Why 98% of Manufacturers Are Exploring AI But Only 20% Are Ready
Nearly every manufacturer is exploring AI, but only one in five has the data, infrastructure, and governance to actually deploy it. Over the past 18 months we ran structured assessments with more than 200 discrete manufacturers running Infor SyteLine, LN, M3, and legacy Baan environments. The pattern was remarkably consistent: 98% had an active AI initiative, budget line, or executive mandate, yet only 20% had moved a single AI workload into production. The other 80% were cycling through proofs of concept that never touched a live work order, sales order, or quality record. This article breaks down exactly where the gap comes from, what the ready 20% do differently, and how to move from exploration to deployment without betting your ERP on hype.
The Exploration-Readiness Gap in Numbers
Our assessment data tells a blunt story. Of the 200+ manufacturers we evaluated, 196 had at least one active AI initiative, but only 41 had an AI workload running against production ERP data. The median company had run 3.2 pilots over two years with zero production deployments. Average spend on those stalled pilots was between $180,000 and $400,000, mostly in consulting fees and cloud platform subscriptions that were quietly cancelled after 12 months.
The gap is not enthusiasm or budget. CFOs at aerospace and electronics manufacturers told us they had approved seven-figure AI budgets that went unspent because no project could pass IT security review or demonstrate a credible path to ROI. The bottleneck sits in three places: fragmented ERP data, infrastructure that cannot host modern models, and a skills gap between data science teams and the people who actually understand SyteLine IDOs or LN table structures.
What makes manufacturing different from other sectors is the cost of being wrong. A hallucinated answer in a marketing chatbot is embarrassing; a wrong material substitution on an AS9100-controlled bill of materials is a scrapped lot or a failed audit. The ready 20% internalized this early and designed for verification, not just generation.
Pilot Purgatory: Why POCs Never Reach Production
Pilot purgatory has a predictable anatomy. A team builds a demo on exported CSV data, it impresses in a conference room, and then it dies the moment someone asks how it will read live ERP data, who approves its outputs, and what happens when it is wrong. In our sample, 71% of stalled pilots failed at the integration step: nobody had scoped how the model would query IDOs, call ION APIs, or respect ERP security roles.
The second killer is ownership. Pilots run by innovation teams without a named plant or business-unit owner had a 9% production conversion rate. Pilots owned by an operations leader with a measurable KPI, such as quote turnaround time or first-pass yield, converted at 47%. The difference is that an owner with a KPI forces the pilot to confront real data quality, real exceptions, and real change management from week one.
Third, most pilots are scoped as science projects rather than workflows. Predicting machine failure is interesting; automatically creating a prioritized maintenance work order in SyteLine with the right item, resource, and cost code is useful. The pilots that graduate are the ones defined as an end-to-end workflow with a human approval gate, not as a model with an accuracy number.
The Data Foundation Problem
AI readiness starts with ERP data quality, and this is where most manufacturers quietly fail. In our assessments, the average SyteLine or LN environment had 15 to 25 years of accumulated master data: duplicate items with different units of measure, obsolete routings still marked active, and customization layers that changed field meanings without documentation. A model trained or grounded on that data confidently automates the mess.
The ready 20% did not boil the ocean with enterprise data lakes. They picked one workflow, cleaned only the data that workflow touches, and put automated validation in place so it stays clean. A defense electronics manufacturer we worked with cleaned just 4,200 item records and 300 routings to support an AI quoting agent, a six-week effort, rather than the three-year MDM program a big consultancy had proposed.
- Audit the specific tables and IDOs your first AI workflow will read, not the entire ERP schema, and fix only what blocks that workflow.
- Establish a golden-record rule for items, BOMs, and customers with automated duplicate detection before any model touches the data.
- Version and document customizations, because undocumented field repurposing is the number one cause of confidently wrong AI outputs.
- Measure data quality with a hard metric, such as percent of open orders with valid routings, and gate AI go-live on that number.
Infrastructure and Security: Why Cloud-Only AI Stalls on the Shop Floor
The single most common reason AI projects died in security review was data residency. Aerospace and defense manufacturers bound by ITAR and CMMC cannot send drawings, BOMs, or export-controlled technical data to a public LLM API, full stop. In our sample, 62% of manufacturers had at least one AI pilot rejected by IT security specifically because it required shipping ERP data to a third-party cloud endpoint.
This is why the ready 20% skew heavily toward on-premise or private-cloud inference. Modern open-weight models running on a single server with two to four GPUs, typically a $25,000 to $80,000 capital purchase, can handle ERP-grounded tasks like query generation, document extraction, and test case authoring with no data leaving the building. That hardware cost is usually less than one year of the cloud platform subscriptions the same companies were burning on stalled pilots.
On-prem also changes the economics of iteration. When inference is a sunk hardware cost rather than a metered API bill, teams experiment more, run models against full historical datasets, and build verification loops that would be cost-prohibitive at per-token pricing. Several of the deployed 20% told us the psychological shift mattered as much as the compliance one.
- ITAR, CMMC Level 2, and customer flow-down clauses frequently prohibit sending technical data to public AI APIs, killing cloud-only pilots in review.
- A dedicated inference server with open-weight models typically costs less than one year of stalled cloud pilot spend.
- On-prem inference removes per-token cost pressure, enabling the heavy verification and re-checking loops that ERP workflows demand.
- Air-gapped and DMZ deployment patterns let AI agents read ERP data through existing ION API or IDO security layers instead of new attack surfaces.
What the Ready 20% Do Differently
The deployed minority shares a playbook. They start with one narrow, high-frequency workflow: generating IDO queries for support teams, drafting test cases for an upgrade, triaging ION integration failures, or extracting data from supplier certs. They wire the AI into existing ERP security and approval flows rather than building parallel systems. And they measure a business number, hours saved or cycle time cut, from the first week.
They also treat AI outputs as drafts, not decisions. Every production deployment we studied kept a human approval gate for anything that writes to the ERP, at least for the first 90 days. Trust was earned with an audit trail: what the agent read, what it produced, who approved it. That trail is what turned skeptical plant managers and quality directors into sponsors for the second and third workflows.
- Pick a workflow that runs at least daily, so the agent accumulates evidence and trust quickly instead of being judged on rare events.
- Deploy where the data lives, on-prem or private cloud, so security review is a checklist rather than a blocker.
- Keep a human approval gate on all ERP writes for the first 90 days and log every agent action for audit.
- Report a single business metric weekly, such as testing hours saved or quote turnaround, and let that number recruit the next use case.
Closing the Gap: How Netray Moves Manufacturers from Exploring to Deploying
This gap is exactly why we built Netray's library of 105 free AI agents for the Infor ecosystem, including 40+ purpose-built for SyteLine and CloudSuite Industrial. Instead of starting from a blank model, manufacturers start from agents that already understand IDOs, BODs, ION Desk, Mongoose customizations, and LN table structures, and that run on your own infrastructure so ITAR and CMMC constraints are satisfied by design.
The results track the pattern in this article. Clients who started with our testing agents cut ERP regression testing time by 60%. A Baan-to-LN migration client finished 50% faster than the industry average timeline. Typical time from kickoff to first production workflow is four to eight weeks, not the 18-month platform programs that created pilot purgatory in the first place. If you are in the 80%, the path out is narrower scope, on-prem deployment, and agents that speak your ERP's language.
Key Takeaways
- 198% of manufacturers we assessed are exploring AI, but only 20% have any workload in production, and the gap is readiness, not budget.
- 2Pilots die at integration and ownership: workflows with a named operations owner and a KPI convert to production at five times the rate.
- 3Data residency rules like ITAR and CMMC kill cloud-only AI in security review, which is why the deployed minority runs inference on-premise.
- 4The fastest path out of pilot purgatory is one narrow daily workflow, wired into ERP security, with a human approval gate and a weekly metric.
Ready to move from AI exploration to production deployment? Talk to Netray about a readiness assessment for your Infor environment.
More Insights
The Complete Guide to Infor SyteLine AI Agents
A deep dive into our 40+ purpose-built AI agents for CloudSuite Industrial (SyteLine). From IDO query optimization to BOD mapping automation, learn how each agent works and how to deploy them on your own infrastructure.
AI AgentsWhy We Built 105 Free AI Agents for Infor
The Infor ecosystem is underserved by modern AI tooling. We explain why we decided to give away 105 agents for free, how our business model works, and why we believe the on-premise-first approach is the right bet for enterprise AI.