AI field mapping with human approval
AI field mapping is the use of a model to draft source-to-target field pairs and transformation rules from system metadata and plain-language descriptions, each proposal carrying a confidence score and a rationale (semantic fit, type compatibility, length, value distribution). Used safely, every draft passes a sandbox test, a named human approval, versioning and an audit trail before it is loaded — cutting the mechanical mapping effort for a complex conversion object from about 30 expert-days to about 12 without loosening governance.
Updated · Visehub
What the AI does — and what it must not
Field mapping is the slowest, most expert-locked step of an ECC → S/4HANA or cross-ERP migration: hundreds of source-to-target pairs and the transformation logic behind them. AI is a good draftsman for it. From system metadata and a plain-language description it proposes field pairs, transformation rules and — from a sentence — validation rules, each with a confidence score and a rationale: semantic fit, type compatibility, length, value distribution.
What it must not do is decide. The expensive work in a migration is defining what correct means for each field and proving it before the load. That stays with people, and the operating model around the AI is what makes it safe.
The control loop
| Stage | Step | Who / what |
|---|---|---|
| 0 | Plain-language specification | The consultant describes the mapping need |
| 01 | AI draft | vise FieldMap proposes pairs and rules |
| 02 | Sandbox test | Every rule runs in a tester on real records |
| 03 | Human approval | A named reviewer accepts, edits or rejects each line |
| 04 | Version and audit trail | Each decision carries an owner and a date |
| 05 | Governed load | The approved mapping feeds the simulation and the load |
Proposals sort into three groups — approved cleanly, approved with problems to review, declined — and the user keeps the final decision in a target-driven view: every mandatory target field is shown as resolved, defaulted, enrichment, manual-load or unresolved, so nothing is forgotten.
The numbers
On the same complex conversion object: about 30 expert-days by hand versus about 12 with FieldMap — roughly 60% fewer expert-days on the mechanical part, with the same deliverable and the same governance. The figures are illustrative and should be validated on your object; the direction is consistent. For comparison, fully in-house preparation of one object such as Business Partner commonly exceeds 1,000 hours.
Two failure modes to avoid
Over-trusting the AI — unverified output moving toward production, errors the team cannot trace surfacing inside the cut-over window. Dismissing the AI — senior consultants writing every rule by hand, the slowest step done the slowest way. The loop above is the answer to both.
In production
The same pattern — suggestion, sandbox, approval, versioning, audit, four-eyes on execution — runs on a multi-country programme at a global brewing group, and FieldMap’s first production customer, Karlsberg, maps material master data from Business Central to S/4HANA with it. The longer argument is in What AI actually does in an S/4HANA data migration.
Frequently asked
- Does the AI clean the data?
- No. It drafts mappings and rules. Defining what correct means, proving it before the load and deciding which records are acceptable stay with people. Treated as a cleaning button, AI moves risk out of view instead of removing it.
- What does a confidence score tell me?
- How well the proposed pair fits on semantic meaning, data type, length and value distribution, with the reason in plain words. Low confidence or a flagged problem means a person looks before the pair is approved; high confidence does not skip the approval.
- What kinds of mistakes does it catch?
- An internal SAP address reference (ADRNR) that looked like a clean match to a document-number field; a 35-character source mapped to a 30-character target; codes like 00123 that lose their leading zeros in a numeric field.
- Can business consultants do mapping work with it?
- Yes — that is one of the main effects. The expert moves from typing rules to judging them, and consultants without classic ETL experience take part in mapping under the same governance as developers.