The bottom line
The 2025–26 software repricing (the 'SaaSpocalypse') punished seat-based SaaS on the fear that AI agents make the login obsolete. The market is now sorting software into vulnerable 'convenience layers' and defended 'deep workflow platforms.'
Why the enterprise-software repricing of the last year is a correction, not an extinction — and what the "deep workflow" thesis means if you own an ERP estate.
The views below are my own, drawn from public market commentary, vendor roadmaps, and my work in the ERP ecosystem. Reasonable people read the same signals differently, and I welcome that.
The repricing everyone felt
Between the middle of 2025 and the low in March 2026, the public software sector went through a violent repricing. The fear was clean and specific: if autonomous AI agents can do the work a human used to log in and do, then software priced per seat is selling a unit — the login — that is about to matter far less. Analysts started calling it the "SaaSpocalypse," and it hit anything that looked like a per-user subscription.
A useful frame for what survives came from an unexpected place: a Bloomberg Odd Lots conversation with BlackRock's operating chief, who described how the firm runs on Aladdin — a single proprietary platform that unifies portfolio management, trading, compliance, and operations on one database and runs risk at massive scale. Aladdin is not a convenience app. It is BlackRock's system of record and its operational backbone. The distinction he drew is the one every ERP owner should now internalize.
Convenience layers vs. deep workflow platforms
The market is separating software into two buckets:
- Convenience layers — applications that mostly aggregate public or shallow information behind a nice interface. If an agent can assemble the same answer directly, the interface is exposed. These are what the repricing punished.
- Deep workflow platforms — systems built on proprietary data, with institutional workflows embedded and switching costs measured in years. These are structurally defended, because the moat is not the UI; it is the data and the process.
An ERP is the archetypal deep workflow platform. Its ledgers, master data, and compliance-bound transaction history are proprietary, integrated, and expensive to leave. That is the moat of truth — the reason the core system of record is one of the best-defended assets in the enterprise against AI displacement. The threat was never to the ERP. It was to the ring of point solutions bolted around it.
That reframes the roadmap on both sides of the table:
- If you run ERP: be skeptical of standalone SaaS add-ons that are really convenience wrappers around data your core already owns. Keep the core clean and lean on the platform vendor's embedded AI — Joule inside SAP, the native agents inside Oracle and NetSuite — before you buy an external point tool that duplicates them.
- If you build ERP: the winning move is to activate the data moat. SAP's "autonomous enterprise" blueprint puts specialized agents across finance, HR, and supply chain on top of its business data; Oracle is unifying structured and unstructured data with semantic vector search so agents can reason over the whole estate. The pattern is identical: turn the system of record into an execution platform.
The first draft principle
The obvious objection to putting AI near a general ledger is that generative models are non-deterministic and finance tolerates almost no error. The reconciling idea is the First Draft principle: the agent produces the first draft — the reconciled entry, the sales order, the variance narrative, the proposal — and a human reviews and signs before anything is final.
For the people who run the estate, that is a role change more than a headcount change: from content creator to content editor. An agent drafts the period-close reconciliation or the vendor-dispute resolution; the analyst validates and approves. For it to work, the platform has to make the draft auditable — citations grouped in a side panel, the specific records and documents that informed the output visible on demand. Trust in a system of action is built on traceability, not on faith in the model.
Tokenomics: the budget line nobody had
There is a cost consequence to all this that lands squarely on the CIO and CFO. As enterprises move from paying "per query" to paying "per workflow" — where a single agent run executes many steps — the compute behind each action scales, and consumption becomes the metric that matters. Licensing is drifting away from seat counts and toward what you actually use: compute, storage, tokens.
That is not automatically bad — you stop paying for logins that do nothing — but it demands new discipline:
- Instrument the unit economics. Know what an AI-drafted invoice, a resolved dispute, or a generated RFP actually costs to produce.
- Budget compute as its own line. Consumption belongs in its own envelope, not buried in the software line where it can run away unmonitored.
- Model the long run, not the launch discount. Base commitments on your real workflow history, not a bulk offer sized to next quarter's sales target.
The organizations that win here are the ones that treat token consumption as an operational metric to manage — the same way they already manage any variable cost tied to work performed.
The talent shift
The BlackRock discussion made one more point worth carrying over: at a firm that is roughly a quarter technologists, the prized hires increasingly bring analytical and communication strength, not just code. When low-code agents compress the distance from concept to implementation, the constraint stops being how fast you can build and becomes how clearly you can frame the problem.
For ERP consultants and integrators, that is a migration from technical configurator to process architect. AI handles the baseline configuration and code generation; the human value moves to designing the business process, orchestrating the workflow, and translating operational reality into the system. The teams that lean into that shift — arming themselves with the vendors' own consulting-grade AI — pull ahead of the ones still selling hours of manual setup.
What the recovery is telling us
Since the March low, the sector has clawed back roughly half the lost value — but unevenly, and the unevenness is the signal. Infrastructure, security, and data platforms — Datadog, CrowdStrike, ServiceNow and their peers — recovered hard, because they are the indispensable foundation under any serious AI deployment. Seat-based horizontal apps like the familiar point solutions stayed discounted even while reporting healthy growth. The market is paying up for deep utility and backend integration, and marking down the interface layer. That is the convenience-vs-deep-workflow thesis being priced in real time.
As the BlackRock executive put it, "there is no fighting technology." The SaaSpocalypse was never the end of enterprise software. It was the market re-sorting it by where the durable value actually sits — in the proprietary data and the embedded workflow — and the ERP sits squarely on the right side of that line. The question is no longer whether your system of record survives. It is whether you activate it as a system of action, and govern the cost and the risk while you do.
Analysis by Ravi Madhavan, offered as one practitioner's reading of what the enterprise-software repricing means for ERP. Originally published on LinkedIn. Substantive disagreement or alternative reads welcome. Allari runs the ERP support layer for JDE, SAP, Oracle Fusion, and NetSuite estates on a consumption-based model — measuring the repeatable manual work these agents absorb and finding where it can be automated.
