- 19
- August
"Is n8n dead?" — the short answer is no, and the 2026 numbers point firmly the other way. In May 2026 SAP invested in n8n at a valuation of roughly $5.2 billion, alongside a multiyear deal to embed n8n's tooling inside SAP's own AI agent builder. What actually changed is not that n8n died but that its role moved — from the place where all the logic lives, to the connective layer that orchestrates AI agents. This article separates the work agents have genuinely taken over from the work workflow automation still wins outright.
In one line: AI agents did not kill n8n; they changed its job. More than 80% of workflows built on the platform now have an AI agent somewhere inside them.
Why people started writing the obituary
The instinct is not unreasonable. Through 2026 the agent side moved squarely onto territory workflow tools used to own:
- April 2026 — Claude Managed Agents shipped with the infrastructure layer teams previously assembled themselves: sandboxed code execution, credential management, scoped permissions and end-to-end tracing.
- The same month — Routines made it possible to run an agent on a schedule, via an API call, or on an event such as a GitHub webhook. That is one of a workflow tool's core jobs.
- July 2026 — ChatGPT Work launched on the promise of working across applications and files from a single instruction until there is a deliverable.
When Managed Agents launched, opinion split immediately: one camp declared workflow tools finished, the other declared agents an overpriced toy. Several months on, both camps look wrong.
The numbers that say otherwise
| Period | What happened |
|---|---|
| 2025 | $180 million Series C at a valuation of roughly $2.5 billion |
| May 2026 | SAP invests, doubling the valuation to roughly $5.2 billion, with a multiyear deal to embed n8n's tooling inside SAP's AI agent builder |
| 2026 | Annual recurring revenue reported to have passed the $100 million mark |
| Current state | More than 80% of workflows built on the platform now involve AI agents |
That last figure is the whole story: if agents were killing workflow tools, the share of workflows containing an agent would not be climbing. It says people did not abandon workflows for agents — they put agents inside the workflows they already had.
What agents genuinely took over
It is worth being honest that some things people used to build as long workflows simply do not need to be built that way any more.
- Dozens of nodes to classify then branch — the ten-layer if/else tree that handled every message variant collapses into one model call that reads and decides.
- Work whose steps emerge as you go — root-cause investigation, or reading a long document and reporting where it contradicts itself.
- Scheduled work that is purely about code — nightly reviews, dependency updates — which Routines handles without a workflow layer in between.
If your automation is entirely of these three shapes, dropping the workflow tool is a defensible decision.
Where workflow automation still wins outright
| Concern | Why a workflow is still the better answer |
|---|---|
| Cost per item | A hardcoded path does not pay to re-reason on every run — at thousands of runs a day that gap is enormous |
| Consistency | Identical output every time, with no drift in interpretation — which matters a great deal for anything with accounting consequences |
| Legacy integration | Systems exposing only REST, SOAP or a database still need something to receive webhooks, reshape payloads and queue work |
| Human approval gates | Human-in-the-loop pauses are first-class, and real approval processes require them |
| Governance and audit | Role-based access, audit history, and control over which workflow may touch which secret |
| Redaction before egress | Filter and redact inside your own perimeter, so only sanitised input ever reaches a cloud model |
| Failure handling | Retries, timeouts and dead-letter routing are long-solved problems in a workflow engine |
Three failure modes dominate agent workloads in production: model timeouts, tool errors and infinite loops. None of them is fixed by a smarter model. They are fixed by a control layer that caps iterations, caps spend, and defines what happens on failure — which is exactly what a workflow engine is for.
The advantage that gets under-counted: self-hosting
The community edition self-hosts for free with no execution or workflow limits. That matters for two groups in particular: organisations under rules requiring certain data to stay inside their own network, and organisations that simply want control over what leaves for a cloud model.
The pattern that works is to let a self-hosted workflow filter and redact first, then send only what is necessary to the model — you get cloud-model capability without the raw records leaving your perimeter. For teams who want to go further still, running models in-house is a viable path; we covered it in Ollama API — connecting AI to your enterprise systems.
The market context that lets both sides grow
The reason agents and workflows are growing together is that the market is expanding rather than being redivided. Gartner projects that by 2026 roughly 40% of enterprise applications will include task-specific AI agents, up from less than 5% in 2025.
In practical terms that means a typical organisation will soon run not one agent but many, scattered across several systems. At that point the question shifts from "should we use agents" to "who controls which agent can touch what, and can we reconstruct it afterwards" — which is exactly why the connective and control layers become more important, not less.
The line that actually helps you decide
| Shape of the work | Use |
|---|---|
| Fixed steps, high volume, daily repetition | Workflow only — keep the model out of the hot path |
| Fixed steps with a point that requires reading and judging | Workflow calling a model at that point — the most common 2026 pattern |
| Unpredictable steps; genuine exploration | Agent-led, with the workflow as a bridge to legacy systems |
| A human approval step mid-process | Workflow owns the process, always |
| Sensitive data that must not leave raw | Self-hosted workflow filters first |
| Scheduled, code-only work | Routines — no workflow needed unless several systems are involved |
Notice that almost every row resolves to a combination rather than a winner — the same conclusion we reached approaching the question from the other side in Is Claude Agent self-sufficient?
What this looks like inside an ERP
ERP processes make the split obvious, because nearly every one of them mixes both kinds of step. Take a supplier quotation arriving by email: pulling the attachment, converting it, filing it and notifying the right people is deterministic and belongs in a workflow. Reading a document whose layout varies and extracting line items and prices is where a model earns its cost. And the approval itself stays with a person — not because a model could not do it, but because legal responsibility sits with a human.
To be straight about our own product: the Saeree ERP AI Assistant is still in training and is not shipped as a customer feature. What the system provides today is the foundation both workflows and agents depend on — configurable approval routes with ordered signers, budget checks before approval, reserved, committed and outstanding amounts held as distinct states, and an auditable trail at every step. For the broader context, see What is Agentic AI and Forrester's 2026 predictions on AI agents in enterprise software.
Conclusion
"Dead" is the wrong word for what happened. This was a change of role — from the place where all logic is drawn as boxes and arrows, to the layer that owns what agents handle badly: system integration, queueing, human approval gates, redaction before data leaves, and an auditable trail. Meanwhile the logic that once needed ten nodes collapses into a single model call.
The most useful advice for a team currently hesitating: do not rip anything out. Take the single longest, most condition-heavy process you run, replace its branching block with one model call, and measure two things — time saved and cost per item. If it pays, expand. That is how you capture the upside of both approaches without betting on which one wins.
n8n was not killed by AI agents. It was moved — from writing all the logic to guarding the doors, which is arguably the more important seat.
- The Saeree ERP team
References
- Bloomberg — SAP invests in AI automation startup n8n at $5.2 billion value
- Trending Topics — SAP bets big on AI, invests in n8n at a $5.2B valuation
- n8n Blog — Series C: getting AI closer to value with orchestration
- Atlan — n8n for enterprise AI workflows
- IronPlate — Claude Code Routines vs n8n
- Zarif Automates — Claude Managed Agents vs n8n
Not sure which of your processes should be a workflow and which should be an agent?
Talk to the Saeree ERP team about approval routes, pre-approval budget checks and audit trails — the foundation both workflow automation and AI can safely build on.
Request a Free ConsultationTel 02-347-7730 | sale@grandlinux.com
