We discovered $2.1M of hidden COPQRead the case study

Blog · Perspectives

The Last Migration (and how it caused a gummy bear shortage)

July 7, 2026 · Shreesham Mukherjee · Forward Deployed Engineer

Building the Great Turret
Joseph Pennell, Building the Great Turret, 1916. Lithograph.

8%

Of migrations finish on time

18–36

Months for a typical program

~1/3

Of ECC customers have moved

The bill comes due December 31, 2027.

That’s the day SAP ends mainstream maintenance for ECC, the ERP that runs an enormous share of the world’s manufacturing. The move to its successor, S/4HANA, is marketed as an upgrade. It isn’t one. It’s a re-implementation of the system your business runs on: a different database, a different data model, every table mapping and custom program reconciled by hand. Industry benchmarks put a full migration at 18 to 36 months. A 2025 Horváth study found that 8 percent of these projects finish on time. Only about a third of ECC customers have completed the move, and by several ecosystem surveys a majority haven’t started. Consultants know it, which is why rates climb every quarter the deadline compresses demand into the same shrinking window.

So before you sign the biggest IT contract of your decade, it’s worth asking a question that sounds naive: what exactly does the money buy?

I have an unusual amount of context on that question, because I’ve spent my career doing the work you’re about to pay for.

The Intelligent Part of an ERP Is the People Feeding It

I was a data engineer at Amazon Robotics, a tech lead at TeePublic (second-highest performing print-on-demand site for artist-to-customer reach), and a data engineer at Palantir, where the job description boiled down to one sentence: connect to every system a large enterprise owns and make the data mean something. Different industries, identical discovery. The system of record is not a brain. It’s a very strict filing cabinet, and all the intelligence lives outside it: in the people who pre-chew reality until the cabinet accepts it.

Consider what “using SAP” actually means day to day. Nobody writes directly to it. Ask why, and even SAP’s own ecosystem will tell you that writing to a live ERP risks corrupting records, so everything moves through a staged pipeline. Extract from the CRM, the MES, the planner’s spreadsheets. Transform until it fits structures the system will tolerate. Load. Reconcile. Repeat forever. That pipeline has a software-shaped name — ETL — but it is mostly labor: mapping workshops, validation cycles, middleware licenses, and consultants whose entire careers are translation between how your factory works and how the ERP insists it should.

Teammates of mine lived this from the inside at a manufacturer. Operating the system of record meant memorizing transaction codes like incantations such as MB51 just to see incoming material movements. And when the architect who had spent two decades building their SAP landscape retired, he left people with a quarter of his experience to untangle the web he’d woven. Every plant has this person. Many have already lost him.

The Migration Bill Is a Labor Bill

Scale that up and you get the real anatomy of an S/4HANA program. Decades of duplicate vendors, obsolete materials, and inconsistent BOMs can’t be carried into S/4HANA’s stricter model as-is. They have to be cleansed, deduplicated, and remapped, and data preparation alone can consume a year of the program. Thousands of custom ABAP programs have to be inventoried one by one, checked against the new data model, and rewritten or retired, usually without documentation and often without the people who wrote them.

The industry’s own numbers tell you how this goes. Data work gets treated as the last 10 percent of an ERP project when it’s really the first 40. Gartner predicts that by 2027, more than 70 percent of recently implemented ERP initiatives will miss their business case, with a quarter failing catastrophically. And when it fails, it fails in public. Lidl walked away from its SAP program after seven years and roughly €500 million — and went back to its legacy system. Revlon’s go-live left $64 million in orders it couldn’t fulfill, and its own shareholders suing. Haribo’s S/4HANA cutover put gummy bears out of stock and cut sales of its flagship product by a reported 25 percent.

Read the post-mortems and the pattern is consistent: the software mostly worked. What failed was the assumption underneath it: that armies of humans could re-shovel decades of operational data into a new schema, on schedule, without dropping anything that matters.

Why We Ever Agreed to This

The monolithic ERP was a rational answer to 1992’s constraints. Integration was expensive. Compute was centralized. The only way to get one view of the business was to force every process into one schema and make everyone speak its language. That was the bargain: contort your operations into the system, and in exchange you get a single place to ask questions.

Both halves of that bargain have now collapsed. The contortion is more expensive than it has ever been: that’s the migration quote sitting on your desk. And the questions that actually decide whether you make money — can we take this rush order? what happens to delivery if that machine is down for a week? which of these BOM changes breaks the schedule? — were never answerable inside the ERP anyway.

“It records. It does not reason.”

What Changed

AI is genuinely good at the part of this that used to consume the man-hours. Reading a messy export. Proposing how fields map. Spotting that these two vendors are the same vendor, that this routing contradicts that BOM. Agents can now read data where it lives (ERP exports, MES logs, spreadsheets) and negotiate schema at read time instead of demanding that the enterprise reorganize itself first.

What AI shouldn’t do is make those calls unsupervised. The expensive part of ETL was never typing the transforms; it was deciding them, and trusting the result. That’s why the model that works is human-in-the-loop: the agent does the shoveling, shows its work, and a person who knows the plant approves the judgment calls. This is no longer a fringe position: McKinsey is publishing essays titled “The end of ERP as we know it,” and Gartner expects a third of enterprise software to ship with agentic AI by 2028.

What This Looks Like When It Works

This is the premise we built ProDex on. Dexter, our agent, connects to the sources you already have: the ERP’s exports, the MES logs, the planner’s spreadsheet that actually runs the plant. ETL becomes a reviewable conversation. He proposes the mappings, flags what won’t reconcile, and shows you the diff before anything counts.

One manufacturer we work with learned that the export layer of their ERP had been silently dropping about a third of their BOM lines. Years of pipelines had never flagged it — pipelines don’t get suspicious. An agent with a human in the loop caught it on the first pass.

The payoff isn’t cleaner rows in a stricter database. It’s your operation reconstructed as something you can use: BOMs, routings, work centers, and inventory rebuilt into a living model of the factory, one you can simulate, plan against, and interrogate. Not “where is the data stored,” but “what should we do.” In weeks, not fiscal years.

Keep the Ledger. Lose the Monolith.

I’m not telling you to delete SAP. Finance, payroll, compliance — systems of record exist for good reasons, and if the ledger must move to S/4HANA, move the ledger. But be honest about what the rest of that 18-to-36-month program is: rebuilding your decision layer inside a system that was never good at decisions.

The system of record and the system of decision are different things. The first might be mandatory. The second, as of about two years ago, is a choice.

The 2027 deadline is real, and it’s a useful forcing function — just not the way SAP intends. Before you commit your best people to two years of feeding the monolith, ask what all that labor was ever for. I spent my career moving data into systems like this. The most useful thing that experience taught me is how little of it deserves to survive.

If 2027 is on your calendar, we’ll show you what your data looks like as a decision layer instead of a filing cabinet.

Book a demo

Sources

All entries