How to Use AI as an Electrical Contractor: Admin Yes, Code Answers No

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Use AI as an Electrical Contractor: Admin Yes, Code Answers No.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Use AI as an Electrical Contractor: Admin Yes, Code Answers No.

Use ChatGPT for admin, not for code answers. Quotes, follow-up emails, job records, supplier chasers and first drafts of paperwork suit it, or Claude. Code questions, cable sizing, protective-device choices and certification stay with the qualified person, working from the code edition their authority has adopted, because a chat answer can be fluent, specific and wrong.

In a chat window a quote question and a code question look identical, which is why the firm needs a written rule that sorts every job before anyone types it. The sorting is simple once it is on paper, and the admin side alone is worth the effort: the illustrative week worked through below frees about four hours without going near a circuit.

Follow me on Instagram@sagnikteaches

Sort every job into green, amber or red

The test for each job is one question: if an inspector, an insurer or a customer's lawyer asked who decided this, would the answer have to be a named qualified person? If yes, the job is red and no assistant answers it. If the answer is "I wrote it, and I read it before it went", the job is green. Amber jobs are drafts that a qualified person edits and owns.

Connect on LinkedInSagnik Bhattacharya
ColourJobWhyWhat the person does
GreenQuote email from site-visit notesPrices come from your list, not the assistantReads every line, checks totals, sends
GreenFollow-up after a quoteWording onlyAdds the real dates and the next decision
GreenCustomer summary of work done todayFacts come from your notesConfirms every fact against the job
GreenSupplier orders and chasersQuantities come from your listChecks part numbers and quantities
GreenDictated note into a job recordLayout only, with test fields left blankFills the readings from the test sheet
GreenSite-visit notes into an action listSorting, not decidingAssigns owners and dates
AmberScope and exclusions wordingAn exclusion can carry safety or cost meaningEdits technical lines, owns the final text
AmberMethod statement or risk assessment draftHazards and controls are the competent person'sWrites the hazards and controls, uses the draft for layout
AmberTerms and conditions draftA contract has legal effectHas an adviser review before use
AmberPlain-English explanation of work for a customerA friendly sentence can still be technically wrongChecks each technical claim
AmberSummary of a tender or manufacturer manualA summary can drop the one clause that mattersOpens the source for each quoted sentence
RedCable size, protective device, voltage drop, maximum demandDesign decisions with safety consequencesCalculates from the adopted code and the site
RedWhat the code requires here, or whether work compliesEdition, amendments and site decide the answerReads the clause in the adopted edition
RedTest readings, pass or fail, certificates and declarationsThe signer carries the responsibilityTests, records and signs
RedSafe isolation, lock-out and live-work decisionsPeople can be hurt on the dayFollows the firm's procedure
RedCalling a defect fine or dangerous from a photoA photo hides what a test showsInspects and tests in person

Your licensing body or insurer may draw the lines differently, and the rule is yours to adjust. What matters is that it exists in writing, and that red means red. The tutorial on what AI can draft and you must sign covers the certificate side in more detail, and the risk assessment and method statement tutorial explains why the amber risk-assessment job keeps a named signer.

Subscribe on YouTube@sagnikteaches

Why a confident code answer can still be wrong

A code answer is the product of an edition, any local amendments, the type of installation and the conditions on the day. An assistant sees only the words in your question. Ask it, for example, what cable a 7 kW car charger needs on a 25 m run (illustrative), and a typical reply names a size, cites a table and a clause, and closes by saying the installation complies. It reads like a senior colleague. But the answer depends on how the cable is installed, the ambient temperature, grouping with other circuits, the voltage drop over the run, the protective device and the edition your authority has adopted. The assistant knows none of those from your question, and it will not say so unless you make it.

Three things make it worse in practice. First, wiring codes are republished on a cycle, and each authority adopts a new edition on its own timetable, usually with amendments. For one widely used code the cycle is three years; some authorities adopt within months and some take years, so the edition in force on your job may not be the newest. Second, an assistant learns from text published at different dates, including forum posts and trade articles that quote older editions, so it can cite a clause number that belongs to a different edition or does not exist. Third, a summary of a clause is not the clause. Amendments change numbers inside clauses, and a tidy paraphrase hides the change.

None of this means the assistant is useless. It means a code answer needs a person who can open the adopted edition, read the clause, apply the site conditions and put their name to the result. That person was always the safeguard. An assistant that answers first only makes the safeguard feel optional.

The Complete Excel Course with AI Integration thumbnail

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 now

Check the adopted edition before anyone quotes a clause

Even a firm that never asks an assistant a code question needs this routine, because the same edition risk applies to staff, to apprentices and to old habits. Here is the routine, in order:

  1. Find who decides. This is the authority that issues your permits or sets your licence conditions. Get the edition, the amendment level and the date it took effect, from their website or by phone.
  2. Write it on one line: "Adopted edition: [edition and amendments], in effect from [date], confirmed with [who] on [date]." Put the line in the van, on the firm's AI rule page and in any assistant instructions.
  3. Match your shelf to it. The edition you hold, in print or as a licensed file, should be the one on the line. If your authority has moved on, replace your copy.
  4. Re-check every six months and whenever a new edition is published. Put it in the calendar, the way you would for test-instrument calibration.
  5. For learning only, paste a clause from your own copy and ask the assistant to explain the wording in plain English. Do not ask it to produce the clause from memory. Check the licence terms of your copy before pasting text from it.

If you want to see how an assistant behaves on questions where you already know the answer, the tutorial on answering code questions safely sets out a ten-question test. Run it once, on your own edition, with no customer data, and keep the score. It tends to settle arguments inside a firm quickly.

A quote email drafted the safe way

Quotes are the best green job because the numbers come from your price list and the assistant only arranges words. The prompt below does two extra things: it forbids electrical specifications, and it makes the assistant list every technical assumption for a qualified person to confirm, instead of burying them.

Draft a quote email from these site notes and my price list. Use only
the prices in my list. Do not choose cable sizes, breaker ratings or
any other electrical specification. Do not say the work complies with
anything. List every technical assumption the quote depends on under
"To confirm on site" and mark each [QUALIFIED PERSON]. Mark anything
you could not price as NOT PRICED.

NOTES: Utility and kitchen. 6 double sockets on a new circuit from the
existing distribution board, 8 downlights in the kitchen ceiling, extractor
fan connection. Access through the loft. Two days, power off for about
half a day. Customer asked if we could also look at the garden lights,
no detail given.
PRICE LIST: socket $85 each; downlight $62 each; extractor connection
$120; new circuit allowance $380; testing and certification $150.

A reply (illustrative) puts the lines in a table: sockets 6 × $85 = $510, downlights 8 × $62 = $496, extractor connection $120, new circuit allowance $380, testing and certification $150, total $1,656. Under "To confirm on site" it lists a spare way in the existing distribution board [QUALIFIED PERSON], the condition of the existing earthing and bonding [QUALIFIED PERSON] and loft access and insulation clearance for the downlights [QUALIFIED PERSON]. The garden lights are NOT PRICED because no detail was given. You read the arithmetic yourself, because the total is where a chat assistant can slip: 510 + 496 + 120 + 380 + 150 does come to 1,656.

Two lines in a first draft like this need to go. One says "all work will comply with current regulations and standards". The firm should not promise compliance on an assistant's say-so, so it becomes "on completion we will test the work and issue the certification required for it". The other names a cable size and a breaker rating for the new circuit. Neither is in the notes, both look entirely plausible, and both are exactly what the assistant must not decide. Replace them with "cable and protective device for the new circuit are sized and specified by our qualified electrician after the site survey". The prompt's ban on specifications is what stops this happening twice. For a fuller template, the domestic quote template tutorial goes through scope and exclusions line by line.

Two more green jobs: the follow-up and the job record

The follow-up. A message that says "just checking in" gets a shrug. Give the assistant the quote total, the slot you can offer and the one decision you need, and ask for three sentences. The reply (illustrative): "Hi [first name], following our quote of $1,656 for the utility and kitchen work, we can hold a two-day slot from [date]. The power would be off for about half a day. Do you want us to book it, or is there anything you would like us to change first? [your name]". The slot and date come from your diary. The price comes from your quote. The assistant supplied only the order of the sentences.

The job record. Dictate what you did, then ask for a record with fixed fields, and put the red items into the template as fields the assistant may not fill. A note such as "Fitted the six sockets, tested it all, all good, customer happy, left Thursday" goes in. The record comes out with the date, the work done, the customer sign-off line and the following header:

Work done: six double sockets fitted on the new circuit
Test results: NOT STATED (enter from the test sheet)
Tested by: NOT STATED
Customer: satisfied, per note
Follow-up needed: NOT STATED

Without that rule, an assistant can turn "tested it all, all good" into "Tests passed". That phrase would then sit in your file as a statement about test results that nobody measured. Leaving the fields as NOT STATED, and filling them from the test sheet by the person who tested, keeps the record true.

A holding reply is also a green job, even when the customer's question is red. A customer messages to ask whether an electric shower can go on an existing circuit. The question is a design decision. The assistant can still draft the reply: "Thanks for asking. That depends on the existing installation, which we need to see. A qualified electrician will check the supply and the circuit and give you a clear answer. Can we book a visit?" The wording is green; the answer stays red.

An amber job: scope wording the qualified person edits

Scope and exclusions text looks like admin, which is why it needs the amber label. Ask an assistant to turn the quote notes into a scope paragraph and it will produce something tidy that includes commitments you never made. A draft for the utility and kitchen job (illustrative) might read:

Scope: supply and install six double sockets on a new circuit
using appropriately rated cable and protective devices, in accordance
with all applicable standards. Supply and install eight downlights.
Connect the customer's extractor fan.
Exclusions: decoration and making good. Existing wiring is assumed to
be in good condition. Any remedial work found during testing will be
charged at our hourly rate.

Three lines need a person. "Appropriately rated cable and protective devices, in accordance with all applicable standards" sounds technical and promises something only the designer can promise, so it becomes "cable and protective devices specified by our qualified electrician". "Existing wiring is assumed to be in good condition" is a statement about an installation nobody has inspected, so it becomes "existing wiring has not been inspected as part of this quote". "Remedial work charged at our hourly rate" is a commercial term the owner has to choose, not the assistant: you might price remedials separately, agree a call-out before any extra work starts, or cap them. The assistant was right to raise the point and wrong to decide it.

That is the whole of amber work in practice. The assistant produces the shape and the common clauses, you strike or rewrite every line that carries a technical or commercial meaning, and the final text is yours. For a firm that sends a dozen quotes a week, a short list of approved exclusions pasted into the standing instructions saves most of this editing, and it keeps the wording consistent between jobs.

A week of green jobs in numbers

Take a four-person firm with an owner, two electricians and a part-time office person. These are illustrative timings, with checking time included in the "with assistant" column, so replace them with a week of your own:

JobJobs a weekMinutes each beforeMinutes each with assistantMinutes saved
Quote emails82510120
Follow-ups after quotes106240
Customer summaries515645
Supplier emails48320
Dictated notes into job records610436
Total33427 minutes166 minutes261 minutes

That is about 4.3 hours saved from about 7.1 hours of admin. At an illustrative $55 an hour for the owner's or office person's time, it is roughly $240 a week of time, against a plan of $20 a month for one personal seat, or $40 to $50 for two business seats. The saving holds only if the checking stays in. Most of it comes from quote emails, which is why they are the first job to switch on, and none of it touches a code question. If your own week shows half this, the sum still favours the plan by a wide margin, but time your own jobs before you rely on it.

A twenty-minute sorting exercise for the crew

A rule only works if people can apply it quickly. Write five jobs on cards, give each person the colours, and have them sort the cards without talking. Then compare. Five cards to try (answers in brackets):

  • "A customer wants to know if the old fuse board or panel is safe." (Red. It needs an inspection and a person's judgement.)
  • "Write a chaser to the supplier for the breakers missing from the last order." (Green. Part numbers come from the order sheet.)
  • "Explain to a customer what testing involves before they sign." (Amber. The wording is friendly, but each technical claim needs checking.)
  • "Summarise the 60-page tender instructions." (Amber. Ask for the sentence behind each point and open the source.)
  • "Which size cable for the new oven circuit?" (Red. Never typed into an assistant for an answer.)

Disagreements are the useful part. Most arguments come from jobs that look like admin but carry a safety meaning, such as a customer explanation or a summary. When two people sort a card differently, ask which of them would be named if it went wrong. That answer usually settles the colour.

One page for the van: the firm's AI rule

Print one page, pin it up, and have everyone who uses an assistant sign it. A filled-in template:

AI RULE FOR [FIRM NAME], version 1, [date]

GREEN (anyone, read every line before sending): quotes, follow-ups,
customer summaries, supplier emails, tidying notes into job records,
action lists from site-visit notes.

AMBER (draft only; [name] edits and owns the final text): scope and
exclusions wording, method statements, risk assessments, terms and
conditions, customer explanations of electrical work, summaries of
tenders and manuals.

RED (never typed into an assistant for an answer; ask [name]): cable
and device sizing, load and maximum demand, voltage drop, test readings
and pass or fail, certificates and declarations, safe isolation,
anything that says an installation complies, hazardous or special
locations.

ADOPTED CODE: [edition and amendments], in effect from [date],
confirmed with [authority] on [date]. Re-check every six months.

CUSTOMER DETAILS: no names, addresses or phone numbers in personal or
free accounts.

CHATBOTS AND RECEPTIONISTS: they decline code questions and offer a
call with [name].

LOG: note on the job file whenever an assistant drafted a document.

Revisit the page after the first month. Where staff keep asking about an amber or red item, the fix is usually a better green alternative, such as a holding reply or a checklist, not a looser rule.

Grey areas: chatbots, apprentices, photos and manuals

Chatbots and receptionists. A website chatbot or AI receptionist talks to customers without you watching, so it must follow the same rule: decline code questions, offer a visit, and never describe an installation as safe or compliant. The AI receptionist tutorial for electricians separates call-outs from quote requests, and the same split applies to what a bot may answer.

Apprentices. An assistant can help an apprentice study, for example by quizzing them on definitions from their own study material. It cannot be the source for a decision on a live job. Make that distinction explicit, because the same tool is open in both cases.

Red questions from inside the firm. A junior electrician finishing late may type a cable question into a phone assistant because the boss is not answering. Telling people not to is not enough. Make asking cheap: a shared group chat, a standing five-minute call at the start of the day, or a named person who replies to red questions within the hour. Then log the question and the answer. After a month the log is a list of the things your team does not yet know, which is a better training plan than a generic checklist.

Counting from drawings or photos. Counting fittings from a plan is tempting. Anthropic's image documentation says counts of objects in an image can be imprecise, especially with many small items, and that location outputs are approximate. A count from an image is a first pass you recount before it goes into a price.

Photos of installations. An assistant will describe a photo of a distribution board with great confidence. Describing is not inspecting, and nothing in a photograph shows a loose terminal, a poor joint or a test reading. This is red.

Manufacturer manuals. Pasting an installation manual and asking for a summary is amber. Ask for the exact sentence behind each point and open the manual to read it, because a summary can omit the one warning that matters on your job.

The first two weeks are easy to plan. On day one write the page. In the first week switch on quote emails and follow-ups and nothing else. In the second week add job records and supplier emails, time every task, and tick the scoreboard. If you would rather have the page and the first jobs set up alongside someone, that is what my AI implementation consultation covers.

Questions electrical firms ask about AI and code

Is it ever safe to ask an AI an electrical code question?

As a study aid, yes: paste the clause text from your own copy of the code and ask the assistant to explain the wording in plain English, then read the clause yourself. It is not safe as a source of answers for work you will sign. The assistant has no knowledge of your adopted edition, your amendments or the site, and it can sound sure while being wrong.

What if a customer asks our website chatbot or AI receptionist a code question?

Set it to decline and offer a call or visit with a qualified person. A good holding reply says the answer depends on the existing installation, which someone needs to see. Test the chatbot with five code questions before launch, and review its logs monthly, because a bot that answers code questions has made your firm the author of the answer.

Which plan should a small electrical firm start on?

Start with one personal plan at about $20 a month (ChatGPT Plus or Claude Pro) and keep customer names and addresses out of it. Move to business seats, around $25 per user monthly or $20 on annual billing with two seats minimum, once an office person shares the work. Business plans do not train on your content by default.

How do I find out which code edition applies to my work?

Ask the authority that issues your permits or sets your licence conditions, and write down the edition, any amendments and the date it took effect. Editions are republished on cycles and authorities adopt them on their own timetable, so check again every six months and whenever a new edition appears. Keep the answer on one line where the whole team can see it.

Further reads

Sources: OpenAI and Anthropic plan pages for seat prices and training defaults, Anthropic's image-analysis documentation (counts and locations from images are approximate), and industry explanations of how wiring-code editions are revised and adopted authority by authority.

Want a safe AI rule written for your electrical firm?

On a 1:1 call we can sort your firm's weekly tasks into green, amber and red, set up the first two admin jobs in an assistant, and write the one-page rule your team signs.

Book a 1:1 call with me