By Sagar Shankaran, Founder of CallSphere
Controlled drawings cannot go to a cloud AI tool. How on-site AI lets a defense weld shop quote 34 packets in September instead of no-bidding eleven of them.
Key takeaways
Thirty-four request-for-quote packets between the second week of August and the last week of September. That is a normal defense-season load for a mid-size weld shop with one estimator, and it happens every year for the same reason: the federal fiscal year closes on 30 September, and the primes push their subcontract awards before it does.
Most of those packets arrive with a cover page carrying a distribution statement and a marking that says the drawings inside are controlled unclassified information. And that marking is the reason your estimator is still doing weld-inch takeoffs with a scale and a highlighter in 2026, while the machine shop down the road feeds its commercial drawings to an AI assistant and quotes in a morning.
Work the hours honestly. A 60-page packet with structural weldments, a bill of material, a weld map and a finish spec takes a good estimator three to four hours: identify grades and shapes, count weld inches by size and position, work out fit-up and fixturing, price the outside services, check the welding procedure specifications you already have qualified against what the drawing calls for, then build the quote in your system.
Thirty-four packets at three and a half hours is 119 hours. Your estimator has around 160 productive hours in a month and also has commercial work, change order pricing, and a purchasing agent asking about plate lead times. So eleven packets get a polite no-bid, and eleven chances at a $30,000-to-$50,000 weldment order go to somebody else. That is the real cost of the marking on page one, and it recurs every August.
On-premises AI means the model runs on a computer you own, inside your own building, so the drawing, the welding procedure and the customer's part number never cross your property line.
If you hold a Department of Defense subcontract, your flow-down clauses almost certainly include DFARS 252.204-7012, which points at the 110 security controls in NIST SP 800-171 and now, under CMMC 2.0, at a Level 2 assessment. If any of the drawings describe a defense article, the export rules apply on top: putting a controlled technical drawing somewhere a foreign national can read it is treated as an export, whether or not anybody in another country actually opened it.
That is why the honest answer inside a lot of shops has been "don't put it in any AI tool." It is not paranoia. If you upload a controlled weldment drawing to a general consumer service, you cannot easily tell an assessor where the file went, who could reach it, or whether it was retained. Explaining that after the fact is a bad afternoon.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent in your browser — 60 seconds, no signup.
The same problem shows up without a single defense job. Look at the nondisclosure agreements in your customer file: an equipment OEM that treats its frame geometry as proprietary, a food processor whose stainless washdown skid design is its competitive edge, a medical device company that will audit you for it. Those contracts do not care whether the recipient is a person or a service.
flowchart TD
A["Prime sends RFQ packet marked CUI"] --> E["Shop computer in the server closet reads it on site"]
B["Shop drawings, DXF and STEP files"] --> E
C["Qualified WPS and welder qualification records"] --> E
D["Twelve years of closed job costs from the ERP"] --> E
E --> F["Draft takeoff: weld inches by size, plate weight, ops list"]
E --> G["Access log written for the CMMC assessor"]
F --> H["Estimator prices it and only the quote leaves the building"]
Two things converged this year. Chips built for running AI locally — Qualcomm's Dragonwing-class processors are the ones people in industrial settings talk about — got good enough that useful models run on a single box on a shop network instead of in somebody's data center. And large enterprises stopped treating on-premises as a legacy compromise: Cisco's rollout of a personal AI assistant to roughly 90,000 employees explicitly emphasizes on-premises operation for control and data protection.
For a fabricator, the practical version is unglamorous. A tower or small rack unit sits with your ERP server. It reads PDFs, DXF and STEP files off your own drive. Nothing goes out over the wire. High-volume local work runs roughly 90% cheaper than sending the same work to a cloud service, which matters when you are running it over hundreds of drawing sheets a week rather than a handful.
Two years ago the local option meant a small, weak model that could not read a weld symbol or hold a whole packet in view at once. That is the part that actually changed. The local box will now take the entire 60-page packet at once and answer questions about it.
7:05 a.m. Your estimator drops the packet into a watched folder on the shop server. By 7:30 there is a draft takeoff sheet: material list by grade and shape with weights, weld inches split by size and process, part counts against each detail, a flag on every note calling for a procedure you have not qualified, and a page reference beside every line so he can check it against the drawing.
8:15 a.m. He works down the flagged items. The drawing calls for a groove weld in a position none of your current welder qualification records cover, which means either a new qualification test or an outside shop. That single flag is worth more than the takeoff, because that is the item that kills margin when it is discovered in week three.
9:40 a.m. He asks the box a question he could never ask a cloud tool with these drawings open: what did we actually run on the last three jobs of this type, hours and pounds, and where did we lose. It answers from your own closed job costs, and it shows the job numbers.
10:20 a.m. The quote goes out. Nothing left the building except the price, the lead time and the terms. Your access log has one line saying who opened the packet and what was done with it, which is what your assessor asks for.
Stated assumptions: 34 packets in the August-September push, estimator fully burdened at $78 per hour, first-pass takeoff on the local box takes 25 minutes and estimator review takes 45 minutes, historical hit rate 22%, average awarded weldment order $38,000 with a 26% gross margin.
Still reading? Stop comparing — try CallSphere live.
CallSphere ships complete AI voice agents per industry — 14 tools for healthcare, 10 agents for real estate, 4 specialists for salons. See how it actually handles a call before you book a demo.
| Line | Today | With the box on site |
|---|---|---|
| Estimator hours per packet | 3.5 | 1.2 |
| Packets actually quoted | 23 | 34 |
| Estimator hours consumed | 81 | 41 |
| Additional quotes submitted | — | 11 |
| Additional awards at 22% | — | 2.4 jobs |
| Additional gross margin at $38,000 × 26% | — | $23,700 |
| Hardware, one time | — | $9,400 |
| Power, backup and support per year | — | $2,160 |
One September pays for the box, and you still have it in October. The number to watch is not the hardware cost, it is whether the takeoff is accurate enough that the review really is 45 minutes. Measure that on twenty packets you have already quoted, where you know the answer, before you count a dollar of it.
Keeping the drawings in the building satisfies one obligation. It does not satisfy the other 109 controls, it does not write your system security plan, and it does not make you assessment-ready. If shop floor computers still share one login and the ERP password is on a sticky note under the burn table monitor, the AI question is not your biggest one.
It also does not read everything. Hand-marked redlines faxed by a field superintendent, thirty-year-old scanned prints with dimensions written over the linework, and weld symbols drawn wrong by a junior detailer all still need a person. Expect the box to be strong on typed notes and bills of material and weak on somebody's ballpoint.
And it does not price the job. Weight and weld inches are arithmetic. Whether you can hold a flatness tolerance on a 14-foot weldment without a stress relief cycle is experience, and it lives in your shop foreman, not on a server.
No, but you need someone who owns it. In practice this is the same person who already owns your ERP server and your backups, whether that is an office manager with a support contract or a local managed services provider. Budget the support line, not a headcount.
Yes, and most shops running both do exactly that. The trap is the human sorting step. Put the controlled work in a folder structure that makes the wrong move hard — separate drive, separate login, drawings never mixed — because the failure is always somebody dragging a file into the wrong window at 4:50 on a Friday.
Two. Your job cost history and quoted margins are the most sensitive numbers in the company, and any nondisclosure agreement you signed with an OEM covers their drawings whether or not the government is involved. Read one of them and see whether it permits a third-party service.
The 2 August 2026 obligations reach US companies whose systems affect users in the EU. A takeoff assistant reading your own drawings inside your own building is a very different animal from a system making decisions about EU citizens, but if you ship into the EU it is worth twenty minutes with your attorney rather than a guess from a vendor.
Quoting is only half the phone problem in August. The other half is the calls that come in while your estimator has his head in a packet — expeditors chasing ship dates, a GC wanting a budget number, a walk-in asking whether you do aluminum. CallSphere builds AI voice and chat agents that answer those lines around the clock, capture the job details, and book the callback, so the interruptions stop landing on the one person who can price the work.
Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
Supplement brands get adverse events reported by ticket, not by form. On-premises AI reads all 3,400 a month without that text ever leaving the building.
Concealed-damage claims cost a warehouse manager half a Friday. On-site AI searches your own footage and scan trail without any data leaving the property.
Casino surveillance footage cannot leave the property. On-premises AI in 2026 cuts a 90-minute disputed handpay review to nine, with no video sent to a vendor.
Custody orders, health plans, subsidy files and classroom video can stay in the building. What on-premises AI unlocks for a child care center in 2026.
How commercial GCs search restricted SSI and CUI drawing sets using AI that runs on hardware in their own server room, with a worked cost example for one job.
Where session audio physically goes, what test publisher and county contracts actually forbid, and what an on-premises setup costs a 24-clinician practice.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco
Watch how CallSphere handles real customer calls, schedules appointments, and processes payments — live.
Try Live DemoBook a DemoCalculate Your ROI