Today an AI agent can sort a shared inbox, draft replies, check documents such as insurance certificates against a list of requirements, and turn meeting notes into action lists, provided a person reads the output before anything goes out. Unsupervised estimating, quantity take-offs, schedule changes and safety sign-off are still mostly marketing.
The deciding question is whether a mistake is cheap to catch. A wrong label on an email costs seconds; a wrong dimension in a take-off costs a job. Most vendor guides describe the best case on tidy files, so the useful test is your own: your inbox, your documents and a pass mark agreed before the trial starts. "Works" here means three things together: the task is common, a miss is easy to spot, and a miss is cheap. It does not mean a measured accuracy figure, because the published figures I found for this kind of work were supplied by vendors.
An agent is a chatbot that can act, and that changes the risk
A chatbot answers inside a chat window and stops. An agent is given a goal, can use tools (read a mailbox, open files, add calendar entries, fill in a form, draft or send an email) and chooses its own next step until it decides the job is done. ChatGPT Work, which replaced ChatGPT agent in July 2026 and sits in Plus, Pro, Business and Enterprise at their existing prices, is one current example. Claude (formerly Cowork) is another. Both can work through several steps across your files and apps instead of answering a single question.
Three things change when a chatbot becomes an agent. It reaches real systems, so a wrong step can leave the chat window as a sent email or an edited record. Errors compound: if each step is right 95 times in 100 (an assumption for the sum, not a measured figure for any product), a ten-step job comes out fully right only about 60 times in 100, because 0.95 multiplied by itself ten times is 0.60. And instructions can arrive from outside, because an agent that reads your mail also reads whatever a stranger wrote in it.
| Type | What it does | Fits a building firm when | Usual way it goes wrong |
|---|---|---|---|
| Chatbot | Answers questions and drafts text from what you paste in, then stops | You want a reply drafted or a subcontract clause explained | A wrong answer you can see before you use it |
| Automation rule | Does the same fixed thing whenever a trigger fires | Every new enquiry should land in the job sheet, or a reminder should always go out | Breaks without a sound when a form or file name changes |
| Agent | Pursues a goal across tools and picks its own steps | The job needs reading, sorting and several small steps, such as clearing a mailbox or assembling a daily log | A wrong step you never see, or an action in a real system that cannot be undone |
Many jobs people want an agent for are cheaper and safer as a fixed rule plus one chatbot step. The comparison of agent, chatbot and automation covers that choice in full. The question here is narrower: where does an agent earn its extra risk in a building firm?
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 nowReading a vendor's number: what the footnote tells you
The pages that rank for this subject are mostly written by software vendors, and a vendor's job is to show the best case. Take a headline row from a construction platform's homepage that I read while preparing this tutorial: "6+ hrs saved daily", "50% less admin work" and "90% decreased project risk". The small print says the figures were reported by active users, and the time-saving figure by users who had the vendor's AI feature switched on. So the numbers are self-reported, there is no comparison group, and the page does not say how "project risk" was measured. That says nothing against the product. It means the figure cannot go into your business case.
Five questions turn any claim into something you can test.
- Whose files? A demo on the vendor's tidy sample project tells you little about your scanned certificates and six-reply email chains.
- Which exact task? "Handles RFIs" can mean classifying incoming ones, drafting answers or sending them. Those carry very different risk.
- Compared with what? Hours saved should be measured against a stopwatch on the same task, not against a guess.
- Who checked the output? If a person reviewed every result, the time saving already includes that checking. If nobody did, the error rate is unknown.
- What happens when it is wrong? Ask to see the log, the undo and who gets told.
Ask for a live run on ten of your own documents, and choose the ugly ones yourself: a phone photo of a certificate, a drawing with hand markups, a long email chain that ends with "see below". A vendor confident in the product will agree. Hesitation is useful information too.
Four jobs that hold up, each with a check you can run
Sorting a shared inbox and drafting first replies
Say a twelve-person firm doing extensions and refurbishments receives about 55 messages a day in a shared inbox. The office manager spends roughly 90 minutes each morning working out what each one is and who should deal with it. For the trial, the assistant gets read access to that inbox and permission to write into a Drafts folder and apply labels. It cannot send, delete, forward or accept calendar invitations. Its instruction names the categories and says what to do with each. It does not say "handle my email".
The first week is a shadow run. The agent labels every message and drafts replies to new enquiries, while the office manager sorts the inbox as usual and records each disagreement. The pass mark is set beforehand: at least 95% agreement overall, and no complaint or urgent problem labelled as routine. All figures in this example are illustrative.
| Category | Messages in five days | Agent agreed | Disagreed |
|---|---|---|---|
| New enquiry | 40 | 38 | 2 |
| Existing client question | 52 | 50 | 2 |
| Supplier invoice or statement | 61 | 61 | 0 |
| Subcontractor question or certificate | 47 | 44 | 3 |
| Automated notice (portal alert, newsletter) | 66 | 65 | 1 |
| Complaint or urgent problem | 9 | 6 | 3 |
| Total | 275 | 264 | 11 |
Overall agreement is 96%, which would pass a casual glance. The second rule fails: three of the nine complaints were filed as routine. They were the polite kind ("just checking the screed is meant to look like this?" with a photo of a crack), and nothing in the words flagged them. The fix was a short paragraph in the instruction listing what counts as a complaint (any photo of damage, any question beginning "is it meant to", any mention of a leak, a claim, a lawyer or a delay), plus a rule that when the agent is torn between "routine" and "complaint" it picks "complaint". A second shadow week on 250 new messages missed none of eight complaints, and the trial moved on to live drafts. An average can pass while the one case that matters fails, so write the second rule before the first day.
The sums: sorting by hand takes about 1.6 minutes a message, so 55 messages is 88 minutes. With the agent, checking a label takes about 20 seconds (18 minutes for 55) and reading and editing ten drafted replies takes about two minutes each (20 minutes), so the morning is about 38 minutes. That frees roughly 50 minutes a day, but only while the office manager keeps checking. If the checking stops, the saving is an illusion.
Checking a document against a list of requirements
Document checks suit agents because the answer is a comparison with a fixed list and the output can quote its evidence. Subcontractor insurance is the classic case: you hold a requirements list, the subcontractor sends a certificate as a PDF or a phone photo, and someone has to tick the boxes. A prompt that forces evidence and allows "can't tell" works in a plain chat tool as well as in an agent, and the wider routine of tracking insurance expiry and onboarding with AI builds on it.
Check the attached insurance certificate against these requirements
for a job running 12 April to 30 June 2027.
1. Public or general liability cover of at least $2,000,000 per claim.
2. Employers' liability or workers' compensation cover shown, or a
statement that the firm has no employees.
3. The policy period covers the whole job.
4. The named insured matches "[subcontractor name]".
For each requirement reply with PASS, FAIL or CAN'T TELL, the exact words
you relied on, and the page. If a date could be read as day/month or
month/day, reply CAN'T TELL and say why. Do not guess missing information.Below is an illustrative first answer, with what a person should have caught.
| Requirement | Agent's first answer | What a person should see |
|---|---|---|
| 1. Liability cover | PASS: "Public Liability $5,000,000 any one occurrence", page 1 | Correct, and the quote is checkable |
| 2. Cover for employees | FAIL: not shown, no statement about employees | Correct: ask the subcontractor which applies |
| 3. Policy period | PASS: "expires 6 July 2027" | The certificate says 06/07/2027, which is 6 July day-first but 7 June month-first. The job ends 30 June, so the result depends on the reading |
| 4. Named insured | PASS: name matches | Correct |
Requirement 3 is the realistic failure: the agent chose one reading of an ambiguous date and presented it as fact. After the "CAN'T TELL" sentence went into the prompt, the same certificate came back as CAN'T TELL with the reason, and the office clerk phoned the broker. The pattern to copy is that every pass must carry a quotation, and every doubt has a named way out.
Turning voice notes and meeting notes into diary entries and actions
Site managers talk faster than they type, which makes voice notes the cheapest input a building firm has. An agent, or even a plain transcription tool followed by a chatbot, can turn a rambling note into a tidy diary entry and an action list. A before and after shows it. The raw transcript, illustrative:
so its Tuesday, [site manager] and two lads on site, started the blockwork at the back, about half done, rain stopped us after lunch, we need to order the lintels Thursday latest, the client asked about moving a socket so thats a possible variation, scaffolders coming Friday not Thursday now
The agent's first draft:
Tuesday. Present: site manager and two operatives. Blockwork to the rear elevation started, about half complete. Rain stopped work after lunch. Lintels ordered for Thursday. Client asked to move a socket (possible variation, not priced). Scaffold now Friday, previously Thursday.
Almost everything is right, and the mistake is easy to miss: "we need to order the lintels Thursday latest" became "lintels ordered for Thursday". A diary entry is a record, and a record that says something was done when it was only planned is worse than no record. The instruction that fixed it: keep the speaker's tense, treat "need to", "should" and "going to" as actions rather than facts, and list every action with an owner and a date at the bottom. A cheap extra check is to read the actions list aloud at the Friday meeting, because any line nobody recognises is a candidate error. The tools that handle RFIs, logs and reports work from exactly this kind of input.
Comparing supplier quotes on landed cost
Pasting three quotes into an assistant and asking for a comparison table is a good job, as long as the table is in landed cost, meaning the total you would pay delivered for the quantity you need. The trap is easy to fall into. Suppose a job needs 100 sheets of plasterboard. Supplier A quotes $11.40 a sheet with free delivery over $300. Supplier B quotes $540 a pallet of 50 sheets plus $85 delivery. A table that compares $11.40 with $10.80 a sheet picks B. The real totals are $1,140 for A, and $1,080 plus $85, which is $1,165, for B. B costs $25 more.
The prompt line that prevents it is: "Give the total cost delivered for exactly 100 sheets for each supplier, showing the sum, and list anything the quote does not state." Then check the sums yourself, because arithmetic is where assistants slip silently. The four jobs share one shape: the input is in your own documents, the output can quote its source, and a person can see a failure in a minute.
Jobs that still fail, and the wording that hides it
These are the claims to meet with a demo on your own files. None is impossible in principle. Each one asks an agent to act without supervision on something where a miss is expensive and hard to see.
| The claim | What is happening behind it | Who keeps the job today |
|---|---|---|
| "Reads your drawings and produces the take-off" | Text on a drawing is read well. Lengths measured from the image depend on scale and on how the PDF was exported, and long sets may be read as text only | The estimator, with the agent's quantities checked against one hand take-off |
| "Prices the job automatically" | It can total rates you supply. It does not know your supplier prices, your labour rates or what a tight access costs | The estimator |
| "Optimises your schedule" | It can propose a sequence. It does not know the crane is booked, the screed needs to dry or a neighbour objects to Saturday noise | The site manager |
| "AI safety monitoring" | Camera tools that spot a missing hard hat are image-recognition products. They flag, and a person decides | A competent person, as before |
| "Chases late payment for you" | A reminder email is a fair draft. A formal notice that starts a clock under your contract is a legal step | Whoever administers the contract, with your adviser |
| "Negotiates with subcontractors and suppliers" | It cannot commit your firm, and a message that sounds like agreement can be read as agreement | The person with authority to commit the firm |
| "Approves invoices and pay applications" | It can compare an application with the budget and flag mismatches. Approval is an accounting control | The person who signs |
Take the first row, because a wrong quantity does the most damage. Here is the kind of failure to test for. A plan exported "fit to page" still prints 1:100 in its title block, but the sheet is now at 80% of that scale. Every length measured from the image comes out short by the same fraction: a 10 m wall reads 8 m, and because area scales with the square, a 40-square-metre floor reads 25.6 square metres (0.8 multiplied by 0.8 is 0.64). That is a 36% shortfall in a number someone will price. A human estimator reads the written dimensions and a scale bar. To test an agent, give it a drawing you have already measured by hand and see whether a 4.2 m wall comes back as 4.2 m.
Format limits matter as well. Anthropic's upload help page says Claude analyses text and visual elements in PDFs of 100 pages or fewer, and for PDFs of 101 to 1,000 pages it processes text only. A full drawing set can pass 100 sheets, so the agent may be reading the notes and ignoring the linework. The same page lists PDF, Word, spreadsheet, text and image formats but no CAD formats such as DWG or DXF. If the plans live in CAD, the agent sees an export of them, not the model. For the estimating side, bid checks after the take-off shows where AI helps once a person has produced the quantities.
Supervision rules for an agent that touches your inbox
The permission ladder
Give every agent the lowest level that lets it do the job, and move up only on evidence.
- Reads only. It can see the mailbox or folder and report. It changes nothing.
- Drafts. It can apply labels and write drafts, notes and summaries. A person sends or files everything.
- Acts after approval. It proposes an action (send this reply, add this calendar entry) and a person clicks to approve each one.
- Acts alone. No construction job in this tutorial earns this level yet.
Moving from level 2 to level 3 should need a written record: 20 working days with no draft that would have gone out wrong and no missed urgent item. Email connectors now do more than read. Claude's Google Workspace connector, for example, can draft, reply, forward and send Gmail, so check what permissions each connector grants before you connect it, and do not allow sending until the trial has earned it.
Why a stranger's email is part of the risk
An agent that reads email also reads emails written by people who want something. OpenAI's own page on prompt injection uses an example any building office will recognise: an agent told to respond to overnight emails meets one message, written by an attacker, that tricks it into finding the owner's bank statements. The same page advises limiting an agent's access to the data and credentials the task needs, reading each confirmation carefully before approving, and giving specific instructions instead of a broad one such as "review my emails and take whatever action is needed". You can read it at openai.com/index/prompt-injections. Anthropic's safety guidance for its Chrome extension (August 2026) goes further and tells people to strongly avoid using a browser agent on financial accounts, legal documents or contracts, health information, sensitive work accounts and sites holding other people's personal information.
That second piece of advice is about letting an agent roam logged-in sites. Uploading a copy of one document to a chat for a check is a smaller act, though you should still remove bank details first. For a building firm the practical rules are: the inbox agent gets the enquiries mailbox only, never the accounts mailbox; it never opens links or attachments from unknown senders; and a message that asks to change a supplier's bank details goes to a person, who phones a number already on file. That last request is the classic invoice-redirection fraud, and it is exactly what an obliging agent would process.
What can go wrong when agents take actions lists more failure types. Below is a filled-in permission sheet for the inbox example, which is worth keeping on one page next to the mailbox settings.
AGENT: Enquiries inbox sorter
OWNER: Office manager (deputy: operations director)
CAN READ: the shared enquiries mailbox only
CAN WRITE: labels and the Drafts folder
CANNOT: send, delete, forward, accept invitations, open links in
unknown senders' mail, reach the accounts mailbox, the
accounting system, the bank or any e-signature account
REVIEW: owner checks every label for the first 10 working days;
then a random 20 messages every Friday, 15 minutes
LOG: every miss recorded (date, what happened, cause, fix)
STOP RULE: any attempt to send, delete or forward, or two missed
urgent items in a week, drops the agent to read-only until
the owner has reviewed the instruction and the log
NEXT LEVEL: only after 20 working days with no missed urgent itemsReview and stop rules that keep it honest
The log matters more than it looks. A miss that is written down becomes an instruction fix; a miss that is shrugged off becomes a habit. Include near-misses, meaning drafts you would have been embarrassed to send, because they show where the next real miss will come from. Keep one person accountable by name, with a deputy for holidays. When the model behind the agent changes (vendors update them without asking), rerun a short shadow test on 30 recent messages, because behaviour that passed last month is not guaranteed this month. The vendors agree on the review step itself: OpenAI's help article on creating files with ChatGPT Work ends its instructions with "Review the file and ask for changes before sharing or relying on it".
A four-week trial that ends in a yes or no
- Week 0, before anything is switched on. Choose one task from the four above. Write the pass mark, including the second rule for the case that matters most. Collect 30 real examples, with the ugliest ones included. Write the permission sheet.
- Week 1, shadow run. The agent works and the person does the job as usual. Record every disagreement and sort the causes into three piles: unclear instruction, missing information, genuine mistake.
- Week 2, fix and repeat. Change the instruction for each pile, then run it on new items, not the same ones, so you are not marking its homework against answers it has already seen.
- Week 3, live at draft level. The person uses the agent's output and times the job against the old method.
- Week 4, decide. Continue, narrow the task or stop. Write the decision and the numbers beside the pass mark.
The money sum is short. Assume the tool costs $20 a month and the morning sort drops from 90 minutes to 40, with 22 working days in the month. That is 50 minutes times 22, about 18 hours. At an assumed internal cost of $25 an hour, that is around $450 of time against $20 of subscription. The comparison only holds while the checking continues, so count the checking minutes as part of the 40, not as a separate favour. If the numbers do not clear the bar, the cheap answer is to stay with a rule and a chatbot step and revisit when the tools or your documents change.
If a vendor's pitch survives the five questions, the live run on your ugliest ten documents and the permission sheet, it has earned a trial. If it only survives the brochure, it has earned another look next year.
Questions owners ask before trying an agent
Do I need a construction-specific agent, or will a general one do?
For inbox, document and meeting-note jobs a general assistant such as ChatGPT, Claude or Microsoft Copilot is enough to run a trial, because the work is reading and drafting. A construction platform only becomes necessary when the agent must work on live project data such as budgets, RFIs and drawings. Then ask the platform you already use for a demo on your own files before you move anything.
What does a first trial cost?
One subscription and some staff time. As of November 2026, ChatGPT Plus lists at $20 a month and includes ChatGPT Work. Microsoft business plans include Copilot Chat at no extra cost, and Microsoft 365 Copilot Business (the app is now called Microsoft Copilot) lists at $21 a user a month on annual billing. The larger cost is the hour or two a day someone spends checking the output during the first weeks.
Who is responsible if an agent sends something wrong?
You are. A message from your mailbox is yours commercially, whoever or whatever wrote it. How AI-drafted correspondence is treated under your contracts and insurance varies, so ask your adviser or insurer before the agent is allowed to send anything. Keeping the agent at draft level, with a named person who approves each message, avoids most of the question.
Will an agent replace my estimator or office manager?
Nothing in the jobs above removes a person. They change where the time goes, from sorting and retyping to checking. The jobs where agents still fail, such as take-offs, pricing and schedule changes, are the estimator's and site manager's own, and the office manager's checking is what keeps the inbox agent safe.
Further reads
- How to Prepare Your Small Business for AI Agents — Get your files and approvals ready before any agent trial.
- AI Agent Use Cases for Small Businesses: 10 Worked Examples — Ten worked agent examples outside construction, for comparison.
- AI Voice Agent for Contractors: What It Handles and What It Hands Off — What a voice agent should handle on calls and what it passes on.
- What Can AI Agents Do for Roofers Between Lead and Signed Contract? — Agent tasks between a roofing lead and a signed contract.
- AI Automation for Construction Companies: Daily Logs to Invoices — Daily logs to invoices as fixed automation, no agent needed.
- AI Solutions for Construction Companies: One Platform or Point Tools? — One platform or point tools when a vendor says agents are built in.
- AI Estimator for Contractors or a Human Estimator? How to Choose — Where an estimator still beats any AI estimating claim.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI's page on prompt injection and its help article on creating files with ChatGPT Work; Anthropic's file-upload help page; Anthropic and Microsoft guidance as recorded in the series fact sheet (checked October 2026); a construction-software vendor's homepage read in November 2026.