AI in data migration

Where the time goes

Migration effort is not the 'Migrate' button. It is mapping specs, value maps, reading load errors after each mock, and explaining reconciliation numbers. Those text- and pattern-heavy tasks are where LLMs help. The business still signs the mapping, the Migration Cockpit (or your ETL tool) still loads, and you stay accountable.

AI-assisted field mapping

Give the model legacy fields (names, descriptions, value profile) and the Migration Cockpit template fields. It returns a draft mapping with confidence and reasons. Confirm the easy 1:1 lines fast; spend your time on account group to BP grouping and roles, company code versus purchasing org fields, and Z-fields. The output is a draft spec, not a load file.

You assist an SAP data migration consultant (ECC to S/4HANA, Supplier).
SOURCE: legacy fields with description and value profile (JSON).
TARGET: Migration Cockpit template fields for the Supplier object (JSON).

For each source field return:
- source_field, target_field (or null if none fits)
- transformation: "direct", "value_mapping", "derive", "split", "concatenate" or "drop"
- confidence (0-1) and a one-sentence reason
- open_question for the business if the meaning is unclear

Only use target fields from the TARGET list.
Custom Z fields: never guess a target; set target_field null and ask.
Return JSON: {"mappings": [...]}

Value mapping suggestions

Most mock-load errors come from value maps: account groups, payment terms, regions, tax categories. The LLM proposes from legacy values plus allowed target values with descriptions. Your script rejects any target not in the target system's customizing list.

Try it: tighten the value-map rules

A suggested target value is accepted only if it exists in the ______business ownertarget customizing listBasis teamconsultant reviewmodel's training dataauto-accept queuewho signed it off. Suggestions below 0.8 confidence go to a ______business ownertarget customizing listBasis teamconsultant reviewmodel's training dataauto-accept queuewho signed it off. A legacy value with no description is sent to the ______business ownertarget customizing listBasis teamconsultant reviewmodel's training dataauto-accept queuewho signed it off. Every approved value map records ______business ownertarget customizing listBasis teamconsultant reviewmodel's training dataauto-accept queuewho signed it off.

Error clustering after a mock load

Thousands of errors are usually a few problems repeated. Cluster without AI first: group by message class and number, and replace record-specific values with placeholders. Then send the LLM one example and a count per cluster, never record keys, and ask for root cause and where to fix it.

import re
import pandas as pd

log = pd.read_csv("mock2_messages.csv")  # object, record_key, msg_class, msg_no, text

def normalize(text):
    text = re.sub(r"\b\d{3,}\b", "<NUM>", text)         # long numbers / keys
    text = re.sub(r"\b[A-Z0-9]{2,4}\b", "<CODE>", text)  # short codes
    return text

log["pattern"] = log["text"].map(normalize)
clusters = (log.groupby(["msg_class", "msg_no", "pattern"])
               .agg(count=("record_key", "size"), example=("text", "first"))
               .reset_index().sort_values("count", ascending=False).head(15))
# send clusters.to_json(orient="records") to the model: root cause + fix location

Try it: 8,000 errors after mock 2

AI across the mock-load rhythm

Drafts before each mock, triage after, a reconciliation narrative before sign-off. For narratives, every number comes from your reconciliation queries and must be checked; one invented number does more damage than no narrative. At cutover, prompts and scripts are frozen.

Many source systems

With many sources, run the same profile and prompt per source, then ask the model to list cross-source conflicts. That list becomes one harmonization workshop instead of fifteen. Label everything draft, reviewed or approved.

Try it: AI, human, or neither?

What SAP itself offers

Know the headlines without overselling. SAP publishes material on AI-assisted data migration with SAP Cloud ALM (check scope in the docs). Its Q3 2026 highlights describe Transition Data Quality agents that apply fixes only after human approval, in phased GA. At Sapphire 2026, SAP and Palantir announced AIP for data migration as an SAP Endorsed App. Position your scripts as complementary, with the same governance.

Check yourself