.”
   
   → Slots genuinely free in your calendar

5. The case exists

   0.6 s

   Created, assigned and prioritised in your CRM. No re-typing, nothing forgotten.

   Opportunity #2291 created

   Assigned to: Julien (account manager)
   Priority: high · follow-up in 2 days if no reply
   Email thread attached automatically
   Slack notification sent

Your salesperson reads it and hits send. They spend their time selling, not reconstructing the context of every message.

1. Your documents come in

   once, at setup

   PDF, Word, scans, intranet, archived email. Badly scanned documents included.

   Document base indexed

   2,184 documents  ·  12,407 pages
   Contracts, quality procedures, spec sheets, minutes
   Versions and validity dates preserved
   Access rights inherited from your directory

2. They become searchable

   ~40 min for 12,000 pages

   Chunking, vectors (pgvector), keyword index. Search understands meaning, not just words.

   Indexing

   chunks       : 38,942
   embeddings   : pgvector · 1024 dim
   lexical_index: BM25 (hybrid search)
   metadata     : department, version, date, confidentiality

3. Someone asks a question

   instant

   In plain language, the way you would ask an experienced colleague.

   Question asked

   “What is the withdrawal period under the Legrand
   framework contract, and did the 2025 amendment change it?”

4. The answer is built

   2.8 s

   Hybrid retrieval, then an answer written only from the passages found.

   Answer

   “The period is 30 calendar days (clause 8.2 of the
   framework contract). Amendment no. 2 of 04/2025 extends it
   to 45 days for orders above €50k only.”
   
   source: framework-contract-legrand-v3.pdf · p. 14
   source: amendment-02-2025.pdf · p. 2

5. Everything is traceable

   —

   Sources cited, questions logged, permissions honoured. An unsourced answer is not returned.

   Guardrails active

   Every statement points to a specific passage
   No source found → “I don’t know”, never an invention
   Documents outside the user’s rights: invisible
   Question log available for review

On sensitive matters — legal, safety, compliance — the answer stays a decision aid. The assistant shows where to read; the expert decides.

1. The invoice arrives

   instant

   Accounts inbox, scanner, or supplier portal. Any format.

   accounts@ · attachment

   From: [email protected]
   Attachment: INV-2026-11482.pdf (2 pages, scan)
   Subject: November invoice — order PO-8841

2. It is read in full

   4.2 s

   Header, lines, VAT, references — even on a crooked scan.

   Extraction

   supplier : "Delmas Steel Ltd" → vendor #V-204
   invoice  : INV-2026-11482 · 12/11/2026 · net 30 EOM
   order    : PO-8841
   lines    : 14  ·  net_total : €18,402.60
   vat      : 20 % → €3,680.52

3. Matching happens

   1.1 s

   Invoice ↔ purchase order ↔ delivery note, line by line.

   Matching PO-8841

   Line   Ordered   Received   Invoiced   Status
     1     120 u      120 u      120 u      ✓
     2      40 u       40 u       40 u      ✓
     …
    11      15 u       15 u       15 u      ⚠ €6.80 vs €6.20
   Discrepancy found: +€9.00 net on line 11

4. Exceptions surface

   —

   What matches goes through. What doesn’t lands on a desk, with the explanation attached.

   1 exception out of 14 lines

   13 lines compliant → posted with no human touch
   1 line out of tolerance → sent to accounts
   Reason: unit price above the contracted rate
   The supplier has been notified automatically

5. The entry is posted

   0.7 s

   Accounting entry created, document archived, payment due date set.

   Posted to the ERP

   Entry 401-V204 · €18,393.60 net
   Source document attached (original PDF kept)
   Due: 31/12/2026
   Complete audit trail, reversible

Accounting no longer keys data in: it arbitrates exceptions. That is the only place where its judgement is worth anything.

1. The review is detected

   < 5 min after posting

   Every platform, every location, continuously.

   New review · Geneva-Cornavin

   ★★☆☆☆  ·  4 minutes ago  ·  Sofia M.
   
   “Staff were fine but 35 minutes of waiting on a Tuesday
   lunchtime with the room half empty. Shame, because the
   food is genuinely good.”

2. It gets analysed

   1.1 s

   Rating, real theme, severity, and whether a gesture is warranted.

   Analysis

   rating     : 2/5
   theme      : "waiting time" (6× this month)
   positive   : "food quality" — worth echoing in the reply
   severity   : medium · customer recoverable
   gesture    : suggested (second visit)

3. The reply is written

   2.2 s

   In your brand voice, no copy-paste formula, no promise you can’t keep.

   Draft reply

   “Hello Sofia, thank you for taking the time — and glad the
   food landed well. 35 minutes on a Tuesday lunchtime is not
   our standard: we reworked the lunch service this week.
   Say hello next time you’re in and we’ll look after you.”

4. You approve in one gesture

   15 s of your time

   Publish, edit, or decline. Nothing goes out without a human.

   Published · reply time: 11 min

   Reply published on the correct location listing
   Customer history updated
   Location response rate: 100 %

5. The signal comes up

   —

   One review is an anecdote. Six on the same theme is an operations problem.

   Trend alert · Geneva-Cornavin

   “Waiting time”: 6 mentions in 30 days (+200 %)
   Concentrated on Tuesday lunch service
   Location average: 4.1 → 3.8
   Alert sent to the site manager

You keep control of every publication. The goal is not to reply automatically: it is to never again leave a review unanswered for lack of time.

1. The document stays with you

   —

   On your server, or in a Swiss or European datacentre you choose.

   Data perimeter

   Storage      : your infrastructure
   Processing   : your infrastructure
   Outbound traffic to a third-party vendor: none
   AI sub-processor to declare under GDPR: none

2. The model runs locally

   1.8 s per answer

   An open model executed on a dedicated GPU, on site. No external API, no per-request cost.

   Execution

   engine        : local inference on dedicated GPU
   model         : open model, versioned and pinned by you
   external_api_calls : 0
   cost_per_request   : €0 (hardware amortised)
   training_data      : no client data reused

3. The answer is produced

   real time

   The same usefulness as a consumer assistant — on files you could never have pasted into one.

   Case summary

   “Case no. 2026-1187 · summary as of 12/11
   Three missing documents identified, including the employer
   certificate (chased on 04/11, no reply).
   Statutory deadline: 28/11.”
   
   sources: 4 documents from the case file, cited

4. Everything is logged

   —

   Who asked what, when, on which file. An audit trail you can show an inspector.

   Audit log

   2026-11-12 09:41 · officer M.D. · case 2026-1187 · summary
   Timestamp, user, document, action — all retained
   Export for audit or supervisory authority
   Retention configured to your policy

5. Shadow use ends

   —

   Teams finally have a sanctioned tool. Shadow IT loses its reason to exist.

   What this fixes

   No more confidential documents pasted into a public chat
   An AI policy you can enforce, because an alternative exists
   Predictable cost: no per-token billing
   Pinned model: behaviour doesn’t change overnight

Nothing is decided by the machine. On a regulated case, the AI prepares and retrieves; the responsible person signs.