From my conversation with Lindsay Keys of Pavoi and Seb Debrouwere of DOSS. Read the full case study →
AI shows up on nearly every software website now. Some of it is a chatbot on top of an old system. Some of it changes how the work gets done. From a vendor's homepage, the two look the same.
Seb Debrouwere, VP of Business Development at DOSS, gave a clear way to tell them apart. Lindsay Keys, VP of Operations and CX at Pavoi, showed what the difference looks like in a real operation.
What "AI bolted on" looks like
Take an older, fairly rigid piece of software and connect an AI model to its database. You can now ask questions in plain English and get answers back. Models keep improving, so adding a chat layer gets easier every year.
The result is a chatbot sitting on top of the old system. The chatbot can tell you what's in the data, and might even send you an alert. The work itself still happens the old way.
What "AI-native" means
Seb defines AI-native software as software built to be used "both by humans and by agents or by machines in equal measure." An agent here is AI software that can carry out tasks inside the system, beyond answering questions about it.
Seb says AI models are good at two things. First, code and logic. Second, small judgment calls inside a longer process that otherwise follows fixed rules. Matching invoices to purchase orders is a good example. Fixed rules can match about 90% of the documents. For the rest, a model is good at judging which invoice most likely belongs to which order.
In an AI-native system, agents can do two things:
Do the work. Agents work inside the system and check their own work.
Improve the system. Agents can build new things in it, such as reports and tools, so the software keeps getting better.
DOSS's version of this is Dynamic Apps, which it launched the week before we spoke. DOSS will demo it live in Part 2 of this series on November 20. Register for Part 2 →
The test: one short shipment

Ask the vendor to walk through a shipment from a supplier. You ordered 1,000 units and 900 arrived. The supplier's other documents still say 1,000. Now someone has to work out what happened to the missing 100.
Some of the work follows fixed rules:
How many units arrived?
What does the contract say about short shipments?
What should the advance shipping notice to the warehouse say?
The rest takes judgment:
Do you claim a credit, ask for a refund or cancel the order?
Who do you escalate to, and how do you follow up?
"If you're not able to do the second one, then all you're doing is you have a chatbot that tells you the first," Seb says. A system that helps a person make the second set of calls, or makes them, is AI-native. Watch where the vendor's demo stops and a spreadsheet would take over.
What the test looks like in a real operation
Lindsay's team is working on this exact case at Pavoi. Lost stock can sit as "in transit" forever. Demand planning keeps counting on it as incoming. At month end it still shows up as dollars in transit, even after a claim has paid out.
Pavoi is building decision trees in DOSS for these cases. When Lindsay decides a shipment is lost, the units move out of "in transit" into a place she can track. The commitment on that stock is released, and there's an audit trail for every unit. "It happens more than people would care to admit," she says.
The system should get better with use
Seb sees each of these decisions as a way for the system to improve. The more real cases it handles, right and wrong, the more it learns about how the business runs. Over time, he says, the system of record "knows so much about how you run your business" and keeps adjusting as the business changes.
More from this conversation



















