Standard software is someone else’s operating model

By the end of 2027, mainstream maintenance ends (opens in a new tab) for SAP’s older ERP generation, and many mid-sized companies will have to decide how to replace it. Most will compare standard products. But the real question is whether buying one still makes sense.

For a company with 250 people working in its core system, building its own AI-first platform can now cost less over five years than implementing a standard ERP. That is the smaller reason to consider it. The larger one is what you can change once the software is tailored to your needs. For organisations with more than 150 to 200 people in their core system, owning an IT system you can redesign your processes around no longer costs more than renting one your employees have to work around.

Every ERP comes with a way of working

Enterprise resource planning (ERP) systems are the software suites that run finance, purchasing, stock, and orders. For twenty years the advice was to buy one and change your processes to fit it. Building your own was slow and expensive, and what you built tended to become the legacy system nobody wanted to touch.

Take a hypothetical wholesaler of technical components with 250 people working in its ERP. An order arrives by email or as a PDF. Someone in the sales office types it into the system, checks stock and the customer’s credit limit, and releases it to the warehouse; later, someone in billing creates the invoice, and someone in accounts matches the payment when it arrives.

Every step happens in a screen the ERP vendor designed, in the order the vendor chose. Customising that is expensive and makes every upgrade harder, so the company adapts to the software instead. In effect, it has bought part of someone else’s operating model: how work is divided between people and systems, who approves what, and in which order process steps happen.

For a multinational with tens of thousands of users, that trade-off often works. Our wholesaler pays enterprise prices for licences, implementation partners, and upgrade projects, and still bends its work around a product built for someone else.

AI has changed what software costs and what it delivers

The first change is in the cost of building software. Senior software engineers now use AI throughout their workflow, from writing code to testing and documenting it. In our experience, a small, senior team can now build in less than a year what once needed a much larger team for several years, provided the scope is clear.

If a tailored system costs about the same to build as a standard one costs to implement, affordability stops being the deciding question. The harder question becomes ownership: who runs, maintains, and improves the system once it is live. Review, integration, and verification still take senior judgement, and upgrades, security patches, and support remain real work after go-live, so running costs fall less than build costs.

The second change is in the work itself. An AI agent can complete a defined task with some independence. At our example wholesaler, an agent can read an incoming order in whatever format it arrives, check stock and credit, book it, and pass anything unusual to a person. In an AI-first workflow, agents do low-risk routine steps like these and people make the decisions that need human judgement.

Standard software was designed for people entering data into screens, and most licences are charged per user. As agents take over routine steps, per-user pricing stops matching the value, and some vendors now charge for their AI features per action or per conversation, so your growing use of AI becomes part of their revenue. The vendor’s own AI features also automate the vendor’s process, which is often the process you wanted to change.

The licence saving is only half the story

Once you own your software, you can redesign your operating model from the result backwards. For the wholesaler, the result is a paid invoice. Orders are read and booked as they arrive, the invoice goes out when the goods are shipped, and payments are matched as they come in, so the sales office, billing, and accounts no longer pass each order between them. One small team owns the whole flow from order to payment and handles the cases the agents hand over.

A rough calculation shows why this matters more than the licence bill. Assume the wholesaler handles 300 orders a day, and 40 people in the sales office, billing, and accounts spend most of their time moving them along. At about €70,000 a year each, including salary, social charges, and overheads, that is €2.8 million a year, or roughly €37 of administration on every order. In the cost model below, owning the system saves about €235,000 a year in running costs.

An AI add-on that only reads incoming orders saves typing, but people still check, release, invoice, and match every order, so it frees perhaps six people’s time. A redesigned flow from order to payment goes much further. If eight in ten orders are booked without a person, every invoice goes out when the goods ship, and nine in ten payments are matched automatically, 20 people could do what 40 do today. Purchasing, supplier invoices, customer enquiries, and the month-end close can follow the same pattern, and freeing 40% of the time of the company’s 100 or so administrative staff would be worth the work of 40 people.

These are assumptions for an illustrative company, so try them with your own team sizes and volumes.

Two caveats apply. The larger gains in the second and third rows come from redesigning the operating model, and your own software makes that possible. Many companies own custom software and are no more efficient for it, because they rebuilt their old process in new code.

Freed time also turns into lower cost only if the organisation acts on it. For most, the realistic result is growth without hiring at the same pace. If orders grow by 10% a year, the old process needs about four more people every year, while the redesigned one absorbs that growth for several years and roughly halves the cost of administration per order. Cutting jobs en masse is rarely where the value lies, and it is a quick way to lose the people who know how your company, your clients, and the whole industry really work.

At 250 users, owning costs less. At 50, it doesn’t.

In our example, the wholesaler’s first option is a standard cloud ERP: €800,000 for a partner to implement it, then about €675,000 a year for licences at €100 per user per month, partner support, release testing, the vendor’s AI add-ons, and internal administration.

The second option is an AI-first platform for the operational core, with parts such as accounting, payroll, and e-invoicing rented from specialists. It costs €1 million to build over about a year, then about €440,000 a year for a two-person internal team, external support, EU hosting, AI models, and the rented services.

Building costs more in the first year and less in every year after. It becomes the cheaper option during year three and is about €600,000 ahead after five years, and the build could overrun its budget by 60% before buying came out cheaper over that period.

The case weakens in two places. Building costs barely change with the number of users, while licence costs rise with every one, so below roughly 150 to 200 users owning does not pay on cost alone; at 50 users, buying is clearly cheaper. The second weak point is the team that owns the system after the build: with three people instead of two, the five-year advantage shrinks to about €100,000.

Owning the system gives you the option to choose your AI

When you own your core software systems, you also own the decision about which AI runs each task and where your data goes. Reading thousands of orders and delivery notes is routine, high-volume work that a small model on a server you control can do well and at low cost. Drafting a careful reply to a customer whose delivery went wrong benefits from the strongest commercial model available, while anything involving employee data can stay on your own hardware.

Sovereignty here means the ability to decide, task by task, which data leaves the organisation and on what terms. Running everything yourself would usually cost more and give worse results. For European organisations the decision includes jurisdiction, since providers subject to US law can be required to hand data to US authorities even when it is stored in Europe.

Suppose the provider behind the wholesaler’s order-reading model raises its prices sharply, retires the model, or changes its terms on how customer data may be used. With its own system, the company tests the alternatives against last month’s real orders and switches within days. Inside a vendor’s suite, it waits for the vendor’s roadmap.

For technical readers

Four pieces make this work. All AI calls go through one gateway, so workflow code never depends on a specific provider and each task’s model is a configuration setting. Each task has a test set of real, solved cases from your own work, and a new model has to match or beat the current one on accuracy, cost, and response time before it is switched on. Personal data is removed or replaced before calls to external providers wherever the task allows. Open-weight models, which you can download and run yourself, handle high-volume routine tasks on EU-hosted or your own GPU servers; in the cost model, that server plus commercial API usage costs about €50,000 a year.

Of the four, the test sets matter most. Without them, nobody can show that a new model is good enough, and the freedom to switch stays theoretical.

Choosing the AI also means answering for it, because someone has to be accountable for what the agents do. AI governance is the set of rules and records that makes this possible: who may approve what, which cases a person must review, and how you can show afterwards what happened and why.

At the wholesaler, the company decides that any order above a customer’s credit limit, or with non-standard terms, goes to a person. It decides what is recorded, and it can explain any booking to an auditor because the record sits in its own database. With standard software and AI add-ons, many of those controls sit with the vendor.

Owning the system also makes two European conversations easier. Where the EU AI Act applies, it expects human oversight and records of what the system did. In Germany, works councils have a say in systems that can monitor employees’ performance, and a log of who reviewed which agent output can count as one, so it helps to agree what is recorded before the system goes live.

For technical readers

AI-first systems design human review around risk and confidence instead of checking every case, because having a person look at everything would cancel out most of the gain. Each task gets thresholds for the amount, the customer, and the agent’s own confidence in its result, and any case that crosses one of them goes to a person. Thresholds will occasionally let a mistake through, so a random sample of the cases that passed without review is audited as well. Two figures then show whether the thresholds are set well: the intervention rate, which is how often a person has to step in, and the override rate, which is how often that person changes what the agent proposed. A threshold should be loosened only once these figures show that the agent handles that type of case reliably. All of this depends on a complete record, so every step logs the model and its version, the input, the output, and the decision the person made.

The same records matter for regulation. Most back-office workflows are not high-risk under the EU AI Act, but Annex III (opens in a new tab) lists systems that allocate tasks to employees based on their behaviour or that monitor and evaluate their performance. A review process that routes cases to particular people and records their decisions can come close to that description, so it is worth checking before agents start assigning work.

How to decide and where to start

The 2027 SAP deadline is one natural decision point. Others are a licence renewal, a steep price increase, or an AI project that stalls because it cannot get the access it needs to the vendor’s system. That is the moment to compare both options over the next few years, with the internal team included.

Most technology leaders first ask why they should not keep the ERP and build the AI-first workflows on top of it. That is often the right first step, with the standard system kept as the record.

Over time, though, you keep paying per-user licences for screens fewer people use, and your agents reach your data only through the vendor’s interfaces, on terms the vendor can change. The vendor’s data model, the way its database structures orders, approvals, and records, is still tied to pre-AI processes. Building on top is the way in and replacement is the destination, one workflow and one module at a time, with the final step timed to the end of a contract.

Whatever the size of your organisation, owning core software is the wrong answer without a senior sponsor who can change both the organisational processes and the associated IT systems, because a new platform running the old process makes little sense. It also needs people who own the system after the build, whether an internal team or a long-term partner.

If you are approaching one of these decision points, test the case before you commit to it. Pick one workflow that matters, such as order to payment, redesign it, and run it beside your current system. Two to three months is enough to see whether owning your core software pays off for your organisation.

We are happy to run the five-year comparison with your own figures. Tell us how many people work in your core system and what your licences and implementation partners cost today, and we will show you how the two options compare for your organisation, including the point at which building would stop paying. Write to us at info@formtwo.ai or enquire at formtwo.ai, and we will get back to you within two business days.

In short

• SAP ends mainstream maintenance for its older ERP generation in 2027, so many mid-sized companies have to choose a new core system soon. AI has changed the economics of that choice.
• With AI-assisted engineering, a small, senior team can build a tailored core platform for roughly what a standard ERP costs to implement. In our example of a company with 250 users, building costs €2.9 million over five years and buying costs €3.5 million.
• The larger return is the operating model. In our example, redesigning the flow from order to payment could free half the time of the 40 people who handle it today, worth about €1.4 million a year and six times what owning the software saves in running costs.
• Owning the system also lets you choose which AI model runs each task and where your data goes, and keeps the rules, reviews, and records that AI governance needs in your hands.
• The case does not hold below roughly 150 to 200 users, or without a senior sponsor who will change the processes and a team that owns the system after the build. The safest start is one workflow, run beside your current system.