AI for engineers: what gets covered and where you cannot trust it
For an engineer the cost of a mistake is high: a wrong calculation or a missed contradiction surfaces on site. So what is needed here is not advice to "give it a try" but a boundary. Below: what is covered reliably, what cannot be trusted, a verification protocol that takes minutes, and what to do if the tool is not officially permitted at work.
For an engineer the cost of a mistake is high: a wrong calculation or a missed contradiction surfaces on site. So what is needed here is not advice to "give it a try" but a boundary. Below: what is covered reliably, what cannot be trusted, a verification protocol that takes minutes, and what to do if the tool is not officially permitted at work.
---
The profession where a mistake costs the most
Among the people who answered my survey there are engineers, including from industry and the financial sector. And their situation differs from everybody else's.
They need the tool, and the cost of a mistake is high. Not "the text came out badly" but a wrong calculation, a missed contradiction in a specification, an error in documentation that surfaces during installation.
So the general "give it a try, you'll like it" does not work here. You need to understand what can be handed over and what cannot, and by what protocol to verify.
Let us go through it by type of task.
What is covered confidently
They all share one feature: you supply the material rather than the model retrieving it from memory. That is the most reliable mode of working, full stop.
Analysing documents you supply
A specification, a statement of work, a set of regulations, somebody else's project. You attach the document and ask for it to be checked against specific points.
What genuinely gets found: internal contradictions, missing items, discrepancies between sections, mismatches between text and tables.
Why it works reliably: the model is not inventing standards — it is comparing what you gave it against itself.
Completeness
Checking against a checklist: is everything present that ought to be. Mechanical work where a human tires and misses things and a machine does not.
Drafts of standard documents
Statements of work, internal memos, process descriptions, instructions. To your template and your inputs.
The saving here is not in quality but in the blank page ceasing to be an obstacle.
Translating the complex into plain language
Explaining a technical decision to a client, an adjacent department or management. One of an engineer's most frequent and least liked tasks.
Structuring raw material
A meeting transcript into outcomes with owners and deadlines. Notes into a coherent document.
Finding contradictions between versions
Two revisions of a document — what changed and whether it broke something in another section.
Where you cannot trust it
The honest part, and it matters more than the previous one.
Standards, norms, document numbers. The model names non-existent ones as confidently as real ones. The mistake looks identical to a correct answer — there will be no signal saying "I am not sure here".
Calculations. The arithmetic comes out, but the choice of formula, coefficients and assumptions is a zone where an error is invisible from the look of the result. A number looks like a number.
Anything that was not in the material you supplied. If the model is drawing on memory, assume it is making things up.
Default values. Grades, cross-sections, tolerances that are "usually this". Usually does not mean in your case.
Decisions with consequences. Sign-off, signature, responsibility — a human.
The verification protocol
Five steps, and it takes minutes rather than hours.
Step 1. Split the answer into two types. Verifiable statements — numbers, values, references, designations. And reasoning — logic, structure, wording. The second you judge with expertise; the first gets checked against the source every time.
Step 2. Require sources in the request. "For each statement, state what it rests on; where there is no source, say it is an assumption." You get a list of what needs checking instead of a wall of text.
Step 3. Set the material frame. The phrasing that is compulsory for this profession: "work only from the attached document; if the answer is not in it, say the data is insufficient".
Step 4. Verify against the source rather than by asking again. Asking the same model "are you sure?" is pointless: it will either agree or reverse itself, and neither means anything.
Step 5. Responsibility sits with whoever signs. If the document goes out in your name, you answer for every value regardless of who found it.
How much this actually saves
The sums have to be honest, because "checking takes longer than doing it" is the main objection.
Going through a specification used to take a day. With a model it is twenty minutes plus verification. And you check not the whole text but the list of verifiable statements: usually five to fifteen items. A minute each is a quarter of an hour.
That is half an hour against a day. Even if the checking takes an hour, the ratio does not change qualitatively.
The saving disappears in one case — when the document contains hundreds of verifiable values. That is a sign the task was chosen wrongly: work like that is not handed over whole, it gets split up.
What to do if it is not accepted at work
A separate situation, and it is common among engineers: a large organisation, strict internal rules, the tool not officially permitted.
Concealment is a bad option. A mistake in work done with a tool nobody knows about becomes your personal problem twice over: the mistake plus the concealment.
The workable order:
Start with your own tasks — the ones only you see. Drafts, structuring notes, preparing questions.
Gather figures over two or three weeks: how long it used to take, how long it takes now.
Come to your manager with a finished result rather than a request for permission. Not "may I try this" but "here is a report, I assembled it differently and it took forty minutes instead of four hours, I verified it against the sources, here is the check".
And firmly: nothing sensitive leaves the building. Internal documents, restricted data, personal information. If you are unsure, assume it must not.
For organisations with data requirements there is a separate route — a model deployed inside the perimeter. Then the information never leaves it at all.
How this gets monetised
First — speed for yourself. The freed-up time goes to work nobody ever got round to.
Second — technical documentation as a service. Small companies are bad at it: it either does not get written or gets written by an engineer who hates it. Analysing, structuring and preparing documents is work people buy.
Third — auditing a client's documents. Contradictions and gaps found are a concrete, verifiable result. And it is also the best way in to a client: you arrive not with an offer of services but with three precise comments on their document.
The price range. From my scrape of 848 listings: a system or pipeline is $490–1 460, against a $30–75 median for one-off work.
What is available for this on the platform
The "Engineering" category — nine assistants: a programming mentor, a software engineer, a bot architect, an IT manager, prompting assistants and specialist ones.
The "Strategic analytics" category — twelve decision-making methods, including analysing a failure in advance: before a project launch it finds roughly a third more risks than the direct question "what could go wrong".
The legal category — thirteen assistants, if the documents are contractual.
Gen AI, "file → text" mode — recognising scans and drawing documentation as text, transcribing meetings. $0.15 per generation.
The express course "Perplexity as an investigative tool" — about the method of verification: how to check a fact, a reference or a value in minutes instead of taking a text at its word. For this profession it matters more than any other. $9, included in Basic.
AI Workflow — if document analysis is a regular task: a chain with a "document analyst" block and a "knowledge base" for searching your own documents.
Registration is free and opens three days of full Basic access.
Where to start today
One action. Take a document you are currently checking by hand and ask for the internal contradictions in it — discrepancies between sections, between text and tables, between versions.
Not standards. Not calculations. Specifically contradictions inside what you attached.
That is the safest task there is — the model has nothing to invent — and the most underrated: a human reading straight through tires and misses exactly those things.
---
The second approach: a pipeline rather than a chat
Everything above is work in dialogue with an assistant. There is another format, and it suits volume better.
Cascading prompts in spreadsheets
Module 1 of the programme — "Earning with ready templates" — holds Google Sheets with an integrated model: a large task split into micro-steps, each living in its own cell and taking the previous one's result as input. A formula can be dragged down a thousand rows and a thousand tasks get processed at once. How that works step by step is covered separately: [Thirty posts in one run](/en/blog/thirty-posts-one-run).
What is available for an engineer:
Eight templates for different tasks. For an engineer the most useful are the HR template for job descriptions and the business strategy template for analysing projects.
Module 1 is included in the Basic plan.
The same chains, visually: AI Workflow
When a chain needs images, video, sound or calls to external services, a spreadsheet is not enough — AI Workflow takes over: the same cascade, but on a canvas where each block's result becomes the next one's input. The quote is calculated before the run and an assembled chain is saved permanently. A detailed breakdown: [Build once, run always](/en/blog/build-once-run-always). Scheduled runs come with the Full plan.
---
What of the ready material applies to an engineer
Module 5, the lesson "Analysing reports" — going through cost and technical documentation. The model compares sections against each other and finds discrepancies.
Qwen VLM for diagrams — turns an image of a flow chart into structured data: nodes, links, directions. A scan of a diagram becomes a set of objects you can work with.
Module 5, lesson 12 — installing a local AI agent on your own machine. For an engineer in an organisation with data requirements that is the only permissible option: nothing leaves the building.
Module 5 is on Basic; Qwen VLM on N8N and Full.
---
What to read next
[Checking a fact with AI search](/en/blog/fact-checking-with-ai) — the verification method that matters more than any other for this profession.
[AI for analysts](/en/blog/ai-for-analyst-premortem) — a pre-mortem before a project launch.
[Local AI models for sensitive data](/en/blog/local-models-sensitive-data) — when nothing can leave the building.
[Transcription and file analysis](/en/blog/transcription-and-file-analysis) — recognising scans and documentation.
[Best open LLMs 2026](/en/blog/open-source-llm-guide) — the full guide to open models.
[Bringing AI into your team when you are not the boss](/en/blog/bring-ai-to-your-team) — what to do when the decision is not yours.