July 1999. The warehouses are full of candy and the orders cannot get out of them. Hershey has just switched on three enterprise platforms at once, SAP R/3 for the core system, Manugistics for supply chain and Siebel for customer relationships, and the order-to-ship path has stopped working. By September the problems are still not fixed. Roughly $100 million of orders go unfilled, with the product sitting in inventory, through the single most important selling season the company has. Third quarter sales fall by around twelve percent and profits by roughly nineteen. In its annual report for that year the company attributed the drop in shipments to order fulfillment difficulties following the startup of the new integrated system.
The detail that matters is the one everybody skips. The project had begun in 1996 and the recommended implementation window was somewhere near forty-eight months. It was compressed to around thirty, driven by a Y2K deadline, and the phases that got abbreviated were the testing phases. Then the go-live was scheduled for July, immediately ahead of peak season.
None of that was a software failure. Every one of those was a management decision, and they are the same decisions that are being made right now in mid-market manufacturers with a tenth of Hershey’s resources and none of its ability to absorb the consequences. If you are earlier in the sequence than this, the prerequisite work is in the guide to modernizing a family-owned manufacturer, and the specific thing you should do before touching a system is in tax books versus management books.
Why do ERP projects actually fail in mid-market manufacturers?
Not because the software is bad. They fail because the company encodes a process it never examined, compresses a timeline to hit a date that was invented internally, and gives ownership to whoever has the most spare capacity rather than the most authority. Three management failures wearing a technology costume.
The encoding failure is the most expensive and the least discussed. An implementation takes how you currently work and makes it permanent in software. If your quoting rule is undocumented, your overhead allocation is unsourced and your scheduling lives in one person’s judgment, you have just paid several hundred thousand dollars to fossilize all three. The vendor will not stop you. The vendor’s job is to configure what you describe, and what you describe is what you already do.
The timeline failure follows from the first. Because nobody examined the process, the configuration phase surfaces a hundred decisions that should have been made months earlier, and those decisions now have to be made under deadline pressure by people who are also doing their day jobs. The date holds. The examination gets cut. Then testing gets cut, because testing is the only phase with no external party demanding it.
The ownership failure is the quiet one. This project gets assigned to the controller or the operations manager as an addition to their existing responsibilities, which means the person accountable for the largest capital commitment in the company’s recent history has no authority to change anything outside their own function.
Watch a related video: Transformation Failure: Why 70% Die at Month 14.
What should you fix before you buy anything?
Three things, none of which cost software money. Document the process you intend to encode. Rebuild the numbers the system will report on. And name a single owner with real authority. If you cannot do those three, you are not ready, and buying the system will not make you ready.
Start with documentation, and be honest about what that means. Not a flowchart drawn in a conference room. The actual rule: how a quote gets priced, what triggers a reorder, how a job gets scheduled, what happens when an order changes after release. Write it down, then have the person who actually does it read what you wrote and correct it. That exercise alone routinely surfaces two or three practices nobody in management knew existed.
Then fix what the documentation reveals. This is the step that gets skipped and it is the entire point. If the overhead rate is stale, fix it now, in a spreadsheet, before it becomes a configuration parameter that costs eleven thousand dollars and six weeks to change later. If quoting depends on one person’s judgment, convert it to a rule now. The system should inherit a corrected process, not a photographed one.
Then name the owner. One person, senior enough to overrule a department head, with a meaningful share of their time formally released from other duties. If nobody in your company can be freed up, that is a finding about your readiness, not a staffing inconvenience to be worked around.
How much system do you actually need?
Almost certainly less than you are being sold. The instinct is to buy the platform that covers everything, because piecemeal feels unstrategic and because every vendor demo is built to make the full suite look inevitable. In a manufacturer under a hundred million dollars, the opposite is usually correct.
Here is the arithmetic nobody runs. A full replacement consumes essentially all of your organization’s change capacity for twelve to eighteen months. During that window you will not be fixing pricing, will not be rationalizing your product range, and will not be integrating an acquisition, because the same handful of people would have to do all of it. That opportunity cost is real and it is almost never put in the business case next to the licence fee.
Meanwhile the benefit is concentrated. In most industrial businesses the overwhelming majority of the value comes from a small number of capabilities: accurate job costing, inventory visibility, and a scheduling function that reflects actual capacity. Everything else is convenience. Buy the concentrated part, get it working, and revisit the rest in three years when you have earned the ability to absorb it.
| Approach | Best when | The real cost |
|---|---|---|
| Full replacement, single vendor | Current system is genuinely unsupported, or you are integrating multiple acquired entities | 12 to 18 months of total change capacity. Highest failure exposure. |
| Core modules only, phased | Most mid-market manufacturers, most of the time | Slower total coverage. Some integration work you own rather than the vendor. |
| Keep the core, add point solutions | The ledger works and the gap is scheduling, quoting or shop floor data | Integration debt. Gets ugly past three or four bolt-ons. |
| Fix the process, buy nothing yet | Process is undocumented, numbers are unreliable, or no owner exists | Nothing but discipline. This is the right answer far more often than it is chosen. |
What does it actually cost, all in?
Far more than the quote, and the gap is predictable enough that you should budget for it deliberately rather than discovering it in month seven. The licence figure a vendor leads with is usually the smallest of the four numbers that will ultimately be spent.
Four cost categories, and only one of them appears on the proposal. Software licence or subscription is the visible one. Implementation services, meaning configuration, data migration and training delivered by the vendor or a partner, typically runs at some multiple of the first-year licence rather than a fraction of it, and it is the number to negotiate hardest because it is where scope ambiguity lives. Third is internal labor, which almost nobody costs: the hours your own people spend specifying, testing, cleaning data and learning, which are real hours diverted from real work even though they never appear as an invoice. Fourth is the productivity dip around go-live, which is a genuine cost and which you should put a number on in advance.
Then add the contingency the proposal does not contain. Data migration is the reliable overrun, because legacy master data in a thirty-year-old manufacturer is always dirtier than anyone estimates and the cleanup is your job, not the vendor’s. Integrations to whatever you are keeping are the second. Custom development is the third, and the discipline worth adopting is that any customization request has to be approved by the project owner against a written reason, because each one is a permanent tax on every future upgrade.
The number that actually matters is not the total. It is the total divided by the capability you are buying. A phased implementation covering job costing, inventory and scheduling delivers most of the available benefit for a fraction of the full-suite figure, which is why the scoping question in the previous section is worth more than anything you will negotiate on price.
How do you choose without letting the vendor run the process?
By writing your requirements from your documented process before you take a single demo. The moment you let a vendor define the evaluation criteria, you are comparing products on the dimensions where that vendor is strong, and you will not notice because the criteria will all sound reasonable.
Build a short list of things this system must do that are specific to your business, expressed as scenarios rather than features. Not “supports job costing” but “quote a three-operation part with an outside heat treat step and show me actual versus quoted margin when it ships.” Give every vendor the same five scenarios and make them demonstrate on your data, not on the demo dataset where everything works.
Then do the thing almost nobody does: call three references who implemented the same modules in a manufacturer your size, and ask them a specific question. What did you find out in month nine that you wish you had known in month one? References are coached to answer general questions. They are rarely coached for that one.
And read the implementation contract harder than the licence contract. Licence pricing is negotiable and comparable. Implementation scope is where projects go wrong, and the clause that matters is what happens when the hours run out, because the hours will run out.
What does the implementation sequence actually look like?
Phased, always, and ordered by how much damage a failure in that area would do. The Hershey pattern is the counterexample: three platforms simultaneously, on a compressed schedule, going live into peak season. Every element of that is reversible as a decision, and every element of it was avoidable.
Sequence by blast radius. Financial and back-office functions go first, because a two-week problem in the general ledger is painful and does not stop trucks. Inventory and purchasing go second. Order management, scheduling and shipping go last, because that is the path that touches customers, and a failure there is the one that ends up in your trade press.
Inside each phase, run parallel. Yes, it is duplicated work for a period and people will complain about it. It is also the only mechanism that lets you discover the system is wrong without discovering it through a customer. Set an explicit exit condition for the parallel period, published in advance: the new system matches the old one on a defined set of transactions for a defined number of consecutive days.
And pick the go-live window against your own seasonality rather than against a fiscal date. If your heaviest shipping months are the third quarter, do not go live in June because the budget year says so. Hershey’s July go-live was chosen against a calendar deadline and executed directly into the ramp for its biggest season.
Who inside your company has to own this?
One named executive with authority across every function the system touches, with at least half their time formally released, and with the explicit power to move the date. That last one is not negotiable and it is the one owners resist hardest.
Authority to move the date is what separates a project that can be run honestly from one that cannot. If the owner can only report status against a fixed date, then every problem discovered late becomes a problem concealed, because the only lever they have is to compress whatever is downstream, which is always testing and training. Give them the date and they will defend it. Take it away and they will defend the appearance of it.
The steering group should be small and should meet weekly with the same one page: scope changes since last week, decisions outstanding, test results, and days of parallel running completed. Anything that cannot be discussed in forty-five minutes is a symptom.
There is a second role worth naming and almost never filled: someone from the shop floor, respected by their peers, given formal time on the project. Not to represent a department in a steering meeting, but to sit in configuration sessions and say that the proposed workflow will not survive a Tuesday. Every implementation I have watched go badly had a moment where someone who actually does the work could have said that, and was not in the room.
Your vendor’s project manager is not the owner. They are a supplier managing a scope of work, their incentives are attached to a contract, and no matter how good they are, they will not tell you that your business is not ready.
What happens at go-live, honestly?
Output drops. Every time, in every implementation, without exception. The only question is whether you told people in advance, because an expected dip is a project on schedule and an identical unexpected dip is a failed project.
Plan for it materially rather than rhetorically. Build inventory ahead of the cutover if your product allows it. Tell your largest customers, by name and in person, roughly six weeks out that you are changing systems and that you have built buffer for them. That conversation costs you nothing and it converts a potential crisis into a demonstration of competence.
Staff the first three weeks as an event, not as normal operations. Extra hands on order entry, someone floating on the shop floor who can answer questions immediately, and a daily stand-up at the end of each shift to collect what broke. The people who will find every genuine defect in your configuration are the ones using it at seven in the morning, and they will only tell you if there is an obvious, low-friction way to do so.
Resist the temptation to declare victory at go-live. Go-live is the midpoint. Adoption is the second half, and adoption is where the return actually lives.
How do you know if it is going wrong?
Four signals, all of which appear months before anyone says the word failure. Watch for them specifically, because the project’s own status reporting will not surface any of them.
The first is scope creep dressed as clarification. Requirements that were not in the original scope arriving as “that was always implied” is the single most reliable early indicator, and it usually means the original process documentation was thin.
The second is testing being rescheduled rather than cut. Nobody ever announces that testing has been eliminated. It gets moved to a later window, then compressed, then run partially. Track completed test cases as a number, weekly, and treat any week where the number does not move as a red flag.
The third is the vendor spending more time with your executives than with your users. That pattern means the relationship is being managed rather than the project being delivered.
Run these as a standing item rather than as an instinct. Four lines on the weekly page, reviewed in front of the same people every week, is the difference between noticing a pattern in month four and having it explained to you in month eleven.
The fourth is the workaround. When you find people maintaining a parallel spreadsheet because the new system does not do something they need, you have learned that the configuration is wrong in a way nobody escalated. Go and find out why they did not escalate it, because the answer is usually that they did and nobody listened.
When should you stop the project?
When the cost to finish exceeds the cost to abandon plus the cost of the alternative, calculated honestly and without regard to what has already been spent. Money already committed is gone whichever way you decide, and every month you avoid that arithmetic makes the eventual number larger.
The specific condition worth naming is this: if you are approaching a go-live and the parallel exit criteria have not been met, you stop and you move the date. Not compress. Move. Every catastrophic implementation in the public record shares that moment, where a team knew the system was not ready and went anyway because the date had acquired more authority than the evidence.
Stopping is not the same as abandoning. In most cases the right call is a pause of one or two quarters to fix the thing the project surfaced, which is nearly always a process nobody had examined. Companies that pause and resume generally finish. Companies that push through an unmet exit condition generally spend the following year explaining themselves to customers.
The conversation that makes stopping possible has to happen before the project starts. At kickoff, in front of the steering group, agree the conditions under which this project would be paused and write them down. A team that has been told in advance that a pause is a legitimate outcome will tell you the truth in month eight. A team that believes pausing is a career event will tell you the date is still achievable right up until the week before go-live.
And if you do abandon, extract the asset. The documented process, the corrected overhead rate, and the cleaned master data are worth having regardless of which system you eventually run, and they are usually the most valuable things the project produced anyway.
About the Stagnation Assassin
Todd Hagopian is a Fortune 500 transformation executive who has generated $3B+ shareholder value across five turnarounds at Berkshire Hathaway, Illinois Tool Works, and Whirlpool. He is VP of Product Strategy and Marketing. Known as The Stagnation Assassin, he is the author of two published books: The Unfair Advantage: Weaponizing the Hypomanic Toolbox and Stagnation Assassin: The Anti-Consultant Manifesto. His blog is published in 15+ languages and read by operators worldwide. Bring him to your stage via his speaking page or connect with him on LinkedIn. His newest book is 10-Minute Transformation (February 23, 2027).
Manufacturing and family-business programs book this material as a session on why systems projects fail and what the board should be asking before it approves one. See the speaking page for formats and availability.

