The AI-native engineer: a starter kit

One

The story, in one line.

The electric motor was ready by 1900. The boom came in the 1920s, when factories were rebuilt around it instead of fitting it into the old line-shaft floor. A faster tool rises to the ceiling set by the workflow around it. Our shafts are digital: the handovers, the re-typing, the approvals, the judgement nobody wrote down. And nobody in the industry is 10 years ahead of you.

A factory floor around 1900, with a line shaft and belts running overhead — A factory floor around 1900. The motor has arrived; the line shaft still turns overhead and the belts still drop to the same machines. Library of Congress.

Both talks carry that argument, with the notes behind every slide.

ICE London · 17 September 2026

AI and the civil engineer: reimagining roles in infrastructure deliveryWelcome and context setting by Bill Guo, ICE London Vice Chair, before Dr Neda Naghshbandi, Anthony Dewar, Lucy Gardner and Amira Damji. The motor story, the capability and adoption curves, the three lanes of work, and the 55-station survey where the key turned.
Download the slidesPDF · 15 pages · 2.5 MB

RCEA × Young Rail Professionals · 24 September 2026

AI, Rail and Engineering: the why, what and howBill Guo's 20-minute talk at The use of AI in rail, alongside Mujadded Alif, Jonathan Ryan and Jiaxi Li. Why it matters to an early career, what actually changes in the job, and how to start with one workflow.
Download the slidesPDF · 13 pages · 2.2 MB

Two

One workflow, two questions.

You do not need a transformation programme. You need one workflow, scored before you automate it.

Question 1 · Can it be automated?

Is the process written down, are the rules dense, are the inputs structured?A form with a standard behind it scores high. A judgement call in a site meeting scores low.

Question 2 · Should it be?

Volume, pain, leverage.Something you do 40 times a year that blocks other people is worth automating. Something annoying you do twice a year is not, however tempting.

Then · Pick, prove, compound

Start where both say yes, and leave the rest alone.Pick one workflow with written rules, small enough to finish. Prove it beside the old way on a real job, check every output yourself, and show your lead the coverage and the iterations, not the hours. Compound: write down the prompts, the checks and the failure cases. The second workflow is faster than the first.

Start smaller than feels impressive. Then tell me what you found.

Three

Six skills that outlast the tools.

From years as a hiring manager in rail. The tools changed every few years. What I looked for did not.

01Judgement under uncertaintyDeciding with incomplete information, and owning the decision.

02Laying out the workSeeing the whole workflow, not the task in front of you. The scarce skill of 1913, and of 2026.

03VerificationChecking an answer without redoing it. Source, method, range, edge, sign.

04AccountabilityThe signature. Knowing what you can hand over, and what you never do.

05Leading across disciplinesCivil engineers end up in the middle of the room. Making change happen without all the certainty.

06Learning the tool before judging itWe would never sign off a material we had not tested. Same rule for AI.

Four

Thirty days, three moves.

Observe, experiment, adopt. In that order, and no further than you can verify.

Week 1 · Observe

Sort one week of tasks into three lanes.AI-replaced: routine, rule-based, checkable. AI-enabled: human and AI in turn. AI-assisted: the decision stays yours. Count the hours in each lane. Most people are surprised by the first.

Weeks 2 to 4 · Experiment

One AI-enabled task, on work you can check.Frame the task, give it the context, let AI draft, then verify every output. Keep a note of what it got wrong. That note is the most valuable thing you will produce this month.

Months 2 and 3 · Adopt

Redesign one workflow around the tool.Verification designed in, accountability kept with the person who signs. Then tell your team what changed, and what did not. Fit it into the old factory, or rebuild around it: this is the rebuild, at the smallest useful size.

Five

Your token quotient.

We talk about IQ and EQ. In this era there is a third: TQ, how fluently you can put a machine's tokens to work and know when not to trust them. Five yes-or-no questions. Tick what is true today.

TQ self-score

0/ 5

I can explain what a token and a context window are, and why they cost money.

Lifts it: any vendor's prompting guide, an hour, for example Anthropic's documentation.

I have built one automation end to end. A script that reads a spreadsheet and writes a report counts.

Lifts it: CS50P, Harvard's introduction to programming with Python, free.

I can say what an API does, and what MCP adds on top of it.

Lifts it: the Model Context Protocol introduction, twenty minutes.

I have chosen a small model over a large one for a task, on purpose.

Lifts it: fit the model to the task. Most engineering value sits in small, specific models on small, specific tasks.

I can verify an AI output without redoing the work.

Lifts it: the five checks in the next section.

Tick the boxes. Your score and a next step appear here.

Six

Verify the output. Five checks.

You can hand over the thinking. You cannot hand over the understanding. This is what the understanding does to an AI output before it goes anywhere.

Source. Where did each fact come from, and can I open it? An answer without a source is a draft.
Method. Would I have done it this way? What did it skip, assume or invent?
Range. Are the numbers physically plausible? Orders of magnitude, units, signs, dates.
Edge. What happens at the boundary case it was not shown? Ask it, then check the answer.
Sign. Would I put my name on it as it stands? If not, it is not finished.

Responsible use, in one paragraph. Never paste client data, drawings, commercially sensitive or safety-critical material into a public tool. Use what your organisation has approved. Record what AI produced and what a person verified. The accountability stays with the person who signs, which is what the ICE Code of Professional Conduct, April 2026, already says.

Seven

Ten words.

Token: The unit a model reads and writes, roughly three quarters of a word. You pay per token, in and out.
Context window: How many tokens the model can hold at once: your instructions, your documents, its answer.
Prompt: What you put in the window. Prompting is the first rung; context engineering is the second.
Agent: A model given tools and a goal, working in a loop until the job is done or it asks you.
API: The doorway software uses to talk to other software. How a model gets into your workflow.
MCP: Model Context Protocol. A standard doorway between a model and your tools, files and systems.
Small language model: A smaller, cheaper model tuned for one job. Often the right one for engineering tasks.
Harness: The scaffolding around a model: the checks, the tools, the limits. Where reliability comes from.
Loop: Draft, verify, correct, again. The unit of AI-enabled work.
Evaluation: A fixed set of test cases you run every time something changes. How you know it still works.

Eight

Where to start.

CS50x, Harvard's introduction to computer science · the foundations, free
CS50P, introduction to programming with Python · the language of engineering automation
Google and Kaggle's AI courses · including the five-day agents intensive
Model Context Protocol · what MCP is, from the source
Paul David, The Dynamo and the Computer, 1990 · the motor story, in eleven pages
DfT Transport AI Action Plan, June 2025 · direction
GBRX, AI in Rail: the Industry Action Plan, April 2026 · delivery
ORR Safe AI Innovation Action Plan, May 2026 · assurance
ICE Code of Professional Conduct, April 2026 · the accountability line

The newsletter

The practical version, every fortnight.

A short newsletter every two weeks on AI in rail: what worked, what did not, and one thing to try. Subscribers also get early access to Scribtive® OS when it opens. Free, and easy to leave.

Consent, UK GDPR Article 6(1)(a). See the privacy notice. The same list as the Scribtive waitlist.

Two questions I am collecting answers to. Which judgement in your work would you never hand to a machine? And which workflow did you pick, and what did the two questions say? If you feel like sharing, one sentence to bill.guo@automatex.uk is plenty. I read every one.