AI data dictionary draft
Object inventory, business-area classification, key columns, confidence levels, sensitivity flags, and notes that still need an SME to confirm.
AI for MES
Your Manufacturing Execution System (MES) holds the operational record. Engineers shouldn't need raw SQL to question it.
Manufacturing databases were built for MES applications, not for AI. Business logic hides in procedures, overnight jobs, reports that don't match, and people who've been on the floor for years. We recover that context before anyone connects a model to production.
The failure pattern
Pointing AI at raw MES tables is a common failure pattern.
Unrestricted table access often produces answers engineers can't trust: yield looks wrong, WIP counts disagree with the floor, and hold status ignores an exception everyone on the line already knows.
The first useful milestone isn't "AI can query SQL." It's "AI can answer manufacturing questions using approved, written business logic." If that isn't written, we stop.
- Schema alone is not meaning Tables, views, procedures, jobs, and report logic all shape what a number means. Schema is just the start.
- Hidden relationships are common Many MES databases barely have formal foreign keys. Relationships live in naming, procedures, and screens
- Governed access or no access We propose read-only paths through approved views and client review, not unrestricted table access.
Where meaning usually lives
Where meaning usually lives on the floor
- WIP trackingLots, wafers, carriers, cycle time
- YieldDefects, bins, rejects, density
- EquipmentUtilization, downtime, chambers
- SPCCharts, violations, out-of-control events
- RoutingOperations, recipes, queue time
- MaintenancePM schedules, MTBF, MTTR
- Holds and qualityHold status, releases, quality gates
- IntegrationsEnterprise Resource Planning (ERP), SAP, equipment, overnight jobs
Platforms vary. Same approach across commercial and custom MES environments. Product names appear only to describe publicly verifiable experience or client-approved project context.
Proposed access path
AI should meet approved meaning, not every raw table.
When the work reaches architecture, we write down a proposed path. Your technical, security, privacy, legal, and governance stakeholders can review it before anything connects.
- 01
Manufacturing equipment and shop-floor systems
- 02
MES databases and application services
- 03
Approved semantic views and business datasets
- 04
Model Context Protocol (MCP), REST API, or middleware layer
- 05
AI assistant layer subject to client-chosen models
The architecture is technology-agnostic. Claude, ChatGPT, Azure OpenAI, Copilot, Gemini, Bedrock, and private LLMs are examples of model providers or options a client may select and approve. Naming them does not imply partnership, sponsorship, certification, or endorsement. We are not selling a model.
What useful discovery produces
Context people can inspect before a pilot begins
Exact deliverables follow the written scope. Below are the kinds of artifacts that make AI and analytics review possible; that careful approach is intentional.
Relationship and lineage map
Documented and inferred links across lots, wafers, routes, equipment, defects, holds, and integrations. Confidence called out so nobody confuses a guess with a fact.
Approved-view recommendations
Proposed semantic views for questions like WIP status, lot history, yield summary, equipment utilization, defects, and lots on hold, not open-ended table access.
Governance and access notes
Read-only account patterns, restriction candidates, audit and logging considerations, and open risks. Your team reviews them. We don't skip that.
Strong first use cases
Start where engineers already argue about the answer
- WIP status Current work by operation, product, or hold condition Usually a clearer first step
- Lot history search Trace moves, status changes, and genealogy for a selected lot Usually a clearer first step
- Yield trend analysis Compare losses, bins, and periods against definitions people actually trust Harder. Still worth scoping carefully
- Equipment utilization Downtime, availability, and chamber or tool patterns Harder. Still worth scoping carefully
- Defect trend analysis Recurring defect patterns against validated sources Harder. Still worth scoping carefully
- Lots on hold Why a lot is held, who owns release, and what changed Usually a clearer first step
How the work is framed
Discover, define, govern, then activate. In that order.
-
Discover
Map the MES environment, stakeholders, trusted reports (and the ones that don't match), and the manufacturing question that matters.
-
Define
Recover schema meaning, relationships, calculations, exceptions, and the places SMEs disagree out loud.
-
Govern
Propose approved views, access restrictions, validation owners, and open risks. Client review before anything goes live.
-
Activate
When separately agreed in writing, we define or implement a bounded pilot around documented context, and not before.
Timing depends on access, evidence, risk, and client decisions. Illustrative MES roadmaps often span roughly ten weeks when scope, access, and participation support that pace. Nothing here is a fixed-duration promise. If access stalls, the calendar stalls.
Working commitments
What we will do, and what we will not claim.
We'll document before we connect
Schema, meaning, ownership, and restrictions come before end-user access to model answers. If that isn't written, we stop.
We'll keep humans in validation
SME review stays part of deciding whether a definition, view, or answer path is trustworthy enough to use. An AI answer with no source isn't good enough.
Client-specific outputs stay with you
Dictionaries, views, documentation, and other client-specific materials are delivered under the ownership terms in the engagement agreement.
We'll report what we find
If discovery shows the blocker is process, ownership, or data quality rather than model selection, we say so. We won't pretend a better model fixes a definition fight.
We do not claim partnership with, certification by, or endorsement from any MES vendor unless expressly stated in writing.
Start with one manufacturing question the system can't answer consistently.
Tell us which MES you run and which decision needs clearer evidence. A short first conversation is usually enough to see whether a focused engagement makes sense.