The brief these systems were designed against
ERP was designed for a world that no longer exists: compute was scarce and expensive, work was batched, integration meant a nightly file, and a screen was something a clerk in one department used to enter transactions correctly.
Against that brief it was brilliant, and the thing it solved is hard: a system of record. One place where the transaction is entered once, checked, posted, closed and auditable years later. Financial accounting, payroll, statutory reporting, the ledger a regulator will ask about. That problem was solved, and it stayed solved.
The trouble started when vendors stretched that shape over everything else a company does. The rest of what a company does has almost nothing in common with recording a transaction.
Where the design brief runs out
Routine recording ends and something else begins: the work that needs judgment, argument, outside information and speed.
Planning and forecasting, where the useful version changes three times before it is agreed. Negotiating an exception nobody anticipated. Qualifying a lead with information that lives outside your walls. Deciding together, in a thread, with the numbers in front of everyone. Reacting to something today rather than in the next release cycle. And now: putting an AI employee on the parts of it that never needed a person.
None of that fits a system designed around a correctly entered transaction. The vendors did not fail; a design brief from a different era is being asked to hold work it was never shaped for. In that half of the operation, a 1980s architecture is a dinosaur, and everyone using it knows.
The evidence is already on everybody's screen
Look at where that work happens today.
It happens in spreadsheets. In a messenger thread. In a third-party tool somebody expensed. In a
document that gets re-attached to an email with _final_v3 in the name.
That is the market's verdict, reached by every company on its own: the flexible half never succeeded inside the ERP, so people built it themselves, outside, in whatever was to hand. It has been living in exile for thirty years, and everyone has agreed not to mention it.
So is this an ERP? Yes and no
The honest answer has two halves.
No, and we are not asking you to move your core. You spent years and real money making financial accounting reliable and compliant. Nobody sane rips that out for a better user interface. Financial accounting stays in SAP or QuickBooks. Payroll stays where payroll is. Statutory reporting stays where the auditors already understand it. That is the ERP core, it works, and it should live where it lives.
Yes, for the other half. Everything that needs flexibility, forecasting, enrichment from outside, collaboration and AI - the half that has been in spreadsheets because it had nowhere else - gets a real home: a described operation with processes, permissions, history, analytics and agents, on a stack designed after the internet.
The two connect instead of duplicating. Your ERP becomes a source like any other: read where it stands or extracted on a schedule, joined to everything else, reported on beside it. What it is authoritative about stays authoritative. See Sources and storage.
The SaaS stack has the mirror problem
The other common answer is a dozen subscriptions: one for the CRM, one for tickets, one for approvals, one for dashboards. Each is better at its job than the ERP module would have been.
Together they are five systems that do not know about each other, so somebody spends their career keeping identifiers matched and lists in step, and the answer to "what is happening" depends on which of them you ask.
| The ERP | The SaaS stack | AIpril | |
|---|---|---|---|
| Fit to your operation | Its process, adapted at a price | Each tool fits its own job | Described, and changed by editing the description |
| Between the parts | Coherent inside its boundary | Integration, forever, as a running cost | No seams: one org, one catalogue, one data model |
| One version of the truth | Yes, inside its boundary | Whichever system you asked | One definition, read by every surface |
| Changing something | A project with a business case | Fast in one tool, a negotiation across five | A version, reviewed and installed |
| Upgrades | Where customisation goes to die | Done to you, on the vendor's schedule | Immutable versions, installed when you choose |
| Where the knowledge lives | A settings tree nobody reads | Twelve settings trees nobody reads | A description a person can read |
Where the alternatives still win
- A regulated core with a certified process. If your industry requires a specific configuration audited by a specific body, the system that arrives certified is worth what it costs.
- A category with a deep specialist tool. Precise tax filing, an industry design suite, a trading platform. Do not rebuild it; connect it.
- Very small and stable. Ten people and two straightforward processes do not need this much machinery.
- An implementation nobody argues with. If your ERP fits, the case for moving is weak. The phrase to listen for is "we work around it": that is the part being discussed on this page.
- The specialist tools people love. Nobody is giving up Slack because a platform has a comment box, and nobody should. Apollo is excellent at what it does with company and contact data; your designers are not leaving their design tool. These are products that are better at one thing than any suite will be, so the right response is to connect them. A message into the channel people already read, a contact enriched as the record is created, a file that arrives from where it was made.
- Maths, models and data science. No built-in algorithm is going to out-compete the Python ecosystem. Cleaning, feature engineering, forecasting, optimisation, anything with a research paper behind it: that work belongs where the libraries are, in the hands of the people who choose the method. What the platform offers is the connection. Your model runs wherever you run it, a process or a schedule calls it as an ordinary step, it gets the data through the same query contract everything else uses, and it writes its result back into a record, a column or a balance the rest of the operation can act on.
The thread through those two: the modern answer is many good tools, connected cheaply, not one system carrying somebody's idea of best practice for every job in the company. A platform earns its place as where the pieces meet: one description of the operation, one org structure, one set of permissions, one audit trail, and integrations that cost a step in a process instead of a project.
What this looks like in practice
Almost nobody replaces everything, and nobody should. The pattern that works:
The system of record stays. The operation that is yours - the processes with your exceptions in them, the planning that lives in spreadsheets, the collaboration happening in a messenger, the reporting somebody rebuilds every month - moves onto one described platform, with the core connected to it as a source.
That turns the thirty-year workaround into a system: still flexible, now with history, permissions, analytics and AI employees in it, and still talking to the ledger the auditors trust.
Time to value
None of the above matters if getting there takes two years, and that is the part of this argument with the shortest half-life.
Nobody has two years any more. Companies assemble something, work in it, change it when the market changes, try an idea, connect a tool that did not exist last quarter and drop one that stopped earning its place. A programme whose first useful day is eighteen months out is answering a question that will have changed by the time it finishes.
There is no implementation phase, because there is nothing to implement. The first working version exists once you approve a specification: after a conversation and a review, not after discovery, a build and a UAT cycle. You then correct it while looking at it, which is the only review that has ever worked.
And no warehouse to build first. The data plane is already there: connect the sources and ask. The months that usually go into modelling, pipelines and a BI layer before anyone sees a number are not compressed here; nobody spends them. See Sources and storage.
The cost follows from that. What dominates the budget of these programmes is people-time on work that does not exist here: an integrator's discovery, custom development, the integration layer, the reporting build, the year of change requests afterwards. That is not a discount on the same line items; most of the line items are absent.
And you are not starting from a blank page. How your company works is already written down: in the spreadsheets people plan in, the processes already running, the catalogues already maintained. Point the assistant at them, and a spreadsheet becomes a dataset, the scan describes what is in it, and a model and a process come back proposed over it. That first version is something to argue with, not a questionnaire to fill in.
Continuity here is the shortest path from what you already do to a described version of it, written by the assistant and corrected by the people who know better.
And it stays cheap afterwards, which is the half people forget. Fast to build and expensive to change is the trap every low-code demo walks into. The system is a description rather than a build, which is why the first version arrives in weeks and why the eleventh does too.
Next
AIpril vs low-code and BPM - the tools you would have reached for to close that gap, and why the result usually ages badly.