Inventory optimisation software has been around for decades, and most of it has the same fundamental limitation: it tells you what to do, but it can’t help you do it.

Until recently, that wasn’t possible without considerable hallucinations. Thanks to the latest AI models that are now faster, more affordable and accurate enough to be genuinely useful in a business setting, that’s finally become a reality.

For us, that unlocked something we’ve been building towards for ten years: all the ERP, WMS, or e-commerce integrations, and inventory, demand, and supplier data across multiple systems, now with an agent on top that can actually use it.

It’s the closest thing to a supply chain specialist working inside your software, and we’re very excited to introduce it. 

Meet our new Supply Chain Agent

When we set out to build Optiply’s Supply Chain Agent, we had a clear set of goals in mind. We wanted to build an agent that could:

• Talk to your supply chain data in plain language, without needing a specialist to interpret the output.
• Execute tasks on your behalf, from updating supplier lead times to working through a structured improvement plan.
• Flag problems before you have to go looking for them, so your team stays ahead rather than catching up.
• Learn from your business over time, remembering past decisions, preferences, and context from every interaction.

We believe we’ve achieved this. The agent connects directly to your existing data and integrations, so there is no lengthy setup or data migration. 

From day one, it knows your suppliers, your stock positions, your ERP systems and your history with us.

It can hold a conversation across multiple turns, understand follow-up questions and adjust its reasoning as your needs become clearer. And because it’s built on ten years of real supply chain data and thousands of customer support conversations, it understands the domain in a way that a general-purpose AI tool cannot replicate.

It’s not perfect, and we’ll be specific about its current limitations further on. But we believe it’s ready, and we think it will quickly become something your team relies on every day.

‍

How be built it: multi-agent from the start 

We built the Optiply Agent on a multi-agent architecture, meaning different specialised agents handle different tasks rather than one large model doing everything. This was a deliberate technical choice, and it has a direct impact on reliability.

If you give a single agent 20 tools, it has to decide which one to use on every request, and the more options it has, the more likely it is to choose poorly or hallucinate. Give a specialised agent a single tool, and it will use it correctly almost every time. 

Sander, Co-founder at Optiply compares it to reading comprehension.

It's like asking someone to read a 500-page book and then name the 10 most important pages. They'll have no idea. But ask them about 20 sentences and they'll know it perfectly.

Specialisation reduces the surface area for errors.

This architecture also maps naturally to how supply chain teams actually work: different people with different responsibilities, collaborating towards a shared outcome.

The agent is also trained on something no new competitor can replicate: ten years of customer support conversations gave us a detailed picture of which problems come up most often, which questions purchasers actually ask and what answers look like in practice.

What makes it different from other AI tools in supply chain

A lot of supply chain software now claims to have AI. Most of it means one of two things: a machine learning forecasting model that has been there for decades, or a chat interface layered on top of existing functionality that cannot change anything in your system.

The most direct way to explain the difference is that chatbots answer stuff for you, the Supply Chain Agent executes things for you.

The table below shows what that looks like in practice.

These are the kinds of issues a consultant might catch in a quarterly review. The agent catches them on a random Tuesday morning.

How the agent knows what to recommend and when to stop

The agent's intelligence comes from three places: 

• Your live data
• A continuously updated memory built from every interaction your team has with ours
• Context from over a thousand daily users

When you speak with our customer success team and agree that reducing overstock is the priority for Q2, the agent has that context the next day. 

When you tell it to refer to suppliers by their Optiply IDs rather than their names during a presentation, it stores that as a preference and applies it in every subsequent session until you change it. It does not start from scratch every time you log in.

Staying in control

The most common concern we hear from customers is about trust. Letting software act on your behalf in a purchasing context is a significant decision, and we’re not going to pretend otherwise.

For now, every action the agent wants to take requires your explicit confirmation. It tells you what it intends to do, and nothing happens until you approve it. That is a hard boundary, and it’s not moving in the near term.

Over time, customers will be able to define their own thresholds, such as the maximum value of a purchase order the agent can raise autonomously, acceptable deviation from a previously agreed replenishment strategy, so that the degree of autonomy reflects the degree of trust you have built with the system.

There’s also a second feedback loop running in the background. When the agent generates a new improvement plan for a customer, it first routes that plan to one of our own supply chain specialists for review before it’s shown to you. 

That expert can approve it, adjust it, or reject it, and whatever they decide, the agent learns from it. It’s how we ensure that the agent’s judgement improves continuously rather than drifting.

Where we’re still improving

We want to be upfront about what the agent doesn’t do well yet, because some of these limitations will matter more to some teams than others.

Speed

First of all, the agent’s responses aren’t instant. It’s doing real work – querying live data, cross-referencing multiple sources, reasoning through the implications before it responds. 

That takes a few seconds longer than a keyword search, and in some cases longer still. We have optimised for accuracy over speed, because in a purchasing context, a fast wrong answer is worse than a slower correct one.

Accuracy

It’s important to note that the agent can make mistakes. It occasionally misreads context, particularly in edge cases that fall outside the patterns it has been trained on. 

The failure mode here is different from a conventional software rule: rather than returning no answer or an irrelevant one, it may return a plausible-sounding answer that is wrong. 

We have a feedback loop built in so that errors are captured and used to improve the model, and we label agent responses clearly so your team always knows they’re interacting with AI.

Finally, we’re confident that for most supply chain teams, the agent’s accuracy is already above the threshold where it adds more value than it introduces risk. And we expect that threshold to keep rising. 

Models are improving with every generation, and we’re learning from real usage every week. The version available today is already meaningfully better than what we were testing six months ago.

Pricing

Pricing is also something we’re still working through. Adding an agent that can actively reduce overstock, prevent stockouts, and replace hours of consultant time does not map cleanly onto a per-user or per-feature model. 

Our current thinking is that the core agent should be part of your existing Optiply plan, since it makes everything else in the product more valuable. Capabilities that go beyond that, like automated supplier communication or autonomous order placement, will likely be priced separately. 

But we want to be honest: we don’t have all the answers here yet, and we would rather get the pricing right than rush it.

What’s available today, and what’s coming next 

We have built an initial version of the Supply Chain Agent which you can try today. 

At the moment, this is what it can do:

• Give you live conversational access to your supply chain data.
• Help with configuration changes like supplier lead time drift, with your approval required before anything is committed.
• Guide you through structured action plans for complex processes like onboarding that would otherwise require a specialist.
• Work across Dutch, English, and Spanish, trained on the specific terminology we use in the application so it understands that "transit time" and "lead time" are the same thing in your context.
• Remember your preferences across sessions, so it gets more useful the longer you use it.

In the coming weeks, we’re releasing proactive insights. The agent will surface issues before you ask. For example, flagging that four suppliers have lead times drifting enough to cause one to two days of stockout risk on specific SKUs, and proposing a correction. 

Next, we’ll release simulation tools, so you can model the impact of a strategy change before committing to it: if you lower service level by 5% across a product category, the agent will show you the projected effect on stock value and fill rate.

Further ahead, the agent will be able to set up its own workflows, draft changes to ERP field mappings, and interview stakeholders directly to keep its understanding of your business current. 

And much more, so stay tuned.

Where supply chain is going

The near-term opportunity is augmentation, not replacement. A purchaser who uses these tools well will outperform a fully automated system, because they can negotiate terms, build supplier relationships, and make judgement calls that no agent will be permitted to make. 

A purchaser who does not use these tools will struggle to compete with one who does. That gap is going to widen quickly. 

We believe today’s announcement marks a shift in the field of supply chain. The future is happening right now. 

Learn more about the Supply Chain Agent here. Some of these capabilities are available in the product already, and we’re rolling out the rest as fast as possible. We’re excited for you to try everything and hear your feedback.