Human AI Judgment
Work with me

Job vs AI · Engineering

Mechanical engineer

Design, analysis and validation work where a wrong answer ends up in a physical object. The artefacts are getting cheaper to produce; deciding whether they are right is not.

  • design
  • simulation
  • manufacturing
  • verification
Get the playbook

A drawing can be generated. The decision to release it carries a name, and that has not moved.

The problem

Start with the problem, not the tool

You already know mechanical engineering. The open question is where technology belongs in it.

Most engineers have tried a chatbot on a work task by now. It drafted something usable, or it was confidently and subtly wrong, and either way the experience did not settle where the technology belongs.

"What can AI do for engineers?" leads nowhere. It starts from a capability and hunts for a use, and it ends in a pilot that demos well and changes nothing. The question that works is narrower: at this specific step, what am I trying to accomplish, which part needs judgment, and what has to be checked?

Order matters. Problem, then workflow, then decision, then technology, then verification. Choosing the tool first is how organisations buy capability they never use.

Lane 01

Automate

Rule-shaped work that is cheap to check.

  • What belongs here

    Repeated information processing against a stable rule: extracting known fields from controlled documents, running an agreed calculation across many cases, assembling a report from information that already exists.

  • What it still costs you

    Defining the fields and their allowed formats, approving the source set, and sampling the output against it. Automated output is never engineering approval.

Lane 02

Augment

A person still decides; the machine widens or accelerates the input. Most of this analysis lands here.

  • What belongs here

    Analysis, research and design exploration: summarising a long requirement set, expanding the concept space, preparing a study matrix, comparing many result sets, organising the evidence in a failure investigation.

  • What stays yours

    The physics. Which source is authoritative, what a simulation is actually capable of proving, what a pattern could mean physically, and the root-cause conclusion.

  • Why so much lands here

    Not because engineering is unusually augmentable, but because the value of the work is in assumptions, trade-offs and verification - and none of those transfer with the output.

Lane 03

Human

Judgment, accountability, and trade-offs someone has to own.

  • What belongs here

    Framing the problem in the first place, choosing between several defensible answers whose consequences differ, and decisions that depend on live context and coordinated action.

  • What AI can still do around it

    Organise stakeholder statements, assumptions and open questions. Build the comparison table. Prepare the briefing. None of that is the decision.

Lane 04

Verify

The check that makes the other three safe - and an obligation inside each of them, not a fourth kind of work.

  • What belongs here

    Determining which standards and requirements apply and interpreting their engineering meaning, and the review that decides whether calculation, design and test evidence justify release.

  • Proportional, and often imposed

    Match the weight of the check to the cost of being wrong; over-checking is how a good workflow gets abandoned. In safety-critical work the check is frequently mandated rather than designed, and the first move is to find out what is already required.

  • What counts as evidence

    An AI output is evidence only when you can trace where it came from and how it was checked. Confidence, fluency and formatting are not engineering evidence.

Situations

Six situations, worked

Six representative situations, chosen to show different lane mixtures. The counts are of these six examples - not a measurement of the profession.

  • 01Automate1example
  • 02Augment3examples
  • 03Human1example
  • 04Verify1example

  • Pulling measurements out of controlled test reports into a structured table

    Automate
    per test campaignStable rule, output checkable against a document that already exists, and a wrong value is caught before it reaches a calculation.

  • Preliminary sizing and first-pass calculations

    Augment
    per design cycleThe arithmetic can be run for you. Choosing the method, the assumptions and the validity range cannot.

  • Exploring more concepts than there is time to draw

    Augment
    early in a projectThe value is breadth, not correctness. A generated concept is a hypothesis and re-enters the normal review path.

  • Investigating a failure

    Augment
    as requiredUseful for assembling evidence and finding patterns across data too large to eyeball. The root-cause statement stays with the engineer who can explain the mechanism.

  • Deciding what the requirement really is

    Human
    per projectThe stated requirement is often not the real need, and resolving that is a conversation between people with competing interests.

  • Releasing the design

    Verify
    per releaseAssembling and cross-checking the evidence is automatable. Deciding the evidence is sufficient is not, and the check is usually mandated rather than designed.

Solutions

Choose the kind of solution, then the tool

Most workflow problems are not generative-AI problems. Naming the class is the decision; the products are examples of it.

  • Conventional software and analytics

    The rule is known and exact, or you need to describe what already happened.

    For example: MATLAB and Simulink, Python or an approved scripting environment

    Watch for: Asking a model to reproduce a rule you could write is the tell that you picked wrong.

  • Optimisation and generative design

    You need the best configuration under explicit geometry, performance and manufacturing constraints.

    For example: Autodesk Fusion Generative Design, Altair DesignAI

    Watch for: Expanding the option space is not selecting the winner.

  • Machine learning on your own simulation data

    Screen many variants quickly, once a credible simulation dataset exists to train on.

    For example: Ansys SimAI, Altair PhysicsAI

    Watch for: A surrogate is valid only inside its training domain. Keep conventional simulation and test for validation.

  • Generative AI

    Language or a document is the deliverable - drafting, summarising, extracting - including inside CAD.

    For example: SOLIDWORKS Design AI, Siemens Design Copilot NX

    Watch for: Design intent, geometry and approval stay with the engineer.

  • No technology at all

    The approved solver, engineering review, physical test - or a process change, or deleting the step.

    Watch for: Automating a step that should not exist is worse than leaving it alone.

Chatbot → agent

From chatbot to agent

The same task at four levels of support. Most engineering work should stop at the second.

  1. Ask a chatbot

    You paste the report in and it summarises. Genuinely useful, and where most people should start.

    Augment

  2. Save the prompt

    The same instructions every time, so the output shape stops drifting. This is where most engineering tasks should stop.

    Augment

  3. Give it the tools

    It reads the requirement register and the report itself, rather than whatever you remembered to paste.

    Augment

  4. Make it an agent

    Multi-step, branching on what it finds, with a fixed output contract and an explicit escalation. Worth it only when the task repeats, needs live access, and can be checked mechanically.

    Verify

Blueprint

What an engineering agent looks like

One worked example, through the same ten components any profession would use. The playbook carries three.

  • Goal

    Produce a requirement-by-requirement review of one test report. Narrow enough that you can tell whether it succeeded.

  • Rules

    Never infer a missing measured value. Distinguish measured from calculated from inferred. Cite the location of every number.

  • Memory

    Component naming conventions, project terminology, validated material properties and approved requirement IDs - the context you would otherwise re-supply on every run.

  • Output contract

    One row per requirement: ID, criterion, measured value, source location, verdict, confidence. A fixed shape is what makes review mechanical instead of a re-read.

  • Verification

    Every value opens to the page it came from, so checking one takes seconds. If the design does not say how a human confirms the output, the work has moved rather than reduced.

  • Escalation

    Missing evidence, contradictory sources, and every safety or design acceptance decision. An agent with no defined refusal will answer confidently in exactly the case where it should have stopped.

Risks

Where this goes wrong

The familiar failure modes - wrong input, wrong assumption, wrong model - plus new ways to hide the mistake.

  • Confident fabrication

    High risk

    A system can produce plausible explanations, equations, citations or specifications that are simply wrong.

    What to do:Treat every critical value and citation as unchecked until you have opened the source yourself.

  • Automation bias

    High risk

    A polished output reduces the reviewer's willingness to challenge it.

    What to do:Match the strength of review to the consequence of being wrong, not to how convincing the output looks.

  • Data leakage

    High risk

    Designs, drawings, customer information and proprietary knowledge may be inappropriate for an external service.

    What to do:Ask whether you would send the file to an external supplier with no NDA. If not, check the approved environment first.

  • Model mismatch

    Medium risk

    An AI surrogate or simplified method may be operating outside its training or validity domain.

    What to do:Keep conventional simulation and test for validation, and check that new designs stay inside a credible domain.

Sources

Research basis

What the profession framing and the risk principles rest on. The lane assignments above are Human AI Judgment analysis, not a classification made by any of these organisations.

+Playbook

The full analysis, in a form you can take into work.

The Practical AI Judgment Playbook: Mechanical Engineering

Six representative situations, the reasoning behind each lane, three agent blueprints, and the failure modes worth knowing before you trust an output.

Cover of The Practical AI Judgment Playbook: Mechanical Engineering

In your own words. This decides which profession I analyse next.

Optional.

Optional. You will receive the requested playbook whether or not you opt in — this box only controls future emails.

Keep going

Other professions

The same structure, applied to different work.

  • Sales

    Account Manager

    The only profession in this series where a completely correct output can still be the thing that costs you the account.

  • Delivery

    Project Manager

    The profession whose actual deliverable is information other people act on - which makes it the one where faster reporting can quietly make the job worse.