I read Anthropic's S-1 as a supply chain person. Here are 10 things to know about it it
I have spent 20 years in supply chain, and lately a lot of time building AI agents within the Supply Chain field and watching what they do.
Anthropic's incredible trajectory and current position makes it one of the most interesting companies to follow at the moment.
So why read an IPO filing?
Because an S-1 is a document where a company must write down what it is building and what can go wrong. Anthropic filed its S-1 on 26 October 2026. I wanted to read it with the question in mind:
What does this mean for a supply chain leader?
The report clearly says that agents are moving from answering questions to doing the work.
Below is the count down ranking of what I see as the most important takeaways in the report. I did also record (my first) Youtube video about it even though that clearly need some improvement LOL.
10. Warehouses and factories are next. Prepare the equipment interfaces now
The report says that frontier AI may reach domains such as robotics, advanced manufacturing, and logistics. They also say that growth plans include robotics and physical work.
What it means is that MCP for the physical world. The WMS and other systems with open, documented equipment interfaces will be the first ones an agent can run. The sooner you are prepared for that, the sooner you can benefit from the advantages of AI.
9. Redesign planner and buyer roles before the tools do it for you
The report mentions at several locations that agents contribute to job displacement. This might be right or wrong. What is true is that because of the human agent interactions. The job of the human changes.
In short what it means is that if you don't rewrite your job descriptions and train the people in the new process, the agent and systems vendors (driven by agent work) will naturally redesign the way humans interact with the technology.
8. Your AI supplier has its own fragile supply chain
What the S-1 says is that compute "depends on a supply chain with multiple potential points of constraint or failure". This includes everything from data centres issues like labor shortages, electricity issues etc as well as other supply chain disruptions, permitting delays. It is also difficult for an AI supplier (similar to any business) to correctly plan the future capacity.
What I recommend to do is to have a clear action plan in place like either dual sourcing, manual work or communication around service reduction to your own team and customers. Also, with the AI provider, make sure to write solid contracts in order to be covered and not de-prioritized when a service issue occurs.
7. Treat your model provider as a critical supplier with geopolitical risk
The report also mentions the governmental block that was put on the more advanced Claude model (Fable). The geopolitical situation around AI is quite turbulent and we can expect quite some regulatory control and movement in the space over the next years.
What it means is that at any points, models or access to certain capabilities can switched off for any customers, in a specific country or maybe even worldwide, by a regulatory decision. The best way to prepare for this is to have a risk mitigation plan in place that can be similar to the service issue on point 8 above.
The difference here with point 8 is just that here it can be a long term regulatory decision, while on point 8 above it is more a contractual service issue.
6. Check security and compliance before agents gets critical access
In the report we can see in several sections that there are risks around cyber security issues. This is of course a really critical issue as the agent may have access to critical business data and if the agent can be compromised or act without control, then it can put a risk at the business
What it means is that when you as a business leader evaluate a supplier, the audit shall include real cyber-security capabilities. Supply chain runs on text written in many cases by external actors: supplier emails, order-confirmation PDFs, ASNs, portals. An agent who reads that text can be steered by it if it is not properly set up and attacked for example by prompt injection.
5. Set autonomy limits like an approval matrix
This can be compared to a classical compliance structure where an agent is producing a certain task, but someone else is validating that task.
Apply your delegation-of-authority matrix to agents. Spend and quantity thresholds. No change to supplier master or bank details without a human. The agent who creates a PO is not the one who approves it. Check "done" in the ERP, not in the agent's own report. Sample-audit agent actions weekly. Give agent read-only by default.
4. Pay per task, and govern the spend
Anthropic claims Claude Opus 5.5 "reduced customer costs per task by approximately 40% compared to the previous Claude Opus model, and Fable 5.1 cut the cost of typical workloads by approximately 25%. And with savings of up to 45% for complex coding and highly agentic tasks.
The problem is that: a business can increase usage, utilize higher subscription tiers, and bring additional workflows onto Claude. And Anthropic is not blocking this.
So the trick is within the business to control and monitor the usage. Also, build the business case per task, per PO processed, per exception resolved etc.
3. Your process know-how becomes the asset, so own it
Agent Skills is like a modular framework that lets organizations document repeatable tasks, workflows, and domain expertise into reusable packages. As organizations encode more of their workflows into Claude, their data, systems, conventions, and prior work, the platform becomes more valuable for them".
What it means is that the more you as a business use Claude, the more valuable it becomes and the harder it is to leave.
In my own experiments the agent writes down what he learns in files.. He cannot forget what he tried. A company needs the same thing.
So the recommendation will be to write SOPs, decision rules and escalation rules as versioned skill files that you store, outside any AI provider location. Make sure to make skills and prompts part of your IP and not the vendors.
2. Your ERP and planning vendors are becoming agent platforms
All big ERP providers now seem to provide AI agents and automated workflows as part of their service. This is not necessarily a bad thing but we can see that the value has moved into pure data storage for the actual software.
Now, depending on the ERP provider, it can be a more or less easy task but in a contract negotiation, it is worthwhile to consider to negotiate to bring external agents. Some ERPs are obviously better set up than others for this.
1. Agents will run whole workflows, not just answer questions
Anthropic says that"We are building towards a world where Claude can be given open-ended tasks lasting hours, days, or weeks, and then execute them autonomously". These agents "carry the access rights, system connections, and context that their organizations grant them".
What it means is that AI agents will become more and more autonomous. They will work for longer time periods without interruption and they will produce better quality output. This is the biggest change and all the previous points kind of build up to this one.
.