Owners hear a lot of numbers about AI and manufacturing, and some are wrong. The line "95% of AI pilots fail" is a misquote of an MIT report that measured something else (see "Why AI projects stall" below). Others come from a company selling the product being measured.
This page collects the figures we think an owner, GM, or estimator at a manufacturer should know this fall. Every number links to the page we read it on, and we checked each one on September 25, 2026. When the only source sells the product, we mark it "vendor data." When a survey is small, old, or a prediction, we say so next to the number. Each section ends with what the numbers mean for your company and one thing to do about them.
Workforce: open jobs, retirements, and hiring
Source: BLS JOLTS, Table 1. July 2026 is preliminary.
- The manufacturing openings rate was 4.4% in July 2026 (BLS JOLTS).
- Manufacturing workers quit at a rate of 1.4% in July 2026, below the 2.1% rate for all private employers (BLS JOLTS, Table 4).
- The median manufacturing worker was 43.9 years old in 2025 (BLS CPS).
- Of the 3.8 million manufacturing jobs to fill by 2033, about 2.8 million come from retirements and other exits, and about 0.76 million from growth (Deloitte and the Manufacturing Institute).
- The median time to fill a job across all industries is about 1.5 months (SHRM).
Openings are rising, headcount is flat, and people are not quitting faster than elsewhere. The gap comes from retirements and unfilled roles, so the extra capacity a manufacturer needs this year has to come from the people already on payroll, and some of what they know leaves with every retirement.
What to do: name the three people within five years of retirement whose judgment you could not replace, such as the estimator who knows which customers always change the drawing. Ask each to keep a folder of the next 20 odd cases they handle, with a line on why they decided what they did. Those cases are the raw material for teaching a new hire or setting up an AI.
ERP projects: cost, overruns, and time
Source: Panorama Consulting ERP Reports 2024 to 2026. Slightly plus significantly over budget, read from the reports' pie charts.
- The median ERP project cost $450,000 in both the 2024 and 2025 Panorama reports (Panorama 2025).
- In the 2026 report, 7.1% of ERP projects ran significantly over budget and 22.3% ran over schedule (Panorama 2026).
- The median ERP project took 9 months in the 2025 and 2026 reports, down from 15.5 months in 2024 (Panorama 2026).
- The top cause of ERP budget overruns in all three years was an unexpected need for more technology (Panorama 2026).
- In 2026, 60.3% of companies realized the compliance benefits they expected from ERP, and 56.3% realized the inventory benefits (Panorama 2026).
- In the 2021 cohort (average revenue $28.4 million), companies expected to spend $841,456 and actually spent $1,115,300 (Panorama 2021).
If you run an ERP, you have likely already approved a six-figure software project, and about 1 in 3 of those run over budget. The top cause is buying more technology after go-live, which is where an AI project on top of the ERP usually lands.
What to do: before you approve any AI project, ask the vendor to list in writing what it needs from your ERP (fields, exports, logins, add-on modules) that you do not already have, and price each item.
Quote turnaround and win rates
- In the same Paperless Parts survey of 400+ US buyers, 53% expected a quote within 24 hours in 2019, and only 6% would wait more than 3 days (vendor data) (Paperless Parts).
- Paperless Parts "early data" shows win rates over 90% for quotes sent in under 2 hours and under 5% for quotes sent after 5 or more days, though fast quotes are likely repeat customers or simple parts (vendor data) (Modern Machine Shop).
- Top Shops machining businesses had a median quote-to-book ratio of 70%, against 51% for other shops (CNCCookbook).
- Xometry reported 81,821 active buyers (up 20%) and $630 million in marketplace revenue (up 30%) for 2025 (Xometry 10-K).
- Protolabs says its software turns a 3D CAD file into a quote, and it serves about 48,000 customers a year (Protolabs 10-K).
At a 30% win rate, an estimator's time on 7 of every 10 quotes produces no order. The buyer's clock runs in hours while the shop's runs in days, and online instant-quote competitors are growing into that gap.
What to do: pull last quarter's RFQ log and work out two numbers for your shop, the median days from RFQ to quote and the share of quotes won. Those are your baseline against the 3 to 4 days and 30% above, and the numbers any quoting project has to move.
Compliance rules and deadlines
- Since October 1, 2025, manufactured products on FHWA projects must be made in the US, and the 55% component-cost test starts October 1, 2026. We found no delay notice as of September 25, 2026 (eCFR).
- Under BABA, iron and steel qualify only if every manufacturing process, from the initial melting stage through the application of coatings, happened in the United States (2 CFR 184).
- FHWA requires certification before steel goes into a project and endorses step certification, where each handler in the chain certifies its own step (FHWA Buy America Q&A).
- The Department of War suspended CMMC Phase 2 on July 13, 2026, about four months before its November 10 start date (Crowell & Moring).
- CMMC Phase 1 self-assessment, in force since November 10, 2025, still applies, along with DFARS 252.204-7012 and NIST SP 800-171 Rev 2 (Inside Government Contracts).
Each of these rules comes down to documents: an MTR tied to a heat number, a cert for every step, a cost breakdown for every component. Block, precast, and equipment makers selling into highway work need that 55% cost file for projects obligated from next week on.
What to do: if you sell into FHWA work, take your top-selling product and list each component with its cost and where it was made. If US-made components come to more than 55% of the total component cost, keep that sheet as the template. If not, you know before the customer asks. Defense suppliers should keep the CMMC Phase 1 self-assessment current while Phase 2 is paused.
Cyberattacks on manufacturers
Source: Verizon 2026 DBIR Manufacturing Snapshot. System intrusion is Verizon's category for breaches through hacking and malware, including ransomware.
- Manufacturing was the most attacked industry for the fifth year in a row, with 27.7% of incidents IBM X-Force handled (IBM X-Force 2026).
- An industrial-sector data breach cost $5.50 million on average in 2026, up from $5.00 million in 2025, against a global average of $4.99 million (IBM Cost of a Data Breach 2026).
- One in four malicious breaches in IBM's 2026 study was AI-enabled, at $6 million on average (IBM).
- Among 332 manufacturers hit by ransomware, the median ransom paid was $1.0 million, 51% paid, and mean recovery cost $1.3 million before the ransom (Sophos 2025).
- Dragos counted 747 ransomware incidents at manufacturers in Q2 2026, about 8 a day, and 45% of all industrial incidents were in North America (Dragos).
A stopped shop floor is expensive, and about half the manufacturers hit by ransomware in Sophos's survey paid. Every AI tool that reads your inbox or ERP is one more login an attacker can steal.
What to do: give any AI its own login (never a shared one), the fewest permissions that do the job, and a log of what it touched. Ask IT this week which of your current logins are shared, and start there.
AI model releases and prices
Source: OpenAI API pricing.
- The eight biggest labs made 52 notable model releases in the 52 weeks to September 25, 2026, one a week on average, and 32 were flagship models, by our count of lab announcements and changelogs, including Anthropic, OpenAI, Google, DeepSeek, and Mistral. Six of the 52 rest on weaker sources, and without them the count is 46.
- The pace picked up in the last six months, with 31 notable releases and 18 flagships between March 25 and September 25, 2026 (same count as above, Google changelog as one example).
- Anthropic's top Opus tier fell from $15 input and $75 output per million tokens to $4 and $20 with Opus 5.5 on September 22, 2026, a drop of about 73% in 10 months (Anthropic pricing).
- Google raised prices on its fast models, from $0.50 and $3 for Gemini 3 Flash to $1.50 and $9 for Gemini 3.5 Flash, and its promo price for 3.7 and 3.8 Flash doubles on January 1, 2027 (Gemini API pricing).
- On September 22, 2026, OpenAI halved prices for its cheaper GPT-6 Sol and Luna models (OpenAI API pricing).
- Manufacturing had the largest jump in AI adoption in the Census Bureau's new business survey series, up 159% (7.5 points) (Federal Reserve).
A model you pick today will likely be replaced within months, and its price can move either way.
What to do: treat AI usage like any other supply cost and review it monthly next to steel and power. Before you sign with any vendor, ask whether you can switch the underlying model without rebuilding the work, and get the answer in writing.
Why AI projects stall
- MIT NANDA's report says "95% of organizations are getting zero return" on $30 to 40 billion of generative AI spending. The report is labeled preliminary and rests on 52 interviews, 153 survey responses, and a review of 300+ public deployments (MIT NANDA).
- In the same report, projects built with outside partners reached deployment about 67% of the time, against about 33% for internal builds, a correlation MIT did not test for cause (MIT NANDA).
- Gartner predicted in 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data, risk controls, cost, and unclear value (Gartner).
- 41.43% of manufacturers surveyed give frontline workers no AI training (NAM Q2 2026).
- Manufacturers in Rockwell's 2026 survey said they use 43% of their data effectively, in a sample of 1,560 firms with $100 million or more in revenue (Rockwell Automation).
- 32% of US firms with 100 to 249 employees and 37% of firms with 250 or more use AI, as of May 2026 (US Census Bureau).
The barriers manufacturers name are data and people, the same two that sink ERP schedules. Our view: a project that starts with one document flow and the people who check the output has a better shot than one that starts with the software, because the test and the reviewers are what tell you it works.
What to do: pick one document flow (cert packets, order entry, RFQ intake), pull 50 past cases with the answer your team approved, and set them aside untouched. That folder becomes the test any AI has to pass, and walking the people who will check its work through it is their first training.
How companies buy software
- 67% of B2B buyers prefer to buy without talking to a sales rep, as of March 2026 (Gartner).
- 51% of software buyers now start research in an AI chatbot, up from 29% eleven months earlier, in a survey of 1,076 software buyers (review-site data) (G2).
- 63% of 1,862 tech buyers used AI while buying, and 94% of those fact-checked what it told them (review-site data) (TrustRadius).
- Among software buyers who regret a purchase, the top vendor-side cause is a poor handoff from sales to implementation (43%) (Gartner Digital Markets).
Buyers do most of their homework before they call anyone, and many now start with an AI chatbot and then check what it said. When your company buys AI, the regret usually starts after the salesperson hands you off.
What to do: before you sign for any AI project, ask for a trial on your own documents, and ask for the name of the person who will run the work after the contract is signed. Bring the people who will use it into the trial, since a typical decision already involves 13 people inside the company.
How to cite this page
Cite it as Second Shift, "Manufacturing and AI in 2026: The Numbers That Matter," Movement Global, September 25, 2026, and link to the original source next to each number you use.
Questions people ask
How many AI models came out in the past year?
By our count, the eight biggest labs made 52 notable model releases between September 25, 2025 and September 25, 2026, and 32 of them were flagship models. Six rest on weaker sources, so the count on strong sources alone is 46.
Do 95% of AI pilots fail?
That is a misquote. MIT NANDA's preliminary report says "95% of organizations are getting zero return" on generative AI spending, based on 153 survey responses and 52 interviews (MIT NANDA).
What does an ERP project cost a manufacturer?
The median ERP project cost $450,000 in Panorama Consulting's 2024 and 2025 reports, and about 1 in 3 ran over budget (Panorama 2025).
Terms in this piece
- Model
- the AI software a lab trains on huge amounts of text, code, and images so it can read and write, like GPT or Claude.
- Flagship model
- a lab's most capable generally available model at the time it ships.
- Token
- a small chunk of text, roughly a piece of a word. AI companies bill by the token, and prices are quoted per million tokens.
- Proof of concept
- a small first test of an AI project, run before anyone commits to a full rollout.
- JOLTS
- the Bureau of Labor Statistics monthly survey of job openings, hires, and quits.
- BABA
- the Build America, Buy America Act, which sets domestic content rules for federally funded infrastructure.
- CMMC
- the Defense Department's cybersecurity certification program for suppliers that handle controlled information.
- Generative AI
- AI that writes text, code, or images, such as ChatGPT or Claude, as opposed to older AI that only sorts or predicts numbers.
