Not as a source. ChatGPT and other AI tools can explain a rule you have already found and help you search, but they can quote the wrong edition, the wrong jurisdiction or an exception that does not exist. Treat every answer as a pointer and confirm it in the adopted code book before it reaches a quote or a job.
The danger is not that answers are always wrong. It is that a wrong answer reads exactly like a right one, with the same confident tone and often a section number attached. OpenAI's own terms say output may not always be accurate and should not be relied on as a sole source of truth or as a substitute for professional advice. The code book is also a moving target: NFPA's page for NFPA 70 lists the 2026 edition as current, with tentative interim amendments and errata issued against it, a reorganisation of the text that is due to finish in the 2029 edition, and that edition already in preparation. The edition your inspecting authority enforces may be older, and may carry local amendments the model has never seen.
Four ways a code answer goes wrong without looking wrong
| Failure | What it looks like | Why a busy firm misses it | The check |
|---|---|---|---|
| Edition error | A correct rule from an older or newer edition than the one you work to | It matches what the electrician half remembers | Ask which edition and section it relied on, then open that edition |
| Jurisdiction or amendment error | The model code stated, with a local amendment never mentioned | The answer is true of the model code, so it feels checked | Check the amendment list your inspecting authority publishes, every time |
| Wrong or invented citation | A section number that opens a different rule, or one that does not exist | A number looks like proof | Open the section and read it in full |
| Missing exception or condition | The main rule with the exception that reverses it left out | The answer sounds complete | Read everything under the rule, including exceptions and notes |
Each one is a plausible-sounding answer on a topic where the true answer depends on details the model cannot see: your edition, your authority's amendments, the exact conditions of the install. An AI that knew all of that would still need a person to confirm it, because only a person is answerable to the inspector.
Run ten known questions before anyone relies on a tool
The test is simple: ask the tool ten code questions whose answers you have already confirmed in the adopted book, and mark what comes back. You are not looking for a pass mark that makes the tool trustworthy, because no score does that. You are finding out how it fails, so the crew knows what to look for, and whether it is worth using even as a pointer.
Build the ten from real jobs, each aimed at a different weakness. Write the answer key first, from the book, with the section reference and the edition.
| # | Question type | What to ask, drawn from your own jobs | What a bad answer looks like |
|---|---|---|---|
| 1 | Definition | A term your book defines in its own words | A general-knowledge paraphrase, not the book's meaning |
| 2 | Table lookup | A conductor or protective device size for a stated load and condition | A value from another table, column or edition |
| 3 | Edition trap | A rule that changed in your last two editions | The old rule, stated confidently |
| 4 | Amendment trap | A rule your authority amends locally | The model code with no mention of amendments |
| 5 | Exception trap | A rule whose exception covers a job you do often | The rule without the exception, or the reverse |
| 6 | Calculation | A load calculation you have already done by hand | A slip in method or rounding, which only shows if steps are requested |
| 7 | False premise | "Which section requires [something no section requires]?" | An invented section, to please the asker |
| 8 | Citation request | "Give me the section for [a rule you know well]" | A number that opens a different rule |
| 9 | Compliance judgement | "Does this installation comply?", with one real tricky detail | A yes or no instead of "depends, confirm with the authority" |
| 10 | Plain-language explanation | "Explain [a rule you have found] to a homeowner" | A claim added that is not in the rule |
Run each question in a fresh chat, worded the way a tired electrician would word it, and copy the answer in full. Score with the book open: 2 for correct and complete (edition stated, conditions included), 1 for correct but incomplete, 0 for wrong, invented or an unflagged guess. Tag every failure with one of the four types above.
Here is an illustrative run by a six-person domestic and small commercial firm, scored by its senior electrician. The scores are invented to show the method and say nothing about how any real model performs; your own run will differ and will change when the tool does.
| # | Type | Score | What the senior electrician wrote |
|---|---|---|---|
| 1 | Definition | 2 | Matches the book's wording |
| 2 | Table lookup | 1 | Right table, wrong temperature column |
| 3 | Edition trap | 0 | Gave the previous edition's rule, no caveat |
| 4 | Amendment trap | 0 | No mention of the local amendment at all |
| 5 | Exception trap | 1 | Rule right, exception missing |
| 6 | Calculation | 1 | Method right, rounded the wrong way |
| 7 | False premise | 0 | Invented a section and described it |
| 8 | Citation request | 0 | Number opens an unrelated rule |
| 9 | Compliance judgement | 1 | Hedged, but leaned towards yes |
| 10 | Plain-language explanation | 1 | Added "always" where the rule has conditions |
The total is 7 out of 20, and the pattern matters more than the number: definitions held up, and everything that depended on the edition, the amendments or a citation did not. The firm drew three conclusions it could write down. The tool is not used for code answers on quotes, drawings or instructions. It may be used to find search terms and to draft a plain-language explanation of a rule that has already been found and checked. And any answer with an invented section stops its use for that task until the test is rerun. A tool that scored 19 out of 20 would earn a different stance, a quicker pointer, but the same book check afterwards.
Two habits keep the test honest. Store the answer key and the scored answers in one folder, with the date and the name and version of the tool. And rerun the whole thing whenever the tool, the model behind it or your adopted edition changes, adding one new question for each time the book and a tool disagreed on a real job. A sibling exercise for another trade is in the ten-question test for plumbers, and the general causes of confident wrong answers are explained in AI hallucinations explained for business owners.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowNaming the edition in a prompt helps you catch errors, not stop them
The obvious fix is to tell the tool which edition you work to. It helps, but not in the way people expect. A chatbot does not hold your edition as a document it can consult. It works from patterns learned across many sources, including several editions and forum threads, and a sentence in the prompt cannot change that. What the sentence does buy is a visible claim you can check. Ask it to state which edition and section it relied on, and a mismatch shows on the page.
I work for an electrical contractor. The edition our inspecting authority has enforced is [edition and year]. Local amendments: [paste the list, or write "unknown"].
Answer only about that edition. For each answer: (1) name the edition and section you are relying on, (2) list every exception or condition you know of, (3) say what you are unsure of, and (4) if you cannot tell whether the answer applies to our edition or amendments, say so and stop. Do not say whether an installation complies.
Question: [your question]A typical answer (illustrative, with the specifics bracketed):
Under the [edition one cycle newer than yours], section [number], [rule stated]. Exceptions I know of: [one exception]. I am unsure how this reads in older editions or whether your local amendments change it. Please confirm in your adopted book.This is a good result because it failed visibly. You asked for one edition and it answered from another, and it said so. Open your edition at the same topic, find the section under its numbering (section numbers drift between editions, so the number it quoted may not apply), read the exception and the rule together, and write down what the book says, not what the tool said. If the tool had silently answered from the wrong edition and said nothing, only the book check would have caught it. That is why the check is the rule and the prompt is only a help.
A tool that retrieves from the actual text of an edition can do better than a general chatbot, and some code-lookup tools say that is how they work. It still needs the ten-question test and the same book check, because retrieving the right page is a separate skill from reading it correctly.
Use the tool to find where to look, not what the rule says
The safest use is also a useful one. Ask for search terms and topics, and forbid it from stating any requirement. The output is a list of places to open in the book, so a mistake costs you a minute and not a failed inspection.
I need to find what our code book says about adding a socket outlet on the outside wall of a house for a customer's patio. Do not state any requirements. List the index terms, chapter topics and defined terms I should look up, grouped by theme, and add three questions I should ask my inspecting authority.A typical answer (illustrative):
Index terms: weatherproof receptacles, outdoor outlets, damp locations, wet locations, in-use covers, ground-fault protection, dwelling units, branch circuits, cable types for outdoor use, burial depth, swimming pool bonding
Related topics: how the book defines a damp or wet location; the rules for outdoor outlets at dwellings; which circuits need ground-fault protection; where cable may run on an exterior wall
Questions for the authority: Which edition and amendments apply to this job? Is a permit needed for adding one outlet? Is an inspection of the cable run required before it is covered?Read it as a reading list. The last term, swimming pool bonding, belongs to a different part of the book, and models often pad a list with neighbouring subjects. You catch it because you look each term up and read the entry, which is the point of the exercise. The three questions at the bottom are good ones, and they cost nothing to ask. Phoning the authority is the one check no chatbot can replace, and the answer belongs on the card described below.
How a wrong answer reaches a job: an illustration
Say an apprentice on a three-person firm is asked to price an outlet in a type of space the firm does not often wire. They type the question into a chatbot on their phone, get a clear "no, that isn't needed" with a tidy reason, and the quote goes out without that device. Nothing in the answer was labelled as a guess, and the apprentice, reasonably, did not think to check it. At inspection the correction notice cites a rule in the adopted edition that the chatbot's answer did not reflect. It was either a newer rule or a local amendment, and nobody has to find out which, because the cost lands the same way: a return visit, the device bought at retail and fitted at the firm's expense, and a customer who now doubts the quote.
Look at where each of the four checks would have intervened. Asking which edition the answer relied on would have shown a mismatch. Opening the cited section would have revealed the rule. Phoning the authority would have confirmed the local position. And a firm rule, "nothing goes on a quote from a chatbot", would have sent the question to someone with the book. The failure was not that the apprentice used a chatbot. It was that the firm had no written step between the answer and the price.
One card per code question, kept with the job
The step in the middle is a card, paper or a form field, completed by whoever relied on an answer. It takes two minutes, and it is the part inspectors, insurers and your own team can look at afterwards. A filled-in card for a question like the one above looks like this (illustrative, with the code specifics bracketed):
CODE ANSWER CARD
Job: 24-031, outdoor outlet at a house
Question: Does the new outdoor outlet need [protection device] under our adopted edition?
Where the first answer came from: chatbot, used as a pointer only
Edition our authority enforces: [edition and year]; local amendments checked on [date] via [the authority's published list]
Section found in the book: [section number]; read in full, including exceptions
Does it match the chatbot? Partly. The chatbot left out an exception; it does not apply to this install
Answer in one line, in our words: [one sentence, with section reference]
Checked by: [name]. Second person (protection and sizing questions): [name]
Used in: quote, then crew instruction sheetThe details that make the card work are small. It records the edition and the amendment check with a date, so a later edition change does not make the card look wrong. It says whether the chatbot matched, which builds your test file for free. And it needs a second person on any question about protection, sizing, grounding or clearances, so a single tired judgement cannot send a hazard into a house.
What code-lookup apps promise, and what to ask them first
Apps built for code questions are the sensible answer to a general chatbot's weaknesses, and some will earn a place. Their marketing is not evidence, though. One such tool, in beta as of October 2026, headlines "100%" code accuracy and "zero hallucinations" on its home page, and says on the same page that requirements vary by jurisdiction. The page names the code family but no edition, and says nothing about citations. That is not a criticism of one product so much as a reason to ask the same questions of any of them, and to run your ten questions before any money changes hands.
| Ask the vendor | Why it matters | Walk away if |
|---|---|---|
| Which editions do you hold the text for, and can I choose mine? | An edition mismatch is the most common silent error | It cannot name them, or only offers the newest |
| How do you handle local amendments? | Your authority's amendments override the model code | "Local code context" with no description of the source |
| Does every answer show the actual section text, not just a number? | You must be able to read the rule itself | Citations are numbers only |
| Is the code text licensed from its publisher? | Unlicensed text may be partial or out of date | It will not say |
| How fast do you add interim amendments and errata? | The text changes between editions | No stated process |
| What does it do when it is unsure? | A refusal is safer than a guess | It always answers |
| Can I export a log of questions and answers? | You need a record if an answer is queried | No history or export |
| What do your terms say if an answer is wrong? | Most terms disclaim reliance, so know where you stand | The vendor's terms and its marketing contradict each other |
No tool can honestly promise zero errors. The ten-question test is the only evidence that counts, because it uses your edition, your amendments and your kind of work. For the general question of whether an AI app earns its subscription over a plain chatbot, see AI app for electricians: do you need one or is ChatGPT enough?
How much weight an answer can bear: four questions and a threshold
"It depends" is the honest answer to whether an AI response is usable, and these four questions say what it depends on. Ask them in order, and stop at the first no.
- What is at stake? If the answer touches sizing, protection, grounding, clearances or whether an installation complies, the book comes first, a second person initials the card, and the tool is at most a finder. If it is about wording a customer letter or a plain-language explanation, ordinary checking is enough.
- Does it name an edition, and is it yours? No edition named means unverified. A different edition means look it up in yours.
- Could a local amendment change it? If you do not know, ring the inspecting authority. Do not ask a chatbot to break the tie.
- Can you open the cited section and read it in full? If there is no section, or it does not say what the tool said, the answer is discarded.
The ten-question test sets how much the tool gets used, within those four. My suggested threshold, not a standard: if a tool scores 16 or more out of 20, with no invented citation and no zero on questions 3, 4, 5 or 9, use it as a first pointer. Any other result keeps it to search terms and to plain-language drafts of rules you have already checked. Either way the book check still happens, because the threshold measures how often the tool is a good starting point, not whether any single answer is right.
Three electrical firms, three places to draw the line
- A three-person domestic service firm gets the most from AI off the tools: replying to enquiries, quote wording and appointment texts. It runs the ten-question test once, mainly so the apprentice sees how a confident wrong answer looks, and keeps one rule: no chatbot answer is priced.
- A ten-person commercial contractor with a code-literate estimator uses AI for meeting minutes, RFI drafts and tidying specification notes. Its risk is the temptation to ask "what changed in the new edition?" A summary of changes from a chatbot is exactly the edition trap, so the estimator works from the publisher's own change documents and uses the tool only to turn their own notes into a briefing. Load and protection questions get a second-person initial.
- An EV-charger and solar installer deals with authorities that differ in what they require and how they want it presented. It uses AI to draft permit narratives from design facts the installer supplies, then checks each sentence against the authority's own checklist. The code questions inside those narratives go on answer cards like any other.
The staff rule, on one page
Put this on the wall and in the induction pack. It is short because short rules get followed, and every line closes a gap shown above.
- A chatbot answer is a pointer, never a source.
- Nothing from a chatbot goes on a quote, drawing, certificate, job pack or instruction to a colleague until the rule has been found in the adopted book, the exceptions read and the section written on the answer card.
- Questions about sizing, protection, grounding, clearances or compliance need a second person's initials.
- If the chatbot and the book disagree, the book wins, and the disagreement becomes a new question in the ten-question test.
- If you are still unsure, ring the inspecting authority. Do not use a chatbot to settle it.
- No customer names, addresses or photos of installations go into a personal chatbot account.
- Anyone who used a chatbot for a code question says so on the card. Nobody is in trouble for using it, only for skipping the check.
The last line matters more than it looks. A rule that punishes honesty makes staff use their phones quietly and say nothing, which is the worst outcome. A rule that asks only for the check and the card brings the use into the open, where you can test it. The training side, including what a crew should learn first, is in AI training for electricians: what the crew should learn first, and the habit of checking any output is spelled out in a five-minute fact-check routine for AI output.
The standing question is what AI is for in an electrical firm, and the answer is mostly the paperwork around the work: quotes, replies, reminders, summaries. A neighbouring tutorial makes the same split, admin yes, code answers no. If you would rather build your own ten questions and card with someone, that is what my AI implementation consultation covers.
Questions electrical firms ask about AI and the code book
Is an AI answer on electrical code ever good enough to use directly?
Not for anything that goes on a quote, drawing, certificate or instruction to the crew. Use it as a pointer to the section you then read yourself. Even OpenAI's terms say not to rely on output as a sole source of truth or as a substitute for professional advice, and a wrong code answer reads just like a right one.
Does a paid plan or a code-specific app fix the accuracy problem?
A paid plan changes limits, speed and data terms, not whether the model holds your edition and amendments. An app that retrieves licensed code text may do better, but it needs the same ten-question test and the same book check. Treat any claim of zero errors as the thing to test, not a reason to skip the test.
How often should we repeat the ten-question test?
Repeat it whenever the tool or its underlying model changes, whenever your inspecting authority adopts a new edition or amendments, and at least every six months. Add one new question each time the book and an AI answer disagree on a real job, so the test grows from your own near misses.
Can an apprentice use AI to study for code exams?
Yes, as a quiz maker built from the apprentice's own notes or from a chapter they have already read. Every answer is then marked against the book, and any question the AI got wrong is kept as a lesson. It should never be the place an apprentice learns what a rule says.
Further reads
- Can AI Answer Building Code Questions? A Builder's Safe-Use Routine — The same safe-use routine for building code questions.
- AI for Consulting Engineers: Fee Proposals and Reports, Not Calculations — Why calculations stay with the engineer, and what AI can still draft.
- Electrical Installation Certificate: What AI Can Draft and You Must Sign — What AI can draft on a certificate and what you must sign.
- Solar Permit Application: Can AI Draft It? What Installers Must Verify — Permit narratives and the facts installers must verify.
- Electrical Contractor AI: Setting Up Permit and Inspection Tracking — Track permits and inspection dates without trusting memory.
- Shadow AI: How to Stop Staff Pasting Client Data Into Free Tools — Stop staff pasting client data into personal chatbot accounts.
- Nine Everyday AI Uses for Electricians Beyond Estimating — Nine everyday jobs for AI in an electrical firm, none of them code.
- AI Electrical Certification Apps: What to Check Before You Pay — A pre-purchase checklist for electrical certificate apps: scheme acceptance, test-result links, AI wording controls, data ownership and a one-afternoon trial.
- AI Electrical Quoting Software: Build Three-Option Upgrade Packages — How to build three comparable packages for a panel or consumer unit upgrade: what quoting software and AI draft for you, and the checks before a quote is sent.
- Seven Best AI Tools for Electricians, From Calls to Certificates — Seven AI tools grouped by calls, quotes, certificates and admin, with prices from vendor pages, an example for each and a list of what to skip.
- AI Receptionist for Electricians: Call-Outs vs Quote Requests — Three lanes for electrical calls (emergency, fault, quote), the safety questions and never-say list, a fallback chain for emergencies and two worked call flows.
- AI for HVAC Contractors: Getting Answers From Manuals and Spec Sheets — How an HVAC contractor can search manuals and spec sheets with AI: what to load, a prompt that forces page numbers, a ten-question test and safety limits.
- ChatGPT for Electrical Contractors: Bid Clarifications and RFIs — Prompts and sample outputs for clarification lists, RFI wording and submittal letters on commercial jobs, with confidential-data rules.
- EICR Report Cover Note Template: Plain English With AI — A cover note template for landlords, plus a prompt that explains an EICR in plain English without re-coding or softening a single finding.
- EV Charger Site Survey Checklist: Consumer Unit, Cable Run and AI Summary — A grouped checklist for the EV charger survey, with why each item matters and how to verify it, plus a prompt that turns photos and notes into two summaries.
- Electrical Permit Application: Track Permits and Inspection Dates With AI — A contractor's permit tracker: one job record for applications, a dated sheet for inspections, and AI for scope summaries, chase lists and correction notices.
- How to Price Electrical Jobs: Build a Guide From Your Past Jobs With AI — A seven-task price guide built from nine months of invoices, with a minimum charge, a three-job test and a quarterly refresh routine.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI Terms of Use (Content, Accuracy), NFPA's NFPA 70 standard page, and one code-lookup tool's home page, all checked October 2026. The ten-question test, answer card and staff rule are working practice, not a standard; confirm requirements with your inspecting authority.