IA.TERRA

Five roles to hand over, not fifty buttons to learn.

Terra's automated assistance is not bolted onto your work as a set of options: it takes on five jobs somebody does by hand today. Here they are, what they produce, and where each one stops.

The machine prepares. You are the one who commits the company.

Elsewhere you will be promised a team that decides on its own. Not here. Everything IA.TERRA produces arrives as an editable draft, with the origin of every figure — because a price submitted, an order acknowledgement or a payment commits your company, and none of those three acts can rest on a reading nobody checked. What you gain is not the decision: it is all the work that comes before it.

The gain

What changes, job by job

  • Answering a tender

    Two to four days of writing and pricing.

    Analysis, pricing and technical statement in under an hour — to review.

  • Pricing

    Old quotes picked over by hand, from memory.

    Prices found automatically, with the evidence behind them.

  • Site report

    Retyped in the evening or the next day.

    Dictated on site, structured automatically.

  • Customer and contract follow-up

    Depends on how alert each person is.

    Automatic, recorded deadlines and reminders.

  • Supplier invoice received by email

    Attachment downloaded, renamed, then retyped by hand.

    The mailbox is checked on its own, the invoice read and created as « to verify ».

  • Invoice / payment matching

    Ticking off the bank statement against open invoices, line by line.

    Receipts extracted and matched, each still to be confirmed.

  • Supplier acknowledgement

    Silence is discovered on the day the delivery fails to turn up.

    Chased at 24 h, 48 h, 72 h — then raised as a dispute.

  • Debt collection

    The chaser goes out when somebody remembers.

    48 h after the due date, then every 72 h, working days only.

  • Disputes and non-conformity

    The discrepancy is settled on the phone and leaves no trace.

    A statement written from the discrepancies and the photos, sent and recorded.

  • Company memory

    Scattered across drives and mailboxes.

    A body of work you can search and reuse.

The augmented team

Who does what, in practice

The reader

Incoming documents

Everything that arrives as a PDF, a photo or a voice note, and that today has to be retyped into a screen.

Supplier invoices
Issuer, number, dates, net, VAT, total and the line-by-line detail — up to sixty lines, carried into the entry form.
Order acknowledgements
Plenty of suppliers reply with a PDF, not a form: the confirmed dates and prices come back onto the order instead of being re-keyed.
Supplier contracts
The clauses that bear on invoicing — caps, penalties, indexation, payment terms, year-end rebate tiers — each with the source text quoted.
Tender specifications
The customer's pack becomes the list of items to put out to tender. It is what stops a request for quotation going out with three items instead of twenty.
Bank statements
CSV, Excel or PDF, whatever format the bank uses: receipts are extracted, then matched against open customer invoices.
Statements of need
A voice note forwarded from WhatsApp, a dictation, a spreadsheet: enough to open a purchase request or a tender without going via the spreadsheet stage.
Product datasheets
A product's catalogue record is created from the manufacturer's documentation.

What it does not do

It creates no data. It pre-fills a form that someone reads over and confirms. An order acknowledgement fixes a contractual commitment: it cannot rest on an automatic reading nobody checked.

The estimator

Tenders and quotes

The price comes from what you have already sold, not from what a model finds plausible.

Breaks the pack down
A tender pack becomes a list of priceable items, with the requirements and the risks noted along the way. Reading a three-hundred-page pack is the job a model is best at.
Prices from history
Prices come from past quotes and the catalogue, by median — with no model involved. The result is reproducible and defensible: two runs give the same figure, and you know where it came from.
Cites its sources line by line
For each item: how many comparable lines were found, and the quote numbers — including the ones that were accepted.
Estimates labour from actuals
From the hours actually clocked on completed job sites, not the budgeted ones — the gap between the two is precisely what derails jobs.
Writes the technical submission
From the real references pulled out of the database. Never from a job site invented to pad the list.

What it does not do

It does not fill in the blanks. An item with no precedent comes out marked “to be priced”, at the top of the screen; if it gets an order of magnitude, that is labelled “benchmark”, is never added into the priced total and still counts as outstanding. This is deliberately stricter than a model that would fill everything in: on a large job, ten plausible lines within 20% are enough to turn a won job into a lost one, and nobody re-reads them once they look priced.

The checker

Discrepancies and disputes

What was ordered, what arrived, what is invoiced — and the gap between the three.

Order against the supplier's reply
Confirmed prices and lead times compared with what was ordered, line by line, as soon as the supplier replies from their portal.
Invoice against order and receipt
The match surfaces the lines that do not tally, with the amount of the difference.
Writes the non-conformity report
From the quantities ordered and received, the invoicing discrepancies and the photos taken on site — the document goes to the supplier instead of staying a spoken conversation.

What it does not do

Detection does not depend on the model. Any price difference above 5% is raised by a calculation, independently of the AI, so that it is never missed if the AI fails or is not configured. The model only writes the readable summary.

The analyst

Management reports

The report nobody has time to write, on the figures the system already holds.

One report per area
Management, engineering, operations, purchasing, HR, sales, assets, accounting, requests — each on its own figures, plus a board-level summary.
An action plan, not an observation
The follow-ups proposed are structured as PDCA and split by time horizon, with a link to the record concerned.
Comes out in the expected format
Word, PowerPoint and Excel in your house style, ready to circulate at a board meeting without reformatting.

What it does not do

A report does not escape the permissions of its area. The HR report — names, certifications, absence, sick leave, expenses with amounts — can only be opened, exported and deleted with HR access, whatever page it is requested from.

The writer

Field and customer relations

What gets said on a job site ends up written down, instead of ending up forgotten.

The job report is dictated
Speech is transcribed with the speakers separated, then laid out: decisions taken, actions with an owner and a due date, key points.
Three levels of write-up
Compact for a callout, detailed for the customer, decision-focused for management — risks, costs and priorities.
Leads into the next step
The report proposes what comes next — a quote, an invoice, another job — and opens the matching screen.
Sets up the technician's day
Today's jobs, absence, approvals pending, certifications about to expire: a morning briefing in a few lines.
Writes the sales campaigns
Proposes campaigns, writes the messages and tailors them to each recipient.

What it does not do

It does not choose who gets contacted: the segments come from verifiable queries on your data, not from the model. And nothing goes out unread — a false sales message commits the company, and “the AI wrote it” is not an excuse that exists on the customer's side.

The full detail

Every function, and what powers it

We distinguish what a model produces from what a calculation establishes. Nobody else does, and yet it is the only question that matters when a figure goes out to a customer.

Model

A language model is involved. The output is a draft, read over before anything is committed.

Computed

No AI at all: queries and rules. Two runs give the same result, and it is defensible — which is why discrepancy detection does not depend on the model.

Voice

Speech transcription, with speaker separation on long recordings.

Reading and extraction

A document arrives, its data goes into the system. Scanned or native, it makes no difference: a text PDF goes through extraction then a text model, a photo through multimodal vision, which does the OCR and the structuring in one pass.

  • Supplier invoice

    Purchasing › Invoices

    Issuer, number, dates, net, VAT, total and up to sixty detail lines.

    Model
  • Order acknowledgement

    Purchasing › Orders

    Confirmed dates and prices carried onto the order, instead of being re-keyed from the supplier's PDF.

    Model
  • Supplier contract

    Purchasing › Contracts

    Clauses that affect invoicing, payment terms and year-end rebate tiers, each with its source text quoted.

    Model
  • Tender specification

    Purchasing › Tenders

    The customer's pack becomes the list of items to put out to tender, as a draft.

    Model
  • Bank statement

    Accounting

    Receipts extracted from a CSV, an Excel file or a PDF, whatever format the bank uses.

    Model
  • Statement of need

    Purchasing › Requests

    A voice note, a dictation, a spreadsheet or free text becomes a draft purchase request or tender.

    Model
  • Product datasheet

    Master data › Catalogue

    The manufacturer's documentation creates the catalogue record — description, part number, classification.

    Model

Decision support

Whether to bid, at what price, with whom, and where the discrepancies are. Detection rests on computation, never on the model: a discrepancy cannot slip through because the AI happened to be unavailable.

  • Tender pack analysis

    Engineering

    Breaks a tender pack into priceable items and notes the requirements and the risks — enough to decide whether to bid.

    Model
  • Pricing from history

    Engineering

    Price by median over past quotes and the catalogue, with the number of comparable lines and the quotes cited. Reproducible and defensible.

    Computed
  • Benchmark order of magnitude

    Engineering

    On items with no precedent only, labelled as such, never added into the priced total and always counted as “to be priced”.

    Model
  • Bid ranking

    Purchasing › Tenders

    Compares replies on price and lead time, but also year-end rebates, penalty and indexation clauses, payment terms and the history of disputes.

    Model
  • Supplier risk profile

    Master data › Suppliers

    The internal trading relationship crossed with external credit standing — insolvency proceedings under way are flagged explicitly, and the summary ends with continue, monitor or heightened watch.

    Model
  • Order versus supplier reply

    Purchasing › Orders

    Any price difference above 5% is raised by computation, independently of the model, which only writes the readable summary.

    Computed
  • Three-way match: invoice, order, receipt

    Purchasing › Invoices

    The lines that do not tally surface with the amount of the difference.

    Computed
  • Receipt reconciliation

    Accounting

    Statement transactions are matched to open customer invoices and stay “to be checked” until confirmed.

    Computed
  • Price and lead-time drift

    Purchasing › Suppliers

    The trend over several months, item by item — a price creeping up 12% with nobody having decided anything is invisible on a single order.

    Computed

Augmented control tower

The indicators from every area on one screen. The choice fits in a sentence: a KPI is not a figure, it is a figure and its cause — an indicator in the red shows what is dragging it down, with a link to the screen concerned.

  • Consolidated indicators

    Management › Control tower

    Each area reports its KPIs into a section, on a single page.

    Computed
  • Attached critical points

    Management › Control tower

    Causes are shown only if the indicator really is degraded — otherwise the page becomes a wall of alerts nobody reads any more.

    Computed
  • Financial trend

    Management

    Revenue, purchases, margin, receivables, arrears, payables and working capital requirement, month by month.

    Computed
  • Supplier service level

    Purchasing › Suppliers

    An order counts as met if the receipt falls within two days of the confirmed date — a logistics tolerance, not punctuality to the hour.

    Computed
  • Figures shared on the portals

    Supplier portal, customer portal

    The supplier sees how they are being judged before their contract comes up; the customer sees what was delivered to them. An assessment done to you becomes something to discuss.

    Computed

Automatic flows, controlled and assisted

Eleven routines run on their own, daily or weekly, without anyone starting them. It is the least spectacular part of IA.TERRA and probably the one that pays best: a forgotten chase costs more than a badly turned report. Triggering is a matter of the calendar, not the model — a deadline cannot depend on an AI being available.

  • Mailbox pickup

    Purchasing › Invoices

    Invoices received as attachments are captured, read and created with the status “to be checked”. Nobody downloads or renames anything.

    Model
  • Chasing orders with no reply

    Purchasing › Orders

    24 h, 48 h, 72 h after sending. Beyond that the order moves to dispute and the team is told — a silent supplier gets discovered before the delivery goes missing, not after.

    Computed
  • Chasing tenders

    Purchasing › Tenders

    72 h then 24 h before the deadline, to every invited supplier who has not yet replied.

    Computed
  • Customer collections

    Sales › Invoices

    First chase 48 h after the due date, then every 72 h until the invoice is settled — working days only; a Sunday chaser does not get you paid faster.

    Computed
  • Spotting a job that needs scheduling

    Operations › Reports

    An “we should also sort out the UPS in rack 2” slipped into a report comes back out as a job to schedule, separate from the one being written up.

    Model
  • Reordering at the threshold

    Purchasing › Stock

    Any movement that takes a bin or a vehicle below its reorder point creates the purchase request, and a daily pass acts as a safety net.

    Computed
  • Certification expiry

    HR

    Alerts at 90, 60 and 30 days. An expired certification means a technician who cannot go out that same morning.

    Computed
  • Fleet renewals

    Assets › Vehicles

    Roadworthiness test and insurance, alerts at 30 and 5 days.

    Computed
  • Customer satisfaction

    Operations › Satisfaction

    A weekly digest to the operations manager: average score, response rate and the detail of critical reviews.

    Computed

Reports

The report nobody has time to write, on figures the system already holds. One per area, plus a board-level summary.

  • One report per area

    Every area › Reports

    Management, engineering, operations, purchasing, HR, sales, assets, accounting, requests — each on its own quantified scope.

    Model
  • Profitability and weak signals

    Management › Reports

    Two cross-cutting analyses, alongside the board summary.

    Model
  • Structured action plan

    Every report

    The follow-ups proposed are split into PDCA and by time horizon, with a link to the record cited.

    Model
  • Exports in your house style

    Every report

    Word, PowerPoint and Excel ready to circulate at a board meeting, with no reformatting.

    Computed
  • Instructions specific to your trade

    Integrations › Custom prompts

    Eighteen customisable scopes: the twelve reports, the three levels of job report, the style of campaigns, tender-pack analysis and the technical submission. What you can do in-house, what you subcontract, the items you never price, your certifications, the wording your main clients impose — a model has no way of guessing any of it.

    Model
  • Walled-off archiving

    Every area › Reports

    A report stays subject to the permissions of its area — reading, export and deletion. The HR report only opens with HR access.

    Computed

Writing and lighter correspondence

What gets said ends up written, and what is written goes to the right person — without an evening spent writing it up neatly first.

  • Job report dictated

    Jobs

    Speech is transcribed with the speakers separated, then laid out: decisions, actions with an owner and a due date, key points.

    Voice
  • Three levels of write-up

    Jobs › Reports

    Compact for a callout, detailed for the customer, decision-focused for management — risks, costs and priorities.

    Model
  • Leading into the next step

    Jobs › Reports

    The report proposes what comes next — quote, invoice or another job — and opens the matching screen.

    Model
  • Dictating a field

    Across the app

    A few seconds of speech fill in a form field, with the result shown live.

    Voice
  • Technical submission

    Engineering

    Written from the real references pulled out of the database, never from an invented job site.

    Model
  • Non-conformity report

    Purchasing › Disputes

    Written from the quantities ordered and received, the invoicing discrepancies and the photos taken on site — the document goes to the supplier instead of staying a spoken conversation.

    Model
  • Today's agenda

    My space

    Jobs, absence, approvals pending and certifications about to expire, in a few lines in the morning.

    Model
  • Sales campaigns

    Sales › Campaigns

    Proposes campaigns, writes the messages, tailors them to each recipient — without ever choosing the target, which comes from verifiable queries.

    Model
  • Requests for quotation

    Purchasing › Tenders

    The invitation goes to the supplier's own space, where the request joins their orders, their invoices and their non-conformities — one place instead of an email thread.

    Computed

What IA.TERRA holds itself to

Five guarantees, not a promise

Automated assistance is only worth anything if you can check where its figures come from. These five rules apply to all five roles.

Anchored on your data, never on generalities

Pricing finds the figures in your past quotes and states, line by line, what it rests on: how many comparable lines, which quote numbers, including the ones that were accepted.

Set on actual hours, not budgeted hours

Labour estimates take the hours actually clocked on your completed job sites — the gap between budget and reality is precisely what derails jobs.

Anything with no precedent is marked “to be priced”

No plausible figure is invented to fill a gap, and what is priced is never added to an order of magnitude in the same total.

The machine proposes, a person signs off

Everything it produces arrives as an editable draft, with the origin of every figure. Nothing goes out unread.

Your data is not used to train any model

That is a contractual commitment, written into article 14 of the terms and conditions — not an intention.

What it costs

Your plan opens all five roles, with no licence supplement. Only consumption is charged, against prepaid credit kept separate from the subscription: you pay only for what actually ran. An empty balance puts the AI functions on hold and touches nothing else — quotes, job sites, purchasing and invoicing carry on.

The trial starts with €5 of free credit — enough to run all five roles on your own documents before deciding.

The best test is your own documents.

A supplier invoice, a tender specification, a bank statement: those are what tell you whether IA.TERRA holds up, not a prepared demo.