A common view among AI-native builders is that the systems of record (SAP and Oracle in the ERP, Salesforce in the CRM, Workday in HR, ServiceNow in IT, the MES, QMS and PLM vendors on the factory floor) will lose the AI era to vertical, AI-native players. This piece makes the strongest available case against that view: that the incumbents win by 2030 with their installed base and agents that are only good enough. It then sets out where the case breaks down, and what evidence would settle it.
AI use : 25%.
The challenger thesis
The thesis has three parts. Earlier essays on this site develop each of them: on unbundling the ERP, on ontology and on SAP's Sapphire keynote.
- The data is not where the vendors think it is. Systems of record were built to record what happened, for finance and compliance. Agents need what is happening and why, and teams deploying agents in factories report that about 40% of that context exists in no system at all.
- The value is in the decision rules. The so-called kinetic layer, the rules that say what a good decision is, sits in spreadsheets, PDFs and people's heads. Whoever writes it down and governs it captures the gains.
- So the record recedes. The system of record stays, as a thinner and quieter ledger, and the value moves to vertical tools that own a specific operational task: the scheduler, the picker, the maintenance planner.
The case for the incumbents
“Winning ugly” does not mean the incumbents build the best AI. It means they keep the customers, the budget and most of the margin while their products stay mediocre. Here is the strongest form of that case, argument by argument.
1. Good enough plus distribution beats best
Enterprise software history has plenty of better products that lost. Slack was the better chat product, and Microsoft Teams overtook it by being bundled into a licence companies already paid for. Teams was not the better product; it was already installed. Slack ended up inside Salesforce, a system-of-record company.
Incumbents do not need the best agent. They need an agent that clears the bar on a process the customer already runs on their platform, and that comes with the contract renewal. A buyer comparing a 70% solution that is already approved by IT, security and procurement with a 95% solution from a company with forty employees will pick the 70% one more often than founders like to admit. That is what “winning ugly” means: winning with the worse product and the better position.
2. The model is a commodity, and commodities favour whoever owns the customer
The challengers' own reasoning works against them here. If the frontier labs keep shipping better models at a falling price per token, intelligence becomes something incumbents buy by the token, like electricity. Every lab wants the incumbents' distribution and signs partnerships with them. An AI-native challenger and SAP rent the same reasoning engine. The difference between them then comes down to what the engine is plugged into, and the incumbent is plugged into everything.
The last five years have shown the pattern elsewhere. Predictions that GAFAM would crumble or that banks would be disrupted aged badly, because AI amplifies institutions that already have distribution and capital. There is no obvious reason the system-of-record vendors should be the exception.
3. The record is the permission layer
An agent that only reads is a report. An agent that matters has to write: release the order, post the invoice, change the routing, close the ticket. In a regulated or audited company, those writes go through the system of record, because that is where the audit trail, the approvals and the legal liability live.
Whoever owns the commit owns the outcome, and the incumbents own the commit. They can also set its price. Salesforce tightened the Slack API terms in 2025 to limit how third parties bulk-export and index customer messages. Every vendor of record can do the same with API fees, rate limits and certification programmes. The AI-native challenger then pays a toll to the company it is trying to replace, on every transaction.
4. Someone builds the missing 40%, and the incumbent buys it
The challengers' strongest field observation is that about 40% of the data an agent needs does not exist in any system today. That may be true, but it only says the data has to be built, not who ends up owning it. Incumbents have been buying the layers around their core for years: SAP bought Signavio for process mining, LeanIX for architecture and WalkMe for the screen layer; Salesforce bought Informatica for data management; ServiceNow bought Moveworks for agents. Each of these deals is an AI-native layer that ended up inside an incumbent.
The forward-deployed engineers that Microsoft, Amazon and the labs are now funding by the billion do not build in a vacuum. They build on the client's existing estate, and the context they produce gets written next to the record they integrate with. The first-equipment wave the challengers describe could make the incumbent's data more valuable, not less.
5. The buyer is paid to avoid regret
“Nobody gets fired for buying IBM” was never a joke about technology. It describes the CIO's incentives. If the incumbent's agents disappoint, it is an industry-wide disappointment. If the start-up fails, it is the CIO's mistake. Those payoffs are not symmetric, and they do not change because the start-up's demo is better.
The calendar also helps the incumbents. SAP ends mainstream maintenance for ECC in 2027, with extended support to 2030. That deadline forces thousands of companies into a large migration project right now, and a company already doing that migration is unlikely to also bet on a new system. It will choose what Angela Strange calls path one: keep the system of record and add agents on top. Challengers call it the losing path. Losing paths can still have most of the market.
6. The share price is not the revenue line
Challengers point to the repricing of listed system-of-record companies, roughly half their value gone in twelve months, as the market agreeing with them. But a lower multiple is a statement about uncertainty, not a cancelled contract. These companies still collect their maintenance and subscription revenue every quarter, and they only need to survive the transition window, not win it. Oracle and SAP survived client-server, the web and the cloud, each of which was supposed to finish them. The market has declared the end of the ERP several times before.
What winning ugly would look like in 2030
A thesis that cannot lose is not worth holding, so here are signals that should separate the two outcomes over the next four years.
| Signal | Incumbents win ugly | Challenger thesis holds |
|---|---|---|
| Where agents commit their writes | Through the vendor's own agent runtime, priced per action | Through open APIs, with the record used as a plain ledger |
| Fate of the best AI-native vertical companies | Acquired by a system-of-record vendor before they reach scale | Replacing modules of the suite and staying independent |
| Incumbent pricing | Seats replaced by agent or outcome pricing on the same contracts | Licence revenue shrinking as fewer people log in |
| Where the operational context lives | Inside the vendor's data cloud and knowledge graph | In separate ontologies the customer owns and can move |
| ECC migrations completed by 2030 | Mostly like-for-like moves to the vendor's cloud | Many companies shrinking the ERP to a financial core |
Where the case breaks down
Made as strongly as possible, the case for the incumbents is right about more than challengers usually admit. It is wrong in four specific places.
- “Good enough” does not work where the money is. Teams beat Slack because a chat message that is 90% right costs nothing. In decisions with low tolerance for error and deep context, such as a production order on the wrong machine or a quality release on the wrong lot, a 70% agent is worse than no agent, because someone has to check every output. SAP's own chief executive said it on stage: for mission-critical processes, “almost right” is not good enough. Winning ugly works for email, summaries and tickets. It does not work for scheduling a plant.
- Buying a company does not buy its fieldwork. The acquisitions in argument four bought horizontal layers: process mining, data management, a screen overlay. The missing 40% is different in every plant. It is built plant by plant, by people who learn how the work is really done, and it does not transfer with a share purchase agreement. In the IBM era the winner was whoever did the deployment, not whoever owned the earlier system.
- A toll is not the same as the value. Argument three is the strongest one, and part of it holds: incumbents will keep the commit and charge for it. But a toll on transactions pays like infrastructure, steady and with a ceiling. The gains from better decisions, such as fewer changeovers, less scrap and less working capital, go to whoever owns the decision rules. A company can be a very good business as a toll road and still not be the winner of the era.
- The challengers will not stay on top of the record. The whole case assumes AI-native players politely build on top of the incumbent and pay its toll. The best of them will not. They will go for the throat: win the decision layer and, at the same time, replace the layer underneath it, the MES, the PLM, the planning module, while the customer is already reopening those choices. AI has cut the cost of writing the deterministic code a system of record is made of, so rebuilding it is no longer a ten-year programme. Every forced migration is a moment when a customer reopens choices it had not questioned in decades. A toll road only pays while there is no other road, and once a challenger owns both the decisions and the record, the incumbent no longer has a window to wait out.
The defensible version of the challenger thesis is therefore narrower than the usual one. In finance, payroll, HR records and anything where the record is itself the product, the incumbents probably do win ugly. In operations, where the value sits in context-heavy decisions that the record never captured, the vertical players are better placed to take the gains. The incumbents will collect a toll on them for a while, until the strongest challengers replace the layer underneath as well. The question for 2030 is not whether the systems of record survive. They will. It is whether the decision layer becomes a module in their suite or a separate company built on top.
What would settle it
Three developments would mean the incumbents are winning, even in operations:
- An incumbent agent running a low-tolerance operational decision, such as production scheduling or quality release, in production at several customers, without an integrator rebuilding the context by hand.
- The best AI-native operations companies consistently selling to an incumbent before they reach scale, at prices that suggest they had no independent path.
- Customers moving their decision rules into a vendor's knowledge graph and accepting that they cannot take them elsewhere.
If two of the three are visible by 2028, the challenger thesis is wrong.
Four scenarios for 2030
The arguments above can be turned into a range of outcomes for the new AI players. The table puts an order of magnitude on each one: the share of new AI and agent spending the challengers capture, and the share of the total budget, meaning everything companies spend on systems of record and on the decision layer above them.
| Scenario | What happens | New players: share of new AI spend | New players: share of total budget | Incumbents: share of total budget |
|---|---|---|---|---|
| Extreme | Challengers go for the throat in every category. They win the decision layer and replace the operational records underneath (MES, PLM, planning) and large parts of CRM and IT service management. Incumbents shrink to the financial ledger, payroll and compliance. | 70–80% | 40–50% | 50–60% |
| Strong | Challengers win the decision layer in operations and replace some operational records, mostly MES and planning. Incumbents keep finance, HR and CRM, sell their own agents there, and charge a toll on the rest. | 50–60% | 20–30% | 70–80% |
| Mild | The incumbents win ugly in most categories. They bundle good-enough agents into renewals and buy the best challengers before they scale. Independent challengers hold niches in low-tolerance operational decisions. | 25–35% | 8–12% | 88–92% |
| Winter | AI in the enterprise disappoints, or funding dries up. Budgets fall back to the safe vendor, most challengers are acquired or close, and agents become a feature of the existing suites. | 10–15% | 2–5% | 95–98% |
Total budgets move slowly even when new spending moves fast, because most of the money still goes to renewing contracts already in place. That is why even the extreme scenario leaves the incumbents with about half of the total budget. The conclusion above, that incumbents win ugly where the record is the product and challengers take operations, lands between the mild and strong scenarios. The signals table is the way to tell which one is unfolding.