At Agentforce World Tour London in June, the manufacturing Customer Panel brought together leaders from RS Group, Forterra, Workdry and Holcim to discuss a practical question: how do manufacturers move from AI experimentation to safe, scalable deployment?
The discussion was deliberately grounded. This was not a conversation about AI hype or distant future ambition. It was about what manufacturing and supply chain organisations are learning now as they test, govern and deploy agentic AI in real business processes.
The clearest message came from RS Group: AI is already in the building.
Employees are experimenting. Teams are testing use cases. Some activity is structured and visible. Some is not. The leadership challenge is therefore shifting. Companies need to understand where AI is already being used, which experiments create value, what risks need to be managed and how the best ideas can be scaled.
That is especially important in manufacturing and supply chain, where customer commitments, product complexity, pricing rules and service obligations all create real operational consequences. AI can create advantage but only when it is connected to trusted data, clear processes and strong governance.
Lesson 1: Start with business outcomes, not AI projects
One of the strongest lessons from the panel was that successful AI deployment starts with the work that matters.
The best use cases were not framed as “AI projects”. They were framed as business problems: how to improve quoting, prepare sales teams for customer meetings, translate product information at scale, reduce repetitive work, improve service productivity and give teams better customer context.
The panel offered a simple test: does the use case help the business grow, improve customer experience, reduce cost, manage risk or make people more productive? If not, it should be challenged.
That discipline matters. AI enthusiasm can quickly create a long list of disconnected experiments. Some will be useful. Some will be interesting but low value. Some may create risk if they are not properly governed. The answer is not to slow innovation down, but to apply enough structure to make it useful: identify the use case, assign ownership, understand the data being accessed, measure the benefit, stop weak ideas and scale strong ones.
The deployment plan starts there: prioritise use cases based on business value, not novelty.
Lesson 2: Turn data into usable context
Manufacturers are often “data rich, information poor”. They hold large volumes of customer, product, pricing, order, service and pipeline data, but that information is often spread across systems, teams and formats.
For a salesperson preparing for a customer meeting, the data may exist, but it is rarely assembled into a clear view at the moment it is needed. They may need to check account records, recent orders, open opportunities, service issues, product interest, pricing context and customer history. That takes time and depends heavily on individual effort.
Forterra described how coworker-style functionality can help users access customer and sales information more naturally. Early use cases include meeting preparation, account summaries, identifying top accounts and surfacing relevant customer context. These are strong deployment candidates because the user, the workflow and much of the data already sit close to Salesforce.
The value is not simply retrieving a record faster. It is helping people understand what matters before they act.
For manufacturers, this is a major opportunity. Better decisions often come from better context: what a customer has bought, what issues they have raised, what opportunities are open, what commitments have been made and where the next conversation should go.
The deployment plan is clear: begin where AI can turn fragmented data into useful, timely context for the people closest to the customer.
Lesson 3: Use focused deployments to expose what needs fixing
The panel also showed that companies are not beginning with fully autonomous agents making high-risk decisions. They are starting with focused, controlled use cases where value is clear and guardrails can be defined.
Holcim shared a practical example around a high-volume, lower-value quoting process. The agent reads incoming emails, identifies the customer, product and delivery site, checks relevant data and creates a quote for human review. Importantly, it does not automatically send the quote to the customer.
That human checkpoint matters. Manufacturing quoting can be commercially sensitive. Pricing may vary by customer type, quantity, geography, delivery location, availability, contractual terms and wider commercial context. The agent can do much of the preparation, but human judgement remains part of the process.
The current fully successful automation rate was described as around 10%. On the surface, that might sound modest. In reality, it is highly instructive. The agent stops when data is incomplete, rules are unclear or records do not support a confident outcome. It is not just automating work. It is exposing where data, process and decision rules need to improve.
With cleaner data and clearer rules, the underlying potential is much higher. That is an important lesson for deployment. Agentic AI does not remove the need for operational discipline. It reveals where discipline is missing.
Lesson 4: Production needs governance from day one
The panel was clear that building an agent is often easier than productionising it safely.
A pilot can show promise. A demo can be compelling. But production deployment requires rigour. The data has to be trusted. The process has to be understood. Ownership has to be clear. Business rules need to be defined. Exceptions need to be managed. Outcomes need to be measured.
One useful analogy from the panel was to treat an agent like a new employee. You would not give a new hire unrestricted access to every system and every decision on day one. You would define their role, train them, set permissions, supervise their work and review their performance. The same logic applies to AI agents.
That framing makes governance practical. Each agent needs a clear role, known data sources, defined permissions, human oversight, measurable outcomes and a route to escalation.
The deployment plan is not “build and release”. It is define, test, supervise, measure and scale.
Lesson 5: Deploy Agentforce where the work happens
The panel did not suggest that every AI use case belongs in one tool. The view was more pragmatic: use the right assistant where the work happens.
For manufacturing and supply chain organisations, many of the highest-value use cases sit close to customer, commercial and service processes. They depend on account activity, pipeline, cases, product information, order history, pricing context, service issues and next best actions. In many organisations, Salesforce is already where much of that work is managed.
That is where Agentforce matters. Its value comes from being embedded close to the workflows, data, permissions and customer context that shape daily decisions. With the right connections into ERP and other enterprise systems, often through MuleSoft, agents can support richer decisions by bringing together customer, product, pricing, order, availability and service data.
The goal is to make manufacturing and supply chain data more usable, conversational and actionable at the moments where decisions are made.
The companies who succeed with agentic AI will not necessarily be those with the most experiments. They will be the ones that connect AI to business outcomes, trusted data, clear processes, strong governance and measurable value.
The opportunity is to help teams make better decisions, act faster and use the data they already have with greater confidence.
Catch the Agentforce World Tour London Keynote on Salesforce+.
