AI Agents for Manufacturing Sales: Where to Draw the Line

A useful first question for a manufacturer considering AI agents is: what would we allow this system to change if its interpretation were wrong? Reading a maintenance document, drafting an RFQ response and changing a production parameter carry very different consequences. Calling all three an “agent” hides the decision that matters.

For a commercial director or operations lead, the immediate task is to draw that boundary. Start with one workflow whose inputs and exceptions people can inspect. Expand its permissions only after testing the next action, not because the demonstration looks convincing.

Separate access from authority

An agent may need access to a product catalogue to prepare a response. That does not mean it should choose a technically equivalent substitute, commit to delivery or send a quotation. Give each permission its own justification.

The distinction matters when a request is incomplete. A system that leaves an operating condition blank may be behaving correctly. A fluent answer that silently fills the gap is harder to spot and potentially more costly.

Test the awkward cases first

Build the evaluation packet from approved records that represent the work, including the cases experienced staff disagree about. Remove unnecessary personal and confidential information. Keep a record of who decided the expected outcome and why.

For RFQ intake, include an attachment that contradicts the email, a missing duty point, a product reference from an old catalogue and a request that belongs to a distributor. For each, specify the acceptable response. Sometimes it should be a question. Sometimes it should be a referral to engineering. A completed answer is not always the desired output.

Run the assistant beside the existing process before allowing write-back. Record missed requirements, invented facts, reviewer corrections and handling time. Separate easy cases from exceptions; a good average can conceal an unacceptable failure on the requests that matter most.

Judge the draft before the adoption statistics

Source message

"Our unit is marked R-17. The attached manual says R-17B. Please confirm the replacement part and whether it can ship on Friday."

Accept this draft

"The equipment identifier and manual revision differ. Service engineering must confirm the applicable parts list. Friday delivery has been requested but not confirmed."

Reject this draft

"R-17B is compatible with R-17. We can send the replacement on Friday." Neither compatibility nor availability is established by the source.

Reviewer decision

Allow preparation of the clarification. Do not permit part selection, stock promises or external sending. Record the unresolved revision before the request moves on.

The rejected draft sounds more helpful. That is why a test based only on completeness or tone would miss the problem. In a service workflow, a useful assistant can shorten the search for evidence without taking over the decision about interchangeability.

In its August 2026 survey, McKinsey reports agent scaling among 40 percent of respondents from organisations above $1 billion in revenue, compared with 22 percent from smaller organisations. That is cross-industry survey evidence, not a readiness benchmark for your factory.

The Manufacturing Leadership Council's 2025 report points to data availability, structure and measurement as practical obstacles. These are reasons to examine the process around a model. They do not establish that every business needs a new data platform before it can try a bounded assistant.

Use research to challenge the investment case. Use your own test packet to decide whether this particular workflow is ready.

The work that remains after the pilot

Someone must own source updates, access changes, failed runs and rollback. Ask what happens when a catalogue changes, an employee leaves, or the model starts producing a different style of output. Budget for that work before describing the pilot as an operating capability.

NIST's AI Risk Management Framework is useful for organising accountability and evaluation. It is not a certificate that an agent is safe. A sensible local control is to suspend write permissions when source traceability fails, while preserving the record for investigation.

Keep safety-critical control, equipment selection and regulated decisions outside a commercial pilot. Those uses need qualified engineering, security and legal assessment of their own. An agency's competence in enquiry handling does not confer authority over plant operations.

Before commissioning the next stage, ask the owner to show one accepted output, one rejected output and the reason for each. If the team cannot explain that distinction, more autonomy would make the uncertainty harder to manage.