The system has been replaced—how can you say it was the right choice?

In an ERP migration project, the time when the most people argue is not the go-live day, but the acceptance meeting two weeks before go-live.

Party B said: The features are all configured, the data has all been migrated, and UAT has passed too. Party A asked: What does "passed" mean? Party B produced a sign-off sheet. Party A stared at the sheet and said a line every implementation consultant has heard—"Who checked these boxes?"

SAP gave a somewhat interesting answer this month. On November 11, ERP News published an interview with Sonja Liénard, who is Senior Vice President at SAP and head of the ABAP platform, while also overseeing AI capabilities for ABAP. She went through the seven types of migration assistants SAP is working on, and none of the seven sections is called “installation”; each section comes with a note on “what the migration team needs to check.”

1. Breaking down the seven sections: each section comes with a review checklist

These seven categories of workflows were first announced at Sapphire 2026 under the name Migration and Modernization Assistants, with SAP Cloud ALM as their common foundation. As of October 28, four of them—data management, custom code, business configuration, and test management—were already supporting early customer cloud ERP projects; system analysis, project management, and go-live rollout are still planned. Liénard said more tools will be released before the end of the year.

AssistantWhat does it doWhat must the migration team verify first
System analysisScan the source system's data, configurations, and custom-built developments into a common profile, with a dependency model at the object and system levelsList of source systems, company codes, interfaces, batch jobs, and self-built applications; the scope and assessment date must be kept on file as a baseline
Data ManagementDetermine what to fix first, which data to migrate, and what to leave in place; prioritize the remediation order by business impact, workload, and dependencies.Key datasets must have an owner; merge and exclusion rules must be set before the batch; during trial migration, verify both record counts and total amounts at the same time
Custom codeUse ABAP Test Cockpit results, usage information, and code dependencies to group technical objects into business extensions, and recommend retiring them or replacing them with standard functionalityEach extension has a business owner and a keep-or-remove conclusion; low-frequency programs must be checked against the business calendar—year-end jobs look unused in a short window.
Business ConfigurationPropose target configurations by requirements, industry, and region, map the organizational structure, identify differences between source and target, apply where permitted, and maintain an audit trailCompany code and plant structure, approval settings, and country requirements require business confirmation; high-impact configurations must be tied to test cases.
Test ManagementInfer what should be tested from process models, requirements, code, configurations, and interfaces, fill coverage gaps, prepare test data, and hand it over to Tricentis for executionEvery key process must have expected business outcomes and signatories; order-to-cash must cover partial shipments, interface rejections, and credit holds.
Project ManagementEstablish a baseline using the project charter and system information, provide role-based guidance, expose risks, and coordinate tasks across workflows.Each milestone defines 'gate-closure evidence': data loading must have reconciliation results and signatures, and code remediation must be linked to changes and retesting.
Launch promotionCompare the global template with local processes and regional requirements, and recommend whether to modify the template or handle it locallyFor each deployment wave, create an exception register: requirements, owner, affected configuration or interface, and evidence required for approval.

The most noteworthy thing about this table is not the left column, but the right column. What is written on the right has always been the vaguest part of implementation projects, and now SAP itself has written it into the product documentation—that is to say, going forward, people buying this solution will very naturally ask the vendor: 'Can your assistant produce these records?'

The safety net Liénard discusses in the interview follows the same playbook: isolated test environments, quality checks, clear instructions, and an audit record that can show who initiated it, what instructions were received, and what was changed. It is uniformly about “being able to check after the fact” capabilities, with not a single sentence about how smart the model is.

II. First ask about the version, then about capabilities

The most useful detail in this material is that it exposes just how unreliable the word "support" really is.

  • The Business Configuration Assistant initially covers only SAP Cloud ERP Public, and will expand to Private only in 2027.
  • The custom code assistant focuses on ECC migration to Cloud ERP Private, along with dedicated Dynpro modernization and RAP artifact generation.
  • ECC's system conversion path only supports Cloud ERP Private and on-premise deployment; for Public, it is always treated as a new implementation.

Put these three points together, and the conclusion is straightforward: for the same set of “seven types of assistants,” under different target versions and migration paths, only some of them are usable. So asking “Do you support AI migration?” is as good as not asking; it must be broken down into four questions—which target version, whether it is a system conversion or a new implementation, whether this segment’s assistant is on the available list, and which sub-function is available.

This is not a phenomenon unique to SAP. Over the past two years, almost all ERP vendors’ AI feature lists have been organized by “application” rather than by “version + path + role.” If the client signs off by application name, they will certainly encounter at acceptance: “This feature is not applicable to your version.”

III. The compliance section is the one most easily glossed over with “successful launch.”

If there is one type of work in migration that most easily produces “formal completion,” it is local compliance.

On November 9, Odoo Italia announced the integration of fiskaly into POS, allowing Italian merchants to report retail sales data to the tax authority via the cloud, no longer requiring a dedicated fiscal printer or electronic cash register (registratore telematico, RT). Odoo says there is no additional charge, and the service behind it is provided by fiskaly's SIGN IT.

Sounds like a clean victory: one fewer device, one less maintenance task, one fewer routine inspection. But when the messages from all three parties are stacked together, things are not so clean.

LevelClearly definedStill unresolved
Compliance EligibilityThis fiskaly software is in a certification pilot and is planned to launch in the first half of 2027Does the currently live version go through the existing portal channel or the new approval framework?
Stock SwitchoverOdoo already supports Germany's SIGN DE and Austria's SIGN AT; Italy is the third marketWhich Odoo versions are supported, hosting methods, and how stores switch from RT
Payment LinkageItaly requires merchants to link electronic payment devices with the reporting system; registration can be done on the tax authority portalHow should the registration of existing payment terminals be handled?
Failure handlingOn November 2, the portal was unavailable; submissions timed out or reported errors, and some requests that "appeared to have failed" may have already been delivered.What the cashier sees, whether it retries automatically, and how to prevent duplicate submissions
Closing reconciliationAbout 95% of tax bureau merchants still use RT; Deputy Minister clarifies no authorization for third parties to automate its web processesHow to confirm which transactions have actually reached the tax bureau before closing the books

The outage on November 2 deserves a bit more discussion. The portal went down, submissions timed out, and the interface showed failure. But fiskaly's explanation made it very clear: some requests that displayed as failed may already have been delivered. Users should first confirm whether they have been registered, and then decide whether to resubmit; otherwise, it would be a duplicate submission.

This is the typical form of compliance-related features: The error is not 'it was not reported,' but 'it is uncertain whether it was reported.' The former can be remedied; the latter makes people guess between two errors—resubmitting is duplication, while not resubmitting is omission.

Compliance is not a module in the system; it is a check that stores must complete every day before closing. What the system can do is make this check take only thirty seconds.

IV. Without Data, AI Has Nothing to Compute

On the same day there was another message, about another matter, but pointing to the same conclusion.

Spanish grape raw material producer JULIAN SOLER has adopted Forterro's MyForterro.ai in Abas ERP. The company manages more than 1,000 materials, runs 30 to 40 production runs per day, and has customers in more than 60 countries, serving those making soft drinks, wine, vinegar, confectionery, and baby food. Recipes used to be calculated manually in Excel; now an optimizer calculates feasible raw material combinations and ranks them by cost and “degree of compliance with customer specifications.”

What are the criteria? Concentration, acidity, pH, color value—all things that need to be measured in the lab. The announcement quotes a sentence from the company side, Antonio López Soler, which I think is the most down-to-earth sentence in all the AI news this week:

Without production, quality, and laboratory data, this plan simply cannot be done.

This sentence makes the order clear. It is not about choosing the model first; it is about first turning the criteria into fields that can be queried in the system. Many manufacturing companies making similar attempts get stuck at the same place: the data is in the lab notebooks, in the team's Excel files, in the experienced workers' judgment—but it never gets into the system. At that point, no matter how powerful the model you deploy, it can only guess.

The performance numbers should also be read by the same standard. The announcement says it expects to save 4 to 8 hours per day on formula calculation and review, and to improve by up to 10% after incorporating raw material costs into selection. But the baseline for the 10% is not provided, nor are post-launch measured numbers. The project started as a two-day hackathon; afterward, the client's own employees configured the agents, refined the prompts, and selected models—this detail is more instructive than the numbers: the reason it could work is that this company's own people know how the formulas should be calculated.

V. Write evidence on the acceptance form, not completion

If you connect these three things, you will find that they answer the same question: on what grounds can it be said to be correct.

SAP's answer is a seven-part review checklist; the answer in this Odoo Italy news story is “not yet possible to say,” because the three matters of cutover, failure, and reconciliation are not clearly described; the answer in that Forterro case is “only when the data is there can calculation be discussed.” All three answers point to the same thing—completion is an action, correctness is a record.

Change the acceptance form to three sentence patterns, and the effect is immediate:

Replace “Data has been migrated” with “Record count XX, total amount XX, 0 discrepancies from the source system, and 3 exceptions confirmed by Engineer Zhang from Finance”.

Replace “Feature is live” with “The full order-to-cash process has been run through for 12 orders, including 2 partially shipped orders, 1 interface rejection, and 1 credit hold; the business signatory has signed.”

Replace 'compliance has been connected' with 'for 10 consecutive business days, verify the number of submitted records and amounts before daily closing, with 0 discrepancies.'

6. Three Things You Can Do Next Week

First, borrow the seven-part checklist and use it, whether or not you are an SAP customer. System analysis, data, code, configuration, testing, project, rollout—these seven parts almost cover the actual work scope of all ERP projects. Run your own project through it and see which part has only “done” but no “evidence.”

Second, ask four qualification questions for each AI feature. Target version, migration path, available sub-features, modifiable scope, and approver. Only if all four can be answered does it enter the candidate list.

Third, define a "closing action" for each compliance item.It's not "we've connected to the tax bureau," but rather "who, at which screen, checks which two numbers, and how discrepancies are handled before closing each day." Writing this step down is more useful than buying any compliance module.


Back to that sign-off sheet from the acceptance meeting. There is nothing wrong with the sheet itself; the problem is that it only has checkmarks and no evidence. In this year of AI entering ERP, what should be changed most is not the system but this sheet—from 'Is it done?' to 'What proves that it is done and correct?'

Only a project that can answer this question would dare to say it has finished the swap.

Sources: ERP News (interview with Sonja Liénard on 2026-11-11; reports from Odoo Italia and fiskaly on 2026-11-12; compilation of announcements from Forterro and JULIAN SOLER on 2026-11-13), SAP Sapphire 2026 and October 28 update materials, and Odoo November 9 release notes. This image is a conceptual illustration.

Guangzhou Zhongdi Information · Enterprise Digitalization Insights · 2026-11-13

关于我们

​我们致力于帮助中小企业实现数字化转型,我们的团队由一群充满激情和创新思维的专业人士组成,他们具备丰富的行业经验和技术专长。

扫一扫获取顾问以及手册

Sign in to leave a comment
Before an agent touches an order, first work out who it will affect.