§ 00Applied AI North

Practical AI for document review built, taught, handed off.

Applied AI North is a build-and-enable practice. We design working systems, train the team that will run them, and leave the keys with you. No black boxes. No vendor lock-in. No six-month discovery phases.

Start a project Watch a system run
We bridge
AI hype real usage
We bridge
Tools fluency
We bridge
Pilots production
We bridge
Vendor-built team-owned
Practical AIClaude SonnetClaude OpusGPT-5Gemini 2.5 ProChatGPTCopilotLlama 3.3 Context EngineeringProblem DecompositionPrompt EngineeringPrompt-as-SOP DesignRAG Agentic AIAgent-2-Agent (A2A)Workflow OptimizationVerification LoopsOutput repair Model & tool selectionHuman judgementAI LiteracyAI UpskillingSensitive DataGTM Engineering n8n · Make · ZapierLangGraphVercel AI SDKpgvector · SupabaseSalesforce · HubSpot PIPEDA-awareSOC2-awareBuilt in Toronto Practical AIClaude SonnetClaude OpusGPT-5Gemini 2.5 ProChatGPTCopilotLlama 3.3 Context EngineeringProblem DecompositionPrompt EngineeringPrompt-as-SOP DesignRAG Agentic AIAgent-2-Agent (A2A)Workflow OptimizationVerification LoopsOutput repair Model & tool selectionHuman judgementAI LiteracyAI UpskillingSensitive DataGTM Engineering n8n · Make · ZapierLangGraphVercel AI SDKpgvector · SupabaseSalesforce · HubSpot PIPEDA-awareSOC2-awareBuilt in Toronto
§ 01Watch a system run

We build the kind of AI that earns its rent.

An illustrative sketch of one shape we ship: a triage agent pulling work off a queue, classifying it, scoring its own confidence, and routing the edge cases to a human. Loops every minute or so. Code lives in your repo, not ours. Your team can read it, change it, and run it without us.

Inbound review queue · 0/8 agent: intake-triage-v3 · running
  • 01intake_2841.pdfQueued
  • 02intake_2842.pdfQueued
  • 03intake_2843.pdfQueued
  • 04intake_2844.pdfQueued
  • 05intake_2845.pdfQueued
  • 06intake_2846.pdfQueued
  • 07intake_2847.pdfQueued
  • 08intake_2848.pdfQueued
~$ agent run intake-triage-v3 [ok] connected: supabase, pgvector, claude-sonnet-4.5 [ok] eval set loaded · 1,200 historical examples [ok] confidence floor: 0.86 · human-review at 1.00   [warn] field edge_case_07 always routes to human [warn] field ambiguous_12 always routes to human   [ok] ready · awaiting inbound documents
Avg time per doc
87s
Field accuracy
94%

An illustrative system, the same shape we ship. We hand over the repo, the eval set, the runbook, and the named owner on your side. How a project runs


§ 02What we do

Seven services. One outcome: AI that gets used.

We don’t sell capabilities. We sell better ways of working. Every engagement comes with a scope, a price, a calendar, and a named outcome on your team’s dashboard. No discovery phases that bill for ten weeks and end in a PDF.

All services and pricing


§ 03Shapes we ship

Three system patterns, demystified.

We don’t publish client case studies. Most of the work is internal, sensitive, or under NDA. Instead, here are the shapes of systems we actually build — abstracted into illustrative mockups so you can see how they hang together. If one of these resembles a problem on your desk, that’s the conversation to start.

structured-extraction · illustrative Queue Review Audit
intake_2841 · standard · 12 pages extracted 87s · 38/38 fields
SubmitterField 010.99
Amount$2,840.000.98
Term12 months0.99
Effective2026-08-010.99
Variationindexed +1.5%0.94
Edge clauseSee section 14(c)human
Add-ons1 unit · included0.96
Hold-back$2,840 deferredhuman
SHAPE 01   Structured extraction  ·  4–6 weeks

Pull structured data out of long, messy documents.

The workhorse pattern. Replace a multi-hour manual review with a 90-second extraction agent that scores its own confidence and routes the edge cases to a human. Evaluated against your real historical documents before it touches production.

Typical accuracy
93–96%
Time / doc
~90s
Human review
5–12%
Claude Sonnet 4.5 pgvector Eval harness Human-in-the-loop
classification-triage · illustrative
#InboundClassConf
01Request item 1041type-A0.97
02Request item 1042type-B0.99
03Request item 1043type-C0.81
04Request item 1044, flaggedurgenthuman
SHAPE 02   Classification & routing  ·  2–4 wks

Intake triage with confidence thresholds.

Classifies inbound calls, forms, and emails into one of ten to twenty buckets with a confidence floor and a human-in-the-loop on the edge cases. The kind of thing a junior used to do all day.

internal-knowledge · illustrative
RefQuestionSourceStatus
#0842Refund window on tier-3policy-v4drafted
#0843SLA on emergency ticketssla-handbookdrafted
#0844Carve-out for legacy plansmemo-2024in review
#0845Onboarding step-skip rulerunbook-v2in review
SHAPE 03   Internal Q&A  ·  3–6 wks

A chatbot that knows your docs and cites its sources.

An agent that answers internal questions from your real policies, handbooks, and tickets — with citations, a confidence floor, and a clean handoff to a human when the question is outside the corpus.


§ 04How a project runs

Six phases. Six weeks. No surprises.

The shape of a typical Chatbots & Agents engagement. Assessments and integrations follow the same rhythm at different lengths. We publish the calendar before kickoff, so there is no mystery about what happens when.

Full approach

Week 1
Assess

On site with the team. Watch the work happen. Decide what is worth building.

Week 2
Shape

Write the evals before the agent. Ground truth from your real historical data.

Week 3
Build

First working pass. Crude but end-to-end. Shows where it fails, which is the useful part.

Week 4
Harden

Tool calls, retries, cost ceilings, observability. Where most pilots quietly stop.

Week 5
Pilot

Live with two or three people. Watch what happens when a real human gets bored.

Week 6
Adopt

Runbook, evals you can rerun, named owner on your side, 30 days of support included.

Kickoff Calendar shown before contract Handoff & 30-day support
§ 05How we work

A partner, not a vendor.

Most AI agencies sell capabilities. We sell better ways of working — and then we hand them to the people doing the work.

Every engagement comes with two deliverables: the system, and the team that can run it without us. We document the prompts, the eval set, the failure modes, and the cost ceilings. We train the named owner on your side until they can change the thing themselves. The off-ramp is built in from week one.

Engagement
Fixed scope and price, named outcome, calendar agreed before kickoff
Ownership
Code in your repo, prompts in your docs, evals you can re-run, no proprietary glue
Enablement
Every build comes with a workshop. Every workshop ends with one of your people running the system live.
What we say no to
Discovery phases that bill for ten weeks and end in a PDF. Anything we’d need to subcontract to deliver well.

§ 06Build & enable

Two tracks. Run them together.

A system without a fluent team gets shelved. A fluent team without a system stays stuck on copy-pasting into ChatGPT. We run both tracks in parallel so the day we leave looks the same as the day after — for the people doing the work.

TRACK A · BUILD

Systems that earn their rent.

Workflow automation, agents, internal Q&A, structured-extraction pipelines, AI-aware product surfaces. Scoped in weeks, not quarters. Evaluated against your real historical data. Shipped to production with monitoring you can read.

  • Evals before agents — we write the test set first.
  • Confidence floors, retries, cost ceilings, and a human-in-the-loop by default.
  • Code in your repo. Prompts in your docs. No proprietary glue.
TRACK B · ENABLE

Fluency, not just access.

Hands-on AI literacy for the people who’ll touch the system every day. Ops, sales, legal, support, ops-adjacent leadership. Workshops are built around your tools, your docs, and the actual workflow we just shipped — not generic prompt theatre.

  • Half-day to full-day workshops. Plain English. No demos that won’t reproduce on Monday.
  • Quick-start guides written for your roles, not the vendor’s docs.
  • A named internal owner who leaves the engagement able to ship the next iteration alone.

§ 07What we keep seeing

Three quiet failure modes. We name them on purpose.

01
Access without fluency

Everyone has a license. Almost no one is fluent.

A Copilot or ChatGPT seat on every desk and a small bump in productivity to show for it. The fix is rarely a better tool. It’s teaching the team how to use the one they already have, on the work they actually do.
02
Pilots without production

A graveyard of demos. Nothing in the workflow.

A great-looking proof of concept on a sales-call screenshare that never made it past one Slack channel. The hardening, observability, and adoption work — the part where it earns its rent — was never scoped.
03
Systems without owners

A workflow that breaks when one person leaves.

An automation living in a private workspace, a prompt in a Notion doc only one person can find, a vendor-built agent no one inside the company can change. We won’t leave one behind.
§ 08Start a project

Tell us the thing that’s taking too long.

A thirty-minute call. We read your brief before we meet. If it is not the right fit, we will say so and point you toward someone better. If it is, you will have a scope and a price by the next morning.

Book an intro hello@appliedainorth.com