AI that understands your shop, not your CRM.

---

What’s actually broken on your floor and in your office

You have one estimator who knows how to quote a certain class of job. Anyone else who tries takes twice as long and quotes 8% high or 12% low. When he’s on vacation, that line of business effectively shuts down. When he retires in four years, the operational risk is sitting in his head.

Your sales engineer spends three hours a day on data entry — pasting between the CAD takeoff, the cost spreadsheet, the ERP, and the customer’s vendor portal — for orders that, once they’re in, you make at margin. Three hours a day, every day. You hired her to win deals. She’s a clerk.

Your ERP report on contribution margin by part hasn’t been updated since 2019. The fields it relies on aren’t being populated cleanly by the shop floor anymore. The CFO can’t get to a real margin number until the 12th of the month, because the data has to be cleaned by hand first. Nobody wants to own the cleanup project because the ERP vendor quoted $80K for the report rewrite.

An RFQ came in on Tuesday morning. The customer wants a response by Friday. Your estimator can get to it by Thursday afternoon, which means the response gets sent at 4:50pm and the customer reads it on Monday — alongside two competitors who got there first. You lose deals because you lose hours.

Three workflows we set up first

RFQ response acceleration

Before. RFQ arrives via email or vendor portal. Estimator opens it, pulls prior-quote history from the ERP, looks up current material costs in a separate spreadsheet, finds the customer’s tolerance and finish specs in a folder somewhere, drafts a quote package, sends to sales engineering for review. Average: 6–8 hours per RFQ. Worst case: it sits for two days waiting for the right person.

After. An agent monitors the RFQ inbox or portal. When a new RFQ arrives, it pulls prior quotes for that customer and similar parts, current material pricing from the source-of-truth spreadsheet or ERP, applicable tolerances and certifications. It drafts a quote package and routes it to the estimator inside Teams. The estimator reviews, adjusts the labor and margin calls, and approves. Average: 40–60 minutes per RFQ. The estimator stays the expert on the parts of the quote that need expert judgment. The agent does the file-pulling, the formatting, and the comparison work.

Tooling. Custom agent built in ACS Automate, with read access to your ERP via the API or a database view your ERP admin signs off on, plus read access to your costing spreadsheet or system. Quote output drops into your existing quote template — no new system for the customer.

Setup cost. One Custom Agent Sprint. Ongoing inside Operate; integration depth determines complexity for Strategic.

Tribal knowledge capture

Before. Your senior estimator knows that a particular finish on a particular alloy, at certain wall thicknesses, has a real-world yield problem that nobody else in the shop knows about, and it’s the reason your quotes on that part class come in 4% higher than the textbook would say. He learned it from a job in 2014. Nobody else has documented it. He retires in 36 months.

After. A structured interview agent runs 30 to 60-minute sessions with him, one part class or one process at a time. The agent asks: what do you check first, what would a junior get wrong, what’s the war story, what’s the rule of thumb, what’s the exception. Output: a structured knowledge base, searchable from inside Teams or your shop-floor portal, that the next estimator can query in real time. The estimator who retires gets to leave a record. The next estimator gets a head start.

Tooling. ACS Automate-hosted interview agent + retrieval system. The interviews are short and weekly so they fit around the work. Knowledge lives in your environment, indexed and retrievable.

Setup cost. First knowledge area is one Custom Agent Sprint. Subsequent ones are faster — interview cadence and time, not engineering work.

ERP question answering

Before. The CFO wants to know: which customers grew 20%+ last year, what’s the gross margin on each, and what’s the trend on rework hours by customer. The data is in Epicor (or Macola, Sage, ECi, JobBOSS, etc.). Getting the answer means a ticket to the IT-adjacent person who knows the ERP, who writes a query, who hands back a spreadsheet, three days later. The CFO sometimes just doesn’t ask.

After. A natural-language interface against your ERP — built with the security model your ERP admin signs off on — lets the CFO, the controller, and the sales manager ask questions in plain English. “What are the top ten customers by gross margin in the last twelve months.” “Show me rework hours by customer for parts shipped in Q1.” The agent translates the question into a query, runs it against an approved read replica, and returns the answer with a note about what data it pulled and when. No write access. No PII without role permission. Every query logged.

Tooling. ACS Automate query agent + your ERP’s API or a read replica. Compatible with most major SMB ERPs; we scope the connection during the Assessment.

Setup cost. Depends on the ERP. Plan on one Custom Agent Sprint for the first system. Reading from Epicor and similar mature ERPs is well-trodden; older systems need more scoping.

Where AI actually doesn’t help (yet)

A short, honest paragraph because every vendor in this space is making claims that don’t survive contact with a small shop.

Shop-floor vision AI — cameras watching machines and parts — is not ready for most small-to-mid manufacturers. The capital cost and the integration work outrun the return for fleets under 100 machines, in our experience. Generative AI does not replace CAD or simulation; the value is in the office, not at the workstation. Predictive maintenance is heavily overstated; for shops with fewer than 100 machines, the data volume is usually not sufficient to train models that beat a careful PM schedule. AI does not solve a broken process; if the RFQ process is broken because nobody owns it, an agent will produce faster bad quotes. We tell you these things before the contract, not after.

Compliance and customer requirements

ITAR exposure changes the architecture. If you handle defense-controlled technical data, your AI tooling has to stay inside the controlled environment, and the people who run it have to be U.S. persons. Microsoft 365 GCC High supports Copilot with ITAR coverage; commercial-tier Copilot does not. We deploy accordingly.

CMMC Level 2 — increasingly required for the DoD supply chain — has 110 controls; AI tooling intersects with about 30 of them, mostly around access control, audit, and data handling. We map the AI workflows against the controls during onboarding. If you’re working toward certification, we coordinate with your assessor and produce the evidence artifacts they’ll ask for.

IP confidentiality with customer-supplied drawings is the most common practical concern. Customers send you STEP files, PDFs, drawings under NDA. Those files need to be processable by your AI agents to do quoting work, and they need to stay out of any cloud service the customer hasn’t approved. We default to in-tenant processing — Copilot in your Microsoft tenant, or agents running in ACS Automate on your behalf — so the customer’s IP doesn’t transit a third-party LLM provider. Where a customer specifies a non-AI handling requirement, we route accordingly and document it.

We tell you what data leaves your tenant, and what doesn’t, in writing. If you can’t answer your largest customer’s “what AI do you use on our drawings” question with the file in front of you, you don’t have governance yet. We give you that file.

Compare the tiers

What could go wrong

Book an AI Readiness Assessment

Three weeks. Fixed fee. A real report and a 90-day roadmap. You keep it whether you hire us or not.

Book an AI Readiness Assessment · Or talk to us for 15 minutes — no pitch.