“Data migration was one line on the quotation. I looked at it and figured that would be quick.”
The owner who said this got stuck in month three of implementation. Not on system features. On the batch of orders in his old system that were “not finished yet” — shipped but not invoiced, half delivered, deposit paid up front, sitting there for two years with nobody willing to touch them. What status should these orders carry into the new system? Three meetings, no conclusion. Before that, he thought data migration meant export from the old, import into the new. A weekend's work.
I'm Johnny Yang, Brightstream Technology in Taiwan. I've led implementation teams for seven years, across 252+ projects in 32+ industries. So I'll state this flatly up front: ERP data migration is the most underestimated item in the entire implementation plan, and it gets underestimated because people think of it as a “moving” job. It is actually a chain of decisions. The time never goes into the technical export-and-import step; that usually runs in a few days. The time goes into deciding, before the move, which data to bring and what shape it takes once it's in — and, after the move, whether anyone is willing to put in the effort to check it, verify it, and sign off on it. And you have to accept something counterintuitive first: for a good share of your old data, the right treatment is not to move it. Moving it doesn't make you safer. It copies the mess in your old system into the new one, untouched, and what you get is a database that looks newer but is just as untrustworthy.
To be clear, “cost” in this post is not the quotation. It's time, extra meetings, and how many people have to stay late reconciling numbers in the month you go live. None of these three appear on any quotation. They are the real price of data migration.
You're right: the data has to move — but moving isn't one step
Your instinct to worry about old data is correct. Customer tax IDs, contacts and delivery addresses took over a decade to build up. Part numbers and specs are the shared language of the whole company. Who still owes you money is how you judge the business. If these can't get into the new system, you've wiped the company's memory and started over. So “the data has to come across” is a fair requirement. I never tell clients to give it up.
What's wrong is thinking of it as one step. In the owner's picture, data migration is “an engineer writes a script, runs it overnight, and the next morning the data is in the new system.” Half of that picture is true — the script does run, and it runs fast. But before that night there's a month of work nobody sees, and that month is spent answering questions.
The same customer has three records in the old system because sales typed the name three different ways. Merge into one or keep three? If you merge, whose address do you keep? The old system's customer tiers have five values — A, B, C, VIP, old customer. The new system's tiering logic doesn't look like that. How do you map them? Old part numbers have specs mixed into them; the new system wants them split into a code plus attributes. What happens to the numbers on old documents?
None of these questions is hard. What's hard is that there are hundreds of them, and only people inside your company can answer them. The consultant can't. So data migration doesn't eat engineer hours. It eats decision time from you and your managers. This is especially easy to misjudge when reading a quotation. I made the same point in “How to read an ERP quote”: the line with the fewest words often takes the most of your own people's time.
Three kinds of data, three levels of difficulty — don't lump them together
“Can our old data come across?” can't be answered directly, because “old data” is at least three completely different things. Only when you separate them can you judge whether this item is going to blow up.
The first is master data. Customers, suppliers, products, part numbers, employees, warehouses. These are the nouns — the words the new system uses to build sentences. Master data is the easiest technically and almost always makes it across. But it's the hardest to clean, and if you don't clean it, it drags down everything after it: dirty master data means every transaction you move hangs off the wrong noun. So the real work on master data isn't migration. It's deduplicating, filling blanks, and standardizing naming rules. No tool can fully automate it, because deciding “are these two records the same customer” takes a person.
The second is historical transactions. Past orders, delivery notes, purchase receipts, invoices, payment records. They're high volume but simple — they're finished, they won't change, and the new system doesn't need to calculate on them. It just needs to be able to look them up. The hard part isn't technical. It's how much you want to move and how rough a version you're willing to accept.
The third is open orders. The most troublesome and the most often ignored: ordered but not shipped, shipped but not invoiced, invoiced but not paid, partial deliveries, deposits received or paid in advance, work in progress, inventory discrepancies from a count that was never finished. What they share is that they're still alive — once the new system takes over, it has to keep acting on them. It has to offset them, close them, and age them.
What makes them hard is that they're not just data. Each one is also an accounting entry. An order shipped halfway: when it enters the new system, does it show how much is still undelivered? Does the half already delivered have a shipping record? If not, anyone looking up that order later sees a broken story. If it does, you've also moved part of the transaction history, and it has to reconcile with the general ledger.
In the implementations I've seen, when go-live gets pushed back, it's usually stuck here. Features tested, training done, and finance won't sign because the status of open orders doesn't reconcile.
| Data type | Technical difficulty | The real work | Common sticking point |
|---|---|---|---|
| Master data (customers, products, part numbers) | Low | Deduplicate, standardize naming, fill in required fields | Nobody has the authority to decide “are these two the same?” |
| Historical transactions (closed orders) | Medium, high volume | Decide how many years and how much detail to move | Want to move everything, end up verifying nothing |
| Open orders | High | Confirm status order by order and reconcile with the books | Finance can't reconcile, won't go live, the date keeps slipping |
“How many years of history should we move” is the wrong question
Every project gets this question, and when the owner asks it, he usually already has an answer in mind: “All of it.”
I understand the reaction. Moving everything sounds safest. But that answer hides an unchecked assumption: that data you've moved is data you can use. Historical data that's been moved but never verified is data you still won't trust for decisions. Verifying takes people, and moving ten years means verifying ten years. Most companies run out of energy by year two. The other eight years become a pile sitting in the database that nobody looks up, and that nobody would trust if they did.
So the right question isn't “how many years.” It's these three.
One: what do you plan to use this data for? There are usually three answers, and each one means a completely different way of moving it. If it's “when there's a complaint or warranty claim, I need to find what we sold to whom,” you need document-level detail, but probably only covering the warranty period plus some buffer. If it's “I want to see trends, how customers have grown or shrunk over the years,” you need monthly summaries by customer and by item, not every single order — an order of magnitude less data, and much faster to verify. If it's “I'm worried about a tax audit later,” the answer often isn't in the new system at all. More on that below.
Two: if you don't move it, can you still look it up? This question sets the real size of the risk. If the old system still opens and can still print reports, the cost of not moving is “it's a bit more annoying to look up,” not “we can't find it.” Those are very different. Most companies overestimate this risk because they never seriously checked whether the old system can be kept around for lookups.
Three: once it's in, who's going to check it? You say move five years. Who's going to confirm those five years of numbers are right? If nobody has the time to check, the only difference between five years and two is database size. Not a single year of it is trustworthy.
Answer these three and the number of years surfaces on its own, usually fewer than the owner said at the start. Most of the projects I've worked on land on the same layered approach: move all open orders, move recent transactions in detail, move older ones as summaries, and leave the oldest in the old system or export them to an archive. The benefit is that your limited verification effort goes into the stretch you'll actually use.
This decision also directly affects the go-live date. It's one of the few variables where a single sentence from the owner can change the length of the project. How to estimate the overall timeline and which pieces affect each other, I break down in more detail in “How long does an ERP implementation take.”
For some data, the right answer is not to move it
This is probably the least popular section in the post, but it's where you save time.
Don't move master data you no longer use. Customers who haven't ordered in five years, part numbers discontinued long ago, codes for salespeople who left three years back. Moving them into the new system gets you three extra pages of scrolling in every dropdown. Worse, new staff can't tell which ones are live and will pick the wrong one. These should be archived, not migrated — export a copy, keep it somewhere, and make sure it can be found.
Don't move wrong data. Sounds obvious, but few people actually do it. Inventory discrepancies you know are off, receivables balances that have sat unreconciled for ages, adjustment entries made back then to make reports look better. Move them over as-is, and the books are wrong on day one in the new system. From that day on, everyone will say “that's an old data problem,” and that line becomes the all-purpose excuse covering everything for the next two years. Better to clear them before go-live with a one-time opening balance adjustment and a written explanation than to let errors seep into the new system.
Data kept only for tax or regulatory reasons goes into an archive, not the system. For many companies, the real motive for moving history is fear of a future audit or tax inspection. But that need is “can be produced, can be read, can't be tampered with” — not “searchable in the ERP.” Export it to a fixed file format, attach a document explaining how to read it, and store it properly. That's enough. Using the ERP as a safe deposit box is expensive and awkward.
Translate the code words in the notes field before moving them. When your sales support staff write “send drawing before shipping for this one” or “this customer doesn't take deliveries on Wednesdays,” that isn't data. It's rules. Copy it straight into the new system's notes field and you've just moved the same reliance on people's memory to a new place. What you should do is pull them out and see which can become fields, flags, or validation checks. Skip this, and six months after go-live everyone is still writing code words in the notes field.
The time goes into verification, not import
Draw the data migration workload as a line, and import is a short segment. Before it is cleanup and decisions. After it is verification and sign-off — and the part after is the part most often skipped entirely.
Verification isn't “open it up and see if there's anything in there.” It has to reconcile: record counts, total amounts, and a random sample of orders compared field by field. A machine can calculate the first two. The third has to be done by a person, and by the person who uses this data day to day — not IT, not the consultant. Only they can see “this number made it in, but its meaning has shifted.”
Even more often skipped than verification is sign-off. After verification, someone has to put their name to “I accept these numbers, and if there's a problem later I won't blame data migration.” Without that, the new system goes live with a built-in disclaimer. Anything that doesn't match can be blamed on the data move, and you'll never find the real cause. One email or one confirmation form is enough. What matters is that someone signs.
And because verification needs people, it can't be crammed into the week before go-live. The sensible approach is to rehearse: move the full set into a test environment, let the real users search it and find mistakes, and leave time to fix them. The production move should be the second or third run. Quite a few of the failed implementations I've seen failed because the first move was the production move, and when problems showed up, they had to fix them while running the business.
So what should you do tomorrow
You don't need to wait for a consultant. Tomorrow you can do two things yourself, and once they're done your read on this item's risk will be completely different.
First, ask your finance staff or sales support staff one question: “How many orders do we have open right now? When is the oldest one from?” If it takes them a long time to find out, or they answer and then add “though a few of them are weird,” you've found the part of this implementation most likely to drag. Pull out those weird orders and deal with them before the implementation starts.
Second, open the customer master in the old system, sort by company name, and look for two or three records that look very similar. Then look at the last-transaction-date column and count how many haven't moved in three years or more. Those two numbers tell you whether master data cleanup will take a week or a month.
Go talk to vendors with these two answers in hand and the conversation is completely different. You're no longer asking “can the data come across?” You're saying “here's how I plan to handle these categories — how will you work with that?” The timeline estimates those two questions get you are worlds apart in accuracy.
If you already have an implementation plan but can't tell whether the data portion is broken down finely enough, talk to me for thirty minutes and we'll scope out the three kinds of data first.




