AI-native products are not normal software with a chatbot attached. They do things for their users.
They are proactive, multimodal, increasingly autonomous and available wherever the customer already is. In the large model era, we are deploying units of intelligence instead of units of compute.
That changes how we build products.
The model is not the product
Models will keep getting better. That is useful, but it also means we should avoid building the company around the quirks of any particular model.
The product is everything around it: context, tools, memory, evals, data, interfaces, permissions, observability and feedback loops.
Those systems should let us adopt better intelligence as it becomes available without rebuilding the product around it.
Evals are product infrastructure
Traditional software is mostly deterministic. AI systems are not, which means "it worked on my prompt" is not a useful standard.
Every important AI capability should eventually have an evaluation dataset that represents the work customers actually expect it to perform. Those datasets should grow with the product.
Production failures become regression cases. Edge cases become tests. Customer feedback becomes signal. New capabilities get evaluated against old ones before they ship.
A thumbs up or down is useful, but it is only one source of data. We also care about whether the task was completed, whether tools were used correctly, whether the output was grounded, how long the work took and whether a human had to fix it afterward.
Over time, this becomes one of the most valuable datasets we own because it describes what good actually means for our customers.
Route intelligence
Not every task needs the smartest model available. Some work needs deep reasoning, some needs vision, some needs speed and some needs a cheap classification. Some should still be deterministic code.
The customer should not need to know the difference.
We can route work between models, tools and traditional software based on the task, cost, latency and required reliability. The underlying system can be complicated while the experience remains simple.
Show progress
Intelligent work often takes time. Do not make customers stare at a spinner while an agent works for thirty seconds, five minutes or an hour.
Show useful progress. ChatGPT normalized streaming text, image products normalized progressive generation and at Durable we stream website creation while the site is being generated.
Long-running agents need the same treatment. Show what is happening when it is useful, keep a trace of what happened and let the customer intervene when necessary.
Then get out of the way.
Work across modalities
The intelligence should not belong to a text box. The same underlying capabilities should work through text, voice, images and other modalities.
They should also be available across the places customers already work: web, mobile, desktop, calls, SMS and whatever comes next.
A customer should be able to start something by voice and review it on their laptop later. The interface is a surface over the intelligence, not the intelligence itself.
Fine-tune when you know why
General-purpose and open-source models improve quickly. Our default is to use that progress rather than compete with it.
Fine-tuning becomes interesting once a use case is stable, the evals are good and we know exactly what we are optimizing. Maybe we need lower cost, lower latency, more consistent behavior or a much smaller model for a narrow task.
Until then, changing prompts, context, tools, workflows and models usually gives us more leverage.
Agents should do jobs
We do not want intelligence that only produces text. We want intelligence that completes work.
An agent should be able to understand a goal, gather context, use tools, make decisions, ask for help when necessary and continue working in the background.
Eventually, much of the best AI UX should be invisible. The customer should not need to watch every step. They should get the outcome.
AI native inside the company too
The same ideas apply to our own systems. An AI-native company should not put agents in the customer-facing product while keeping its internal workflows entirely rule based.
Analytics, customer journeys, support, operations and lifecycle messaging can increasingly be driven by intelligence that understands the individual customer and their context.
Instead of one generic workflow for thousands of customers, we can move toward one intelligent system working on behalf of each customer.
AI native is not a feature. It is an architecture for the product and, eventually, the company.