Guide
What can AI do for my shop?
Every software pitch now has AI in it, and most owners we talk to are somewhere between curious and tired of hearing about it. This page is what we tell them on a first call: what the technology does well today in a construction or manufacturing business, where it can't be trusted, what it costs, and how to start without betting anything that matters.
The useful work is mostly paperwork
The AI that earns money in a shop today is unglamorous. It reads, writes, sorts, and summarizes. That covers more of your week than it sounds like.
- Reading documents. Supplier invoices, receipts, packing slips, purchase orders. AI reads the PDF or the photo and types the numbers into your system, so a person stops re-keying them. This is the most proven category and usually the right first automation. We run parsers like this in production; the well-built ones flag anything they're unsure about instead of guessing.
- Drafting. First drafts of quote cover letters, follow-up emails, job postings, site reports, warranty responses. The person who used to write it from scratch now edits and sends. The time saved is real, and the review step is what keeps the quality yours.
- Summarizing and finding. A forty-message email thread about one change order, reduced to what was decided. A voice memo from a site visit, transcribed and turned into notes. A question answered from your own manuals and old quotes instead of someone's memory.
- Sorting and routing. Incoming leads tagged by trade and urgency. Expenses categorized. Warranty claims split from sales inquiries before anyone reads the inbox.
Where it can't be trusted
Exact numbers.
AI reads documents well and does arithmetic badly, and when it's wrong it is confidently wrong. Working systems split the job: AI extracts, ordinary software checks the totals, a person approves anything that doesn't reconcile. Never let a model's output flow into your books or onto a customer document without one of those two checks.
Gaps.
When information is missing, a model tends to fill the hole with something plausible rather than say so. A quote drafted from thin inputs will contain invented specifics. The fix is the same review step as above, plus building automations that are allowed to say "not sure, look at this one."
Judgment.
Pricing a job, choosing a sub, anything touching safety, hiring. AI can lay out the information; the call is yours. Anyone selling you software that makes these decisions for you is selling past what the technology does.
A pattern worth noticing: everything on the good list has a person or a program checking the output, and everything on this list is what happens without one. That check costs far less than the re-keying it replaces.
Start small and keep score
Use the general chat tools yourself for a month. A business plan runs roughly $20–40 per person per month. Draft emails with it, summarize documents, ask it to explain a contract clause. You'll learn its character faster by using it than by reading about it, including the confident-but-wrong part.
Pick one paperwork task to automate: high volume, low stakes, already checked by a person downstream. Supplier invoices are the classic. Leave quoting and anything customer-facing for later.
Count hours before and after. If nobody can say what the automation saved, you can't decide whether to build the next one.
Ask where your data goes
A document you send to an AI service leaves your building, and the question to ask any vendor is the plain one: what leaves our systems, where is it stored, and is it used to train models? The big providers publish different answers for consumer chat plans and business plans, and the terms change, so read the ones attached to the plan you're buying rather than trusting a summary (including this one).
For most shop paperwork, a mainstream business plan is a reasonable answer. When it isn't (customer lists, anything under an NDA), the sensitive part can stay out of the prompt, or the whole workload can run on infrastructure in Canada. We build it that way for clients who need it; it costs more, and for some data it's the right call.
What it costs
Three tiers, in the order you should spend them:
Chat subscriptions.
Roughly $20–40 per person per month. This is where the first wins come from, and for a small office it may be the whole program for a year.
A built automation.
Our fixed-scope builds run $5,000–25,000, and document automations tend toward the lower end. You own the result, and the ongoing model usage typically costs cents per document.
The audit first, if you're not sure.
$1,500–3,000 maps where your hours go and prices what each fix would save, whether the fix is AI, ordinary software, or a subscription you already have. Some audits come back recommending the $30 chat plan and nothing else for six months.
The order matters because each tier tells you whether the next one is worth it. A shop that skips to a build without knowing where its hours go is guessing with a bigger number. The audit is our automation audit, and it treats "keep the chat plan, skip the build" as a valid finding.