# Making Your Dinosaur ERP Act Its Age: Why "Context" Beats Simple Prompting Every Time



> Stale ERP data and missing shop-floor facts → get context-aware answers that cut errors, speed scheduling and stop bad quotes.
> 
> **URL:** https://www.gosmarter.ai/blog/erp-context-beats-simple-prompting/

**Date:** 2026-06-26
**Author:** BlogSmarter AI

**Categories:** blog

**Tags:** artificial-intelligence, data-strategy, manufacturing

## 



**Your ERP is not stupid. It’s just old and being asked questions it was never built to answer.** If you want AI to help a metals team make sound calls, the short answer is this: **context beats prompting every time**.

The pain is obvious. **Bad schedules, thin quotes, missing certs, and people wasting hours digging through PDFs, emails, and mystery fields like `ATTRIBUTE1`.**

I see the fix as pretty simple. You do **not** need to rip out a 10 to 20-year-old ERP to get useful AI. You need a context layer that pulls in live stock, routings, pricing rules, heat data, and documents before the model answers. That’s where **[GoSmarter](https://www.gosmarter.ai/hubs/ai-for-metals-manufacturing/)** fits for metals manufacturers. Tools like **[MillCert Reader](https://www.gosmarter.ai/docs/digitising-mill-certificates/)**, **[Product Lineage](https://www.gosmarter.ai/hubs/integrated-cert-traceability/)**, **[Business Manager](https://www.gosmarter.ai/hubs/gosmarter-for-metals-operations/)**, and **[Production Planner](https://www.gosmarter.ai/hubs/shop-floor-planning-software/)** give the AI the facts your shop runs on.

What you get is plain enough:

-   **Less guessing** in scheduling and quoting
-   **Fewer manual checks** on mill certs and traceability
-   **Lower risk** when [CBAM](https://taxation-customs.ec.europa.eu/carbon-border-adjustment-mechanism_en) and batch records need backing up
-   **Faster replies** to customers without making up lead times or scrap rates
-   A sane first step that starts in **read-only mode**, not a full ERP circus

The article’s point is blunt: **a smarter prompt will not fix missing shop-floor facts**. If the AI cannot see the machine status, stock position, cert PDFs, or customer rules, it fills the gaps and hopes nobody notices.

Here’s how to fix it.

## What the future holds for software development in metal fabrication

{{< youtube width="480" height="270" layout="responsive" id="Zqfl17EhPzU" >}}

## Where Simple Prompting Breaks: Schedules, Quotes and Compliance

{{< image src="6a3dc2712902db05ecd80572-1782434818088.jpg" alt="Prompt-Only AI vs. Context-Aware AI in Metals Manufacturing" >}}

The cracks show up first in scheduling. Then quoting. Then compliance. Same problem every time: the AI sounds sure of itself, but it **doesn't have the facts in front of it**.

### Why your production schedule is still a guess

Ask an AI, "What should we run today?" and give it nothing else, and it'll hand back something that looks sensible. That means very little on a metals shop floor. **Sensible-sounding is not the same as right.**

Your ERP stores records. It does not always show the live picture. If the model can't see machine status, coil widths, due dates and changeover times, it starts filling in the blanks. So you get a schedule that is wrong, delivered with total confidence. That is useless for live shop-floor decisions [\[2\]](https://casys.ai/blog/context-engineering-guide).

The gap hits the same places every day:

| Decision | AI with Live Context |
| --- | --- |
| Machine availability | Pulled from live ERP and shop-floor data |
| Job sequencing | Sequenced against due dates, routings, coil widths and changeover constraints |
| Priority changes | Re-sequenced when due dates or machine status change |

The same blind spot turns fast quoting into expensive guesswork.

### Why fast quotes go wrong when the AI cannot see the rules

A metals quote lives or dies on the details: alloy grade, thickness tolerances, yield assumptions, scrap rates and customer-specific pricing rules. Miss one, and the margin is wrong before the quote even leaves the desk. A prompt-only AI will happily invent the missing inputs with plausible numbers. Looks tidy. Sounds smart. Still wrong for your grade, your customer and your stock at that moment.

> **A quote built on assumed scrap rates is not a quote - it is a liability with a reference number.**

A lot of built-in ERP AI tools make this worse because they can only see what's on the user's screen right then. They can't check inventory, purchase orders and production schedules in one query [\[6\]](https://twbs.com/resources/blog/why-we-stopped-waiting-for-ai-to-come-to-erp-and-started-building-it-ourselves/). So the quote comes back fast, polished and **expensively wrong**.

Compliance falls over for the same reason. If the model can't see the certs or carbon data, it can't check them.

### Why scrap, certs and [CBAM](https://taxation-customs.ec.europa.eu/carbon-border-adjustment-mechanism_en) checks get messy fast

{{< image src="72bf75d8d513eafd6f7261d44e873d41.jpg" alt="CBAM" >}}

Scrap tracking and compliance look simple until you ask the AI to trace an offcut back to a heat code and match it to a mill certificate PDF. From a vague prompt, with no document access, that's fantasy.

| Risk Area | Prompt-Only AI (No Context) | AI with Live Context |
| --- | --- | --- |
| **Mill certificates** | Confidently claims a cert exists or fabricates heat details | Retrieves and verifies specific mill cert PDFs against the physical batch |
| **CBAM reporting** | Uses outdated or generic carbon values; risks EU regulatory penalties | Injects real-time supplier data and regional policy constraints for accurate filing |

Get the policy context wrong and you get the carbon figure wrong. Then the filing is wrong too.

## What Good Context Actually Looks Like in a Metals Shop

Useful context in a metals shop is not some magic prompt trick. The problem is usually **missing shop-floor detail**, not weak AI. Pair legacy ERP with the right operational inputs and it stops acting like a dusty filing cabinet. It starts helping you make decent calls.

### The 6 pieces of context that change the answer

These are the inputs the model needs before it can make a call.

-   **Job routings:** show the order of operations and where the bottlenecks sit.
-   **Coil and heat data:** confirm whether the material in stock is fit for the job.
-   **Live inventory status:** show what is _actually_ available right now, not what the system said yesterday.
-   **Pricing rules:** apply customer discounts, [LME](https://www.lme.com/) surcharges and freight rules.
-   **Customer specs:** spell out tolerances, packaging and quality rules.
-   **Production constraints:** show capacity, setup time and machine downtime.

Together, these inputs make up the context layer the model needs before it can reason properly.

### What a better request looks like on the shop floor

The gap is not prompt length. It is input quality. A longer prompt full of waffle still gives you a bad answer if the model cannot see the facts that matter.

You see the difference fastest in everyday shop-floor jobs:

| The Manual Way | The Automated Way |
| --- | --- |
| **Scheduling:** "What should we run next on the slitter?" | Sequence the next three jobs for Slitter 2 using current coil inventory from the ERP, the blade change on Slitter 2 from the maintenance schedule, and the 92% yield average for 0.5 mm gauge from the MES [\[2\]](https://casys.ai/blog/context-engineering-guide). |
| **Order expediting:** "When will order #4471 ship?" | Predict the ship date for #4471 by cross-referencing the delayed coil delivery from the supplier's lead-time model with the current production backlog on Line 3 [\[2\]](https://casys.ai/blog/context-engineering-guide). |
| **Compliance:** "Check the certs for this heat." | Validate Heat #882 against Customer Spec X-100, pulling the chemical analysis from the QMS and confirming the CBAM certificate is attached to the original PO [\[1\]](https://annora.ai/articles/a3-why-every-manufacturer-needs-a-company-brain/). |

Context-rich requests give the AI the facts it needs, so the answer matches **your shop, your stock and your constraints**.

## How to Add Context Without Replacing the ERP

Most manufacturing SMBs are still stuck with ERP systems that are 10 to 20 years old. And let’s be honest, ERP migration projects have a nasty habit of chewing through time and cash, then missing the mark anyway. A better move is to add a context layer on top of what you already run. That layer pulls in the right records, documents and rules _before_ the AI tries to answer. The hard part is not replacing the ERP. It’s getting the right context into the answer without tearing the whole place apart.

### Use retrieval and document layers to stop the guessing

Retrieval-Augmented Generation, or [RAG](https://en.wikipedia.org/wiki/Retrieval-augmented_generation), pulls the exact record first, then builds the answer from that source. That could be a mill certificate linked to a heat code, or the pricing rule behind a surcharge [\[4\]](https://p2-innovate.com/blog/forget-fine-tuning-orchestrate-your-erp-ai/)[\[2\]](https://casys.ai/blog/context-engineering-guide). **No guessing. No made-up filler.**

This works well with messy data, which is what most old ERPs are full of. Bent fields, odd naming, half-finished records, PDFs nobody wants to touch. That’s normal. You do **not** need a spotless data warehouse before you start.

### Give each team its own AI assistant, not one bot for everyone

One assistant for planners, quality and sales sounds neat on paper. In practice, it’s useless. Each team needs different context.

-   **Planners** need sequence logic and capacity data.
-   **Quality teams** need mill certificates and heat traceability.
-   **Sales** needs live pricing and lead times [\[7\]](https://www.geniuserp.com/en-gb/resources/blog/ai-manufacturing-erp-genius-cortex/).

Role-based assistants fix this by keeping each user inside the records they actually need. A quality engineer gets mill certificates and traceability data. A sales rep gets pricing rules and lead times. Neither sees the other’s data unless their role allows it. That keeps answers tied to the facts that matter for that job and improves output quality [\[2\]](https://casys.ai/blog/context-engineering-guide). So yes, the assistant should be scoped by team, not dumped on the whole business as one generic bot.

### Where [GoSmarter](https://www.gosmarter.ai/hubs/ai-for-metals-manufacturing/) fits when your system is older than your apprentices

{{< image src="eb99d853cd9ddf944d250f8e08898720.jpg" alt="GoSmarter" >}}

GoSmarter sits on top of your current ERP instead of ripping it out. It adds the context machinery that older systems were never built to handle.

-   **MillCert Reader** uses AI OCR to read and digitise PDF mill certificates, cutting manual data entry errors.
-   **Product Lineage** handles heat-code traceability and links inventory records to the right certificates automatically.
-   **Business Manager** covers inventory tracking, order management and scrap tracking in one place, so the AI can see live stock status.
-   **Production Planner** gives you first-draft cutting plans tied straight into inventory and orders.

Rollout matters as much as the software. Start in read-only mode. Let the AI extract and surface data without writing back to the ERP. That gives your team time to trust it and shows up data gaps without putting operations at risk [\[3\]](https://kamna.vc/2026/03/27/legacy-erp-ai-transformation-manufacturing/)[\[5\]](https://superkind.ai/blog/legacy-ai-agents). Most shops move from that first phase to AI-assisted recommendations within 8 to 12 weeks [\[5\]](https://superkind.ai/blog/legacy-ai-agents). The safest first step is read-only extraction on one job that’s hurting today.

## Start Small: Run GoSmarter on One Painful Job This Week

Once the context layer is in place, **don’t roll it across everything at once**. That’s how good ideas get buried under meetings, setup faff and “we’ll come back to it next month”. Prove it on one live job first.

Pick the single job causing the most pain **right now**. That might be certificate handling, traceability, scrap tracking or manual rekeying. Go for the smallest live task that shows the gap fastest.

If mill cert handling is the main bottleneck, run **MillCert Reader** on the next batch of certificates that lands on your desk. If traceability is the bigger headache, use **Product Lineage** on the next check. Don’t overthink it. Use the job that puts the current process under the harshest light.

The goal this week is simple: prove, on one live job, that **context-aware AI beats the current process**.

Track what changed:

-   time saved
-   fewer manual checks
-   faster customer reply time
-   lower compliance risk
-   less rework

Then use that baseline to pick the next job.

## FAQs

{{< faq question="What is a context layer in ERP AI?" >}}
A context layer in ERP AI acts as the semantic bridge between **messy ERP field names** and the business meaning your team actually cares about. It helps AI read old, opaque system data and map it to plain operational ideas like **supplier reliability** or **production schedules**.

It also gives the AI the right data and role-based limits _before_ it handles a query. That means the system can make sense of your setup as it stands, without forcing you into a full ERP rip-out just to get useful answers.
{{< /faq >}}

{{< faq question="How does read-only rollout reduce risk?" >}}
A read-only rollout cuts risk. Your legacy ERP stays the system of record, while the AI layer only **searches, summarises, explains and recommends** through controlled read access.

That keeps the **blast radius small**. The AI can't directly change money, stock or core records. Read calls sit behind a thin integration layer that normalises the data and fails cleanly if systems are slow or down. If you want write actions later, put them behind human sign-off first.
{{< /faq >}}

{{< faq question="Which task should we automate first?" >}}
Start with jobs where the ERP already has the data, but your team still loses time making calls. Think **replenishment planning, production scheduling, quote generation, and customer exception handling**. These are the sort of workflows that clog up the day without forcing you into a full system rip-out.

For the first rollout, use **shadow mode**. Let the AI layer produce recommendations alongside the live system, but don’t let it touch anything yet. That gives you room to build trust, check the data isn’t a mess, and see how the suggestions stack up before you move to AI-assisted decisions.
{{< /faq >}}