01
Different systems speak different business languages
A supplier sends a category code. Your procurement system expects an internal category. A legacy export uses a description that your new application does not recognize. Someone has to decide how the source information fits the destination.
AI data mapping learns these relationships from mappings your team has already approved. StayCharted AMT uses those examples to suggest destination values for new records, using the terminology your business works with.
02
Teach the model with mappings you already trust
Start with an export containing source values, any context needed to interpret them, and the approved destination values. For example:
| Source value and context | Approved destination |
|---|---|
| ELEC / CBL / DATA — supplier product category | Network cables |
| End-user computing — purchasing description | Computer hardware |
| Office premises upkeep — service description | Facilities maintenance |
Keep the destination labels consistent. If the same source code means different things for different suppliers, include the supplier or other relevant context in your examples. A code by itself may not contain enough information to choose the right mapping.
03
Mapping, matching, and categorization answer different questions
- Data mapping: How does this source value translate into the destination system’s terminology?
- Record matching: Which specific existing item or record does this refer to?
- Categorization: Which business category does this record belong to?
These tasks can overlap. Mapping a supplier’s product category into your internal taxonomy is different from identifying the exact catalog item. Choose one destination field and one clear question for your first model.
04
Keep straightforward mappings straightforward
An exact lookup works well when a source code always maps to one destination. A trained model is useful when descriptions vary or the mapping depends on context that is difficult to capture in a lookup table.
Include representative variations and exceptions in your training examples. Test the model on records it has not seen, and compare its suggestions with mappings your team has reviewed. Look closely at ambiguous descriptions and newly introduced codes.
05
Review suggestions before updating the destination
Start by reviewing mapped results in a batch file. Once the model is useful for your task, your application can call the StayCharted AMT API within its existing import or data-preparation workflow.
- Source record + relevant context
- Your application or batch workflow
- StayCharted AMT model
- Suggested destination value
- Review and apply the mapping
Your application controls validation and updates. Keep records with unclear or unsuitable suggestions available for review, and check that the destination value still exists in your current taxonomy.
06
Start with one recurring translation task
Pick a field your team repeatedly translates between systems: supplier categories, service descriptions, or internal codes. Gather approved mappings and a separate set of examples for evaluation.
- Check how often suggestions match the reviewed destination.
- Look for errors between closely related categories.
- Review exceptions before expanding to another source or field.
As source terminology or destination categories change, update the examples and evaluate again. A useful mapping model should reflect the systems and vocabulary your team uses today.