True North Solutions
What happens inside each tool
One diagram per tool. The colour of each step tells you who is doing the work — fixed rules, an AI model, or a person. That distinction is the whole method.
Input or output
Deterministic — fixed rules, same answer every time
AI agent — judgment under ambiguity
Human decision — kept deliberately
01
Judgment Lane Classifier
Answers "should this be automated, and by what?" before anyone writes code. Note that no AI runs inside this tool at all — the decision about where AI belongs is itself too important to leave to a model.
You describeA task, scored on four sliders: how much judgment it needs, what happens when it goes wrong, how often it runs, how many exceptions appear.
Rule: volume × repeatabilityHigh volume with low judgment is the classic automation case.
Rule: consequence of errorHigh risk pulls the answer toward a human, whatever the volume.
Rule: exception rateMany exceptions means the rules are not really known yet.
Lane: deterministicRules are known and stable. Build software, not a model.
Lane: AI agentLanguage or ambiguity is genuinely present, and errors are recoverable.
Lane: humanJudgment, accountability or relationship. Automating it destroys value.
You getA lane, the reasoning behind it, and what would have to change for a different answer.
Why it matters in the room: most AI disappointment comes from putting work in the wrong lane. This makes that argument visible in about a minute.
02
Automation ROI Calculator
Payback on a specific automation, using published wage data rather than invented numbers.
You chooseRegion, role, how many people, hours lost per week, and the build cost.
Published wage dataUS: BLS OEWS May 2025. Singapore: MOM June 2024. India: ASI aggregate.
Employer load appliedUS ×1.43 (BLS ECEC), Singapore ×1.17 (CPF), India ×1.16.
Currency and paybackConverted, then annual saving against build and run cost.
You getMonths to payback, annual saving, and the assumptions stated openly so they can be argued with.
The uncomfortable finding, left in on purpose: in India the labour case often does not close — fully-loaded clerical cost is low enough that a build cannot be justified on wages alone. The tool says so. There, the case rests on quality, throughput, traceability and risk instead. India has no official occupation-level wage data, so the ASI aggregate is used and flagged rather than substituting self-reported figures that vary threefold for the same role.
03
AI Opportunity Audit
A Lean value-stream analysis of a process you describe, ending in a prioritised view of where AI helps and where it does not.
You describeIndustry, function, and how the process actually runs today — in your own words.
Map the value streamBreaks the process into 4–8 steps as they run today, not as the SOP claims.
Name the wasteEach step tagged with one of the eight Lean wastes, stated concretely.
Assign a judgment laneEvery step routed to deterministic, AI, or human — with the reasoning shown.
Selectivity ruleThe system is instructed that routing most steps to AI means it has done the job badly.
You getValue stream, lanes, 3–6 prioritised opportunities, honest risks, one first move, and how success would be measured before trusting it.
Then a human reads itOutput is a structured starting point for a conversation, not a recommendation to act on.
Built-in honesty: a live run on a forging process returned five deterministic steps, one AI, one human. A tool that sells AI would not have said that.
04
Agent Audit Trail
The answer to "how would we ever trust it?" — an agent doing real verification work with every step recorded.
The scenarioForging batch BF-24817, front axle beam, 42CrMo4. Five supplier certificates in five different formats. Spec: tensile 900–1100 MPa, yield min 700, hardness 269–331 HB.
ReadEach certificate parsed, whatever its layout.
NormaliseUnits converted — kg/mm² to MPa, comma decimals handled — with arithmetic shown.
MatchHeat numbers checked against the batch's required heats.
VerifyEvery value tested against the specification.
Every step loggedCertificate, action, detail and outcome — ok, warning or blocked. Nothing happens off the record.
Release stays humanThe agent is forbidden to auto-release, and forbidden to guess an illegible value. It recommends and flags; a person decides.
You getThe full trace, what it found, what it could not read, and what a human must rule on.
What it caught in a live run: a unit-conversion trap, a heat number belonging to a different batch, an illegible scan it refused to guess at, an out-of-specification value — and one defect nobody planted: two certificates from different suppliers both claiming the same heat number. It held the batch.