● Process automation · document AI · agents · governance
AI and automation measured in hours reclaimed, not demos
We find the work your team does by hand every week, automate the parts that are rules and reading, and leave the judgement with people. Baselined before, reported monthly after, governed to a standard your auditor can read — with a two-week assessment that ranks candidates by payback before anyone builds anything.
Candidate scoring · assessment
payback first
Invoice & PO intake1,400 documents / month
~310 h / yr
Ticket triage & routingSupport + intake inbox
~180 h / yr
Quote draftingFirst pass from spec sheet
~145 h / yr
Record reconciliationERP ↔ bank ↔ shipping
~90 h / yr
Contract clause reviewDeferred · data not ready
hold
Ranked by payback, not novelty
Illustrative scorecard
31 automations in production · 2023–2026
ClaudeAzure OpenAIOdoon8nPostgresAWS Bedrock
EN · FR · AR
What we automate
Three patterns cover most of the real work.
Almost every engagement is one of these, or a combination. Each ships with a confidence threshold, a review queue for anything below it, and an audit trail — so the automation can be wrong without being expensive.
Email / scan arrives
Extract → validate
Low confidence → human
Posted to ERP
A.01 · Document AI
Reading documents
Invoices, purchase orders, delivery notes, claim forms, statements. Extracted, validated against your own records, and posted — with anything ambiguous routed to a person instead of guessed.
- Field-level confidence
- Validated against ERP
- Review queue & audit log
- EN · FR · AR documents
Live in 4–8 weeks
From $18k
Trigger in system A
Rules + classification
Write to system B
Exception → owner
A.02 · Workflow
Moving work between systems
The handoffs that currently happen by email and copy-paste: routing, approvals, status sync, reconciliation, alerts. Built at the API layer so a UI change doesn’t break it.
- Odoo · CRM · accounting
- Idempotent & retryable
- Exception ownership
- No screen scraping
Live in 3–6 weeks
From $18k
Request in plain language
Retrieve your data
Proposed action → approve
Executed & logged
A.03 · Agents
Assistants with real tool access
An internal assistant that can answer from your own documents and take scoped actions — draft the reply, create the record, prepare the report — with approval required before anything is committed.
- Retrieval over your content
- Scoped, permissioned tools
- Approval before commit
- Every action traceable
Live in 8–16 weeks
From $85k
Guardrails
Six controls that make this defensible in an audit.
We came to AI work from ISO 27001 and GRC engagements, so governance isn’t a slide at the end. These controls are designed into the build and documented for your privacy register and your auditor.
Six controls · designed in
G.01
No training on your data
Enterprise endpoints with training disabled and zero-retention terms, or self-hosted open models where residency demands it. Contracted, not assumed.
Terms
Zero retention
G.02
Data residency you choose
Deployed into your cloud tenant in the region you name — Canada, US or EU. Prompts, outputs and documents never transit a Noordev system.
Region
CA · US · EU
G.03
Human in the loop by design
A confidence threshold per field or decision, a review queue below it, and a named owner for exceptions. Nothing with financial or legal effect commits unapproved.
Approval
Required at gates
G.04
Full audit trail
Input, output, model and version, confidence, who approved and when — retained in your infrastructure and queryable. Every automated action can be traced and reversed.
Retention
Your policy
G.05
Evaluated before it ships
A labelled test set built from your real documents, with accuracy and cost thresholds that must be met to go live — and re-run whenever a model or prompt changes.
Gate
Eval suite in CI
G.06
Documented for Law 25 & GDPR
Data flow diagram, purpose, lawful basis, retention and subprocessor list handed over as records you can drop into your privacy register.
Deliverable
Register-ready
How an engagement runs
Assessment first. Nothing built before the payback is on paper.
Most AI projects fail at the first step — automating something that was never worth automating. Two weeks of measurement costs less than one wrong build, and the fee is credited against whatever you decide to do next.
01
Week 1–2
Assess
Process inventory, volume and hours baseline, data readiness check, and a ranked backlog with estimated payback per candidate.
- Hours baseline
- Ranked backlog
- Fixed quote
02
Week 3–4
Prove
The top candidate run against a labelled sample of your real data. Accuracy, cost per item and failure modes measured before any integration work.
- Labelled test set
- Accuracy + cost
- Go / no-go
03
Week 5–9
Build
Integration, review queue, thresholds, logging and dashboards — in your tenant, in your repository, deployed through CI with the eval suite as a gate.
- Your tenant & repo
- Review queue UI
- Eval gate in CI
04
Week 10–12
Run parallel
Automation runs alongside the manual process while your team checks its output. Thresholds tuned on live data before anything is switched off.
- Shadow mode
- Threshold tuning
- Team training
05
Month 4–∞
Measure
Monthly report against the original baseline: hours reclaimed, accuracy, exceptions, run cost. Retire or rebuild anything that misses its case.
- Hours reclaimed
- Run cost at cost
- Next candidate
Engagement models
Fixed-fee builds. Model costs billed at cost.
We don’t resell tokens or take a margin on inference. Your provider bills you directly, we show the run cost per item in the monthly report, and we’re paid for the engineering.
Start here
Automation assessment
A measured, ranked answer to what’s worth automating. Credited in full against your first build, and useful even if you build nothing.
- Process & volume inventory
- Hours and error baseline
- Data readiness assessment
- Ranked backlog with payback
- Governance & residency plan
- Fixed quote for phase two
Most common
Automation sprint
One process taken from manual to production, including the review queue, the logging and the parallel-run period. Where most of our AI work lands.
- One process, end to end
- Document extraction or workflow
- Review queue & thresholds
- Eval suite in CI
- Parallel run & team training
- Audit trail & privacy records
- 90 days of tuning included
Platform
Agent or platform build
Multiple workflows, retrieval over your own content, scoped tool access and an approval layer — plus the governance to put it in front of regulated users.
- Retrieval over your content
- Permissioned tool access
- Approval & escalation layer
- Observability dashboards
- Model-agnostic architecture
- Security review & threat model
- Runbooks & handover
In production
Medians across 31 automations, 90 days after go-live.
Measured against the baseline recorded during the assessment. Two automations missed their projected payback and were retired — those are in the median too.
Median manual hours removed
Field-level extraction accuracy
Median payback period
Actions with an audit record
Client · Argenteau Distribution
Approval above threshold
SB
Simon Bélanger · VP Finance, Argenteau · Laval
AI & automation FAQ
Before you fund a pilot.
We’ll tell you when a spreadsheet formula, a report, or a process change would beat an AI build — that answer comes up more often than the market implies.
What does AI automation actually replace?
+
High-volume, rule-heavy work that currently moves through inboxes and spreadsheets: reading invoices and purchase orders, triaging and routing tickets, extracting data from PDFs and contracts, reconciling records between systems, drafting first-pass replies and reports. We automate the steps, not the judgement — a human still approves anything with money or legal consequence attached.
How do you decide what to automate first?
+
A two-week assessment scores every candidate process on volume, hours consumed, error rate, data readiness and how contained the failure mode is. You get a ranked backlog with an estimated payback per item. We start with the highest-payback process that can fail safely, not the one that demos best.
Does our data get used to train models?
+
No. We deploy against enterprise API endpoints with training disabled and zero-retention terms, or self-hosted open models where residency requires it. Data stays in your cloud tenant, prompts and outputs are logged in your infrastructure, and the data flow is documented for your privacy register — including Québec Law 25 and GDPR records.
What happens when the AI is wrong?
+
Every automation ships with a confidence threshold, a human review queue below it, and a full audit trail of input, output, model version and decision. Failure is designed to be visible and reversible: low-confidence items route to a person rather than being pushed through, and every action an agent takes can be traced and undone.
How much does an automation engagement cost?
+
A two-week assessment is CAD $9,500 and credited against the first build. An automation sprint that puts one process into production is $18k to $45k depending on integration depth. Agent and platform builds with multiple workflows, tool access and an approval layer run $85k to $180k. All fixed-fee, with running model costs quoted separately and billed at cost.
Do you build on our existing systems?
+
Yes — that’s usually where the value is. Odoo, Dynamics, Salesforce, HubSpot, Sage, Shopify, SharePoint and Postgres are the systems we most often connect. Where an API exists we use it; where one doesn’t, we build a documented integration rather than screen-scraping a UI that will change.
Is this RPA?
+
No. Classic RPA drives a user interface with a script and breaks the next time a button moves. We work at the data and API layer, using language models only for the parts that genuinely need judgement — reading unstructured documents, classifying intent, drafting language. The result is cheaper to run and far less brittle.
How do you measure whether it worked?
+
We baseline the process before we touch it — volume, hours, cycle time, error rate — and report the same numbers monthly after go-live. If an automation doesn’t reach the payback we projected, we say so in the report and either fix it or recommend retiring it. Hours reclaimed is the metric, not tokens consumed.
Related services & reading
Related services
Automation types
● Start with the numbers
Bring one process. We’ll bring the baseline.
A 45-minute call on the work that eats your team’s week. If a two-week assessment is the right next step we’ll say so — and if the honest answer is a process change rather than a model, we’ll say that instead.
© Noordev Technologies Inc. · 2012–2026
Montréal · Toronto · Miami · Rabat
