AI in Manufacturing: From Pilot to Production, Fast
Most manufacturers do not have an AI problem. They have a pilot problem.
A forecasting model impresses in a demo. A chatbot answers stock questions in a sandbox. Then six months pass and nothing has changed on the shop floor. Planners still chase shortages by email, buyers still expedite at premium rates, and sales still promises dates nobody has checked.
The reason is rarely the AI itself. It is the data underneath it, the process around it and the missing route from “interesting” to “trusted”. This article explains why manufacturing AI pilots stall, what production-ready AI actually requires and a practical 90-day path to get there.
Why most manufacturing AI pilots stall
Pilots stall for operational reasons, not technical ones. The same five patterns appear again and again.
- The pilot runs on an extract, not on live operations. It works on last quarter’s spreadsheet. When it meets real orders, real stock and real schedules, it needs an integration project nobody budgeted for.
- Nobody on the shop floor owns it. An innovation or IT team builds it. Planners, buyers and supervisors never asked for it, so they never change how they work.
- It produces insight but not action. The model warns that a shortage is coming. The planner still has to open three systems to raise a purchase order or move a work order.
- It cannot explain itself. When the quality manager asks why a recommendation was made, there is no clear answer and no audit trail. In regulated manufacturing, that ends the conversation.
- Success is measured by model accuracy, not business results. A forecast that is 92% accurate means nothing to a COO. Fewer late orders, lower expediting costs and less planner time do.
The real bottleneck is connected data, not the model
AI can only answer an operational question if it can see every record that question depends on. In most manufacturers, those records live in different places: demand in the CRM, stock in the ERP, the production schedule in a spreadsheet and quality results on paper.
Every new AI use case then starts with an integration project. That is where the months go.
Look at what a few everyday questions actually require:
| Question a planner or salesperson asks | Records the AI needs to answer it |
| Can we ship this order by Friday? | Sales order, available and allocated stock, open work orders, inbound purchase orders |
| Are we clear to build this work order? | Bill of materials, component stock, quality holds, supplier delivery dates |
| Which customers are affected by this supplier delay? | Purchase order, linked demand, sales orders, customer accounts |
| Which batches went to which customers? | Lot and serial records, production history, shipments, customer records |
| What is the real margin on this order? | Quote, material and labour cost, landed cost, invoice |
Every answer crosses at least two departments. When customer, order, inventory, production, purchasing and finance data share one data model, AI can read across all of them from day one. When they don’t, each question is its own integration project.
What production-ready AI looks like
Production-ready AI is AI your operations team trusts enough to act on every day. In practice, that means five things.
- Grounded in live records. Recommendations come from today’s orders, stock and schedules, not from a copy taken last week.
- Explainable. Every recommendation shows the records behind it: which order, which component, which supplier date. A planner can check the reasoning in seconds.
- Permission-aware. The AI sees and does only what the person using it is allowed to see and do. A sales rep should not get supplier pricing because they asked a clever question.
- Human approval for consequential actions. Drafting a purchase order or proposing a schedule change is helpful. Releasing it should still need a person to approve it, with a clear record of who did.
- Measured on a business outcome. Each use case has one agreed metric with a baseline, such as late orders, expediting spend, stockouts or planner hours.
If a pilot is missing any of these, it is unlikely to survive contact with the shop floor.
Where to start: use cases that pay back fastest
Start with a decision your team already makes every day, using data you already hold. Avoid starting with a moonshot like a fully autonomous factory. Start with a recurring question that costs time or money when it is answered late.
| Use case | The decision it speeds up | Outcome to measure |
| Shortage summaries | Which components will run short, and for which orders | Production stoppages, expediting spend |
| Clear-to-build checks | Whether a work order can start today without stalling | Work orders started then paused, schedule changes |
| Supplier delay impact | Which customer orders a late delivery will hit | Late customer orders, proactive customer updates |
| Available-to-promise answers | Whether sales can commit to a delivery date | On-time delivery, order cancellations |
| Expiry and batch risk alerts | Which stock to use, move or quarantine first | Write-offs, quality holds released late |
| Traceability queries | Which batches went to which customers | Time to answer an audit or recall request |
The first use cases should recommend rather than act. A planner sees the suggestion, checks the reasoning and decides. Once acceptance is consistently high, the same use case can start preparing actions, such as a draft purchase order, for approval.
A 90-day path from pilot to production
Ninety days is enough to prove or drop one use case, provided each phase has a gate it must pass before the AI is trusted with more.
| Choose and connect Days 1 to 30 | Recommend Days 31 to 60 | Act with approval Days 61 to 90 |
| Pick one recurring decision Agree a metric and baseline Confirm data on one model Name a shop-floor owner | AI suggests, people decide Every suggestion explained Track acceptance and why Fix data gaps it exposes | AI drafts the action A person approves it Compare metric to baseline Plan the next use case |
| Gate 1: baseline and data agreed | Gate 2: planners accept the advice | After day 90: if the metric moved, repeat with the next decision |
The gates matter more than the dates. If a gate is not met, extend that phase or change the use case. Do not move on to letting AI act until your planners trust its advice.
How Axolt approaches AI in manufacturing
Axolt starts from the data problem, not the model. Inventory, manufacturing, procurement, logistics, finance and field service all run natively inside Salesforce, on the same data model as your customers and orders. There is no separate ERP to sync and no pipeline to build before AI can see the full picture.
That is what Axo AI, built on Agentforce, works from. Because it reads live operational records, it can help teams:
- move from stock questions to coordinated action, through Agentforce inventory operations
- check whether production is clear to build before a work order is released
- summarise stock shortages and review the impact of supplier delays
- surface expiry risk and answer batch traceability questions
The principle is simple: AI should work where your people already work, on the records they already trust, within the permissions they already have.
Questions to ask before you scale
Do we need a data science team to get started?
Not for operational AI on connected data. You need a process owner who makes the decision every day and can judge whether a recommendation is right.
What if our data isn’t clean?
It rarely is. Fix the data for one decision first, not for the whole business. AI exposes data gaps quickly, which is useful in itself.
Should AI be allowed to change orders or schedules on its own?
Not at first, and for consequential actions, usually not at all. Ask any vendor, including us, to show exactly where approval is enforced and how it is recorded.
How do we know the pilot worked?
Agree one business metric and its baseline before you start. If that number has not moved after 90 days, change the use case rather than extending the pilot.
Will AI see data a user shouldn’t?
It shouldn’t. Confirm that the AI respects the same record and field permissions as the person using it, and test it with a restricted user before rollout.
From pilot to production
The manufacturers getting real value from AI are not the ones with the most advanced models. They are the ones whose operational data is connected, whose teams trust the recommendations and whose results are measured in late orders avoided and hours saved.
If your customers already live in Salesforce, the fastest route to production AI is to bring your operations onto the same platform.
Bring one decision your team struggles with today, and we’ll show you how Axolt answers it on live data. Book a demo
