It’s ten past six and the workshop has gone quiet. The estimator is still at his desk, and he has just clicked send on a quote he is genuinely proud of. Sharp price, honest lead time, the sort of number that wins work and still leaves something on the job. Then he opens his inbox and finds a two-line reply the customer sent at twenty to one that afternoon: thanks mate, we’ve gone with someone else this time.

He reads it twice. The price wasn’t the problem. He knows his numbers, and his numbers were good. What beat him was that the other shop got a quote back before this one had lunch, and by the time his landed the customer had already signed.

Not the price. The clock.

This is the part owners find hardest to hear, because it stings more than losing on price does. Losing on price at least feels like a market. Losing on turnaround feels like leaving money on the bench. And it is the more common way to lose. Industry data compiled in our 2026 manufacturing brief puts it at roughly 86% of manufacturers losing deals because their quoting is too slow. Not too dear. Too slow.

Think about what the estimator’s day actually looked like. The RFQ pack came in as an email with a PDF and a couple of drawings attached. Before he could price a single line, he had to read the pack, pull the part numbers, key them in, cross-check the drawing against the bill of materials, find last year’s price on that grade of steel, and chase a supplier for the one material he wasn’t sure of. Hours of it. Careful, unglamorous, error-prone hours. And a manual re-key isn’t free even when it looks finished: data-entry error rates run around 1 to 3%, and each wrong figure can cost roughly $200 downstream once it’s baked into a quote and a purchase order (Artsyl, 2025, on APQC benchmarks). By the time he reached the part he was actually hired for, the light outside had gone.

What the estimator was hired for

Here is the quiet waste at the centre of it. You did not hire an estimator to type. You hired him because he can look at a job and know things a spreadsheet can’t: that this tolerance will eat two extra setups, that this customer always adds a revision, that the shop is tight in week three so the lead time needs padding. That judgement is the product. It is the difference between a quote that wins and holds margin and one that wins and bleeds.

And it is the thing that gets the least of his day. Estimators spend something like 60% of their time on data extraction and verification rather than pricing (StartProto, 2025). The pricing judgement — the reason he’s in the chair — gets squeezed into whatever’s left, often the last hour, often after the job’s already gone cold. A quote that takes three to five days to compile by hand can carry maybe twenty minutes of genuine pricing thought (StartProto, 2025). The rest is data entry wearing an estimator’s job title.

None of this happens in a comfortable market. Australian manufacturing contracted about 2.6% over the year and its labour productivity fell 3.7% across the decade to 2023-24 (Ai Group, Performance and Outlook Report 2025). Trades and technician roles are the hardest in the economy to fill, with a fill rate of just 54.3%, the lowest of any major group (Jobs and Skills Australia, 2025). When more than 90% of Australian makers are small and medium businesses (AMGC, 2025), the estimator who is drowning in data entry is often the only estimator. There is no bench to bring on. There is just him, and the clock, and the quote that goes out too late.

The same shape, three more times

Once you see it in quoting, you see it everywhere the office meets a document. An order arrives as an unstructured email and someone re-keys it into the system by hand. A supplier invoice turns up and someone matches it, line by line, against the purchase order and the goods received. Manual invoice processing costs somewhere between $12 and $30 each; automated, it’s more like $1 to $5 (APQC, 2024, via Artsyl). Same shape as the quote: high-volume, document-in, human reading and typing, margin leaking out the side.

The AI everyone is selling you is the wrong one

Now, if you have been to a trade show this year, none of this is what the AI vendors wanted to talk about. They wanted to talk about the shop floor. The machine that senses its own bearings wearing and warns you before it breaks. Predictive maintenance is real and the savings are real, on the order of 8 to 12% over a preventive schedule (US DOE, via NIST 2024 guidance), and unplanned downtime is genuinely brutal, with a median cost cited around US$125,000 an hour (ABB Value of Reliability survey, 2023).

But for most 30-to-80-person Australian shops it is the wrong place to start, and the reason is boringly practical. Predictive maintenance runs on a stream of machine telemetry, and that data lives in a read-mostly shop-floor tier that most shops this size have not instrumented yet. There are no sensors feeding it. Whether AI pays you back is decided less by how clever the model is than by whether it can actually reach your data and do something with it. The most-marketed use case sits on the data you don’t have. The use case that already costs you deals sits on the data you do.

Where your data actually lives

Because here is the thing about the quoting problem: the data it needs is already somewhere an AI agent can reach. Your price list, your job history, your customers, your invoices live in cloud systems built for other software to talk to them. The inventory and MRP layer (Katana, Cin7 Core, Unleashed, MRPeasy, SimPRO, Pronto Xi), the accounting back-office (Xero, QuickBooks Online, MYOB) and the CRM all expose proper read-and-write connections today. Xero even ships an official MCP server with live tools for invoices and bills, which makes it about the cleanest integration in the whole stack. The integration path runs through your cloud ERP and your ledger, not through your design tools.

That last distinction is where honesty earns its keep. If your drawings live in SolidWorks PDM, that vault is genuinely closed: it exposes only a local desktop connection on the machine the software runs on, with no cloud endpoint an outside agent can call. So nobody should promise you a magic pipe into your CAD vault, and we won’t. The quoting win doesn’t need one. It reads the RFQ and the drawing pack the way they already arrive, as an email and a PDF, and writes the priced result back into the ERP and ledger that already hold your prices and history.

What actually changes

So picture the same 6pm, rebuilt. The RFQ lands at nine. An assistant reads the pack, pulls the part numbers and quantities, matches them against your price list and last year’s jobs, and hands the estimator a structured draft quote with every line already populated. It does not price the job. It does the extraction that used to eat his day, and it cites where each figure came from so he can check it in seconds rather than rebuild it from scratch. He spends his morning doing the thing he’s good at: reading the job, adjusting for the setups and the revisions and the tight week, deciding the number. He signs it and it goes out before lunch, while the job is still winnable.

The rule underneath this is not negotiable, so let me be plain: the machine extracts and drafts, a person prices and sends. Nothing is auto-quoted. And the control isn’t only “a human signs off,” because the real danger isn’t the obvious error, it’s the confident-looking quote that gets rubber-stamped by a busy estimator. So low-confidence extractions get flagged, every spec and price is cited back to its source document, and a sample of signed-off work gets a second-pass check. Nothing silently finalises a quote or posts to your ledger.

The honest ledger

Two things I won’t dress up. First, RMAI is new to manufacturing, and I’m not going to hand you a before-and-after from a shop like yours that doesn’t exist. The proof I can point to is other people’s. Komatsu Australia had a three-person team drowning in more than 52,000 invoices a year; built on Power Automate and AI Builder, they were live in four weeks and saved around 300 hours a year on a single supplier’s invoices (Microsoft / Komatsu case study). That’s third-party, and the same read-and-write ERP-and-ledger mechanism the quoting win uses. The wider evidence is sobering and worth respecting: MIT found about 95% of corporate AI pilots delivered no measurable profit impact, and the ones that worked were process-first and back-office, with bought tools succeeding around 67% of the time against roughly a third for internal builds (The GenAI Divide, 2025).

Second, if your systems are disconnected and your data is messy, integration can quietly become the real budget. That’s exactly why the work starts with a free 30-minute discovery call, not a build. On the call we walk one real path through your own quotes and invoices and name the highest-value fix, and any build gets quoted fixed-scope after that, on your numbers, so it lands inside a $10k to $60k band and ships in three to six weeks rather than ballooning later. Payback across that band tends to run somewhere around two to twelve months (illustrative, recalibrated to your own numbers, not a promise).

Where to start

You don’t need a transformation programme to find out whether this is real. You need one honest look at where your quotes actually come from, how long they take, and which of your systems an agent can genuinely reach.

Two ways in, both low-risk:

• Read the map first. We’ve written a plain-English brief on the Australian manufacturing software landscape — which of your systems can actually be reached, by what mechanism, and what “closed” really costs you. → https://realmindsai.com.au/guides/manufacturing/
• Book a free 30-minute discovery call. We’ll take one real path — an RFQ to a priced quote — name the tier your core system sits in, and show you the highest-value place to close the gap, before anyone promises a build. → https://outlook.office.com/book/[email protected]/?ismsaljsauthenabled

Somewhere in your shop tomorrow morning, an RFQ is going to land in an inbox. The only question worth asking is whether your best estimator gets to spend the day pricing it, or typing it in while the customer signs somewhere else at lunchtime.