Workflow Overview for AI Products
The main difference between the two workflows is not the technology a team uses. It is how the work is organized and who participates in which decisions.
Predicting Who Shapes the Product
When people talk about building software that uses machine learning, the conversation often centers on the model: which architecture to pick, how to train it, and how to measure its accuracy. A model, however, is only one part of a working product. Someone still has to decide what the product should do, how users will reach it, and when a result is good enough to ship. How a team divides that work, and who takes part in which decisions, shapes what gets built and how fast the team can learn.
What do you think happens?
In a traditional ML team, who usually decides what the product should offer and what counts as success?
Reveal answer
Decisions are often made elsewhere, and the ML engineer receives a product direction rather than helping to define it.
At many organizations, ML engineers are rarely involved in product decisions under the traditional pattern. Choices about what the product offers, how it behaves for users, and what counts as success are often made elsewhere.
Two Ways to Organize the Work
This article compares two ways of organizing AI product work. The first is the traditional ML engineering workflow, in which model development and product development are often separate processes. The second is the AI engineering workflow that has emerged alongside foundation models, in which the people building the AI are much more involved in building the product itself. The goal is to describe the difference clearly and to recognize which pattern a team is following.
The main difference between the two workflows is not the technology a team uses. It is how the work is organized and who participates in which decisions.
Handoffs in the Traditional Workflow
In the traditional pattern, the product direction is set first, and model development happens as a separate stage. The ML engineer receives that direction and works to produce a model that serves it. Once the model is ready, it is passed along, and product development continues from there. Work moves from one group to another, which is why this pattern is described as disjointed.
Follow the path in the diagram and notice where the ML engineer's role ends. The engineer starts after the direction exists and finishes when the model leaves their hands. Choices about what the product should offer, how it should behave for users, and what success means are made at the top of the chain, not by the person training the model.
What Changes With Foundation Models
The central claim of the source material is that the situation changes with foundation models. With foundation models, AI engineers tend to be much more involved in building the product. The model is no longer built in isolation and then passed along. The engineer who works with it is often also deciding how it is used inside the product.
Read the two columns side by side. In the traditional column, the ML engineer and the product decisions sit apart. In the AI engineering column, the engineer sits next to the product decisions. The role title is not what changes; the scope of participation is. The AI engineer is involved in building the product, not only in producing the model that the product later receives.
Why Iteration Becomes the Rewarded Skill
This shift also changes what gets rewarded. The new AI engineering workflow rewards those who can iterate fast. Iteration here means trying an idea inside the product, seeing how it behaves, and adjusting quickly. Because the engineer is close to the product, each cycle of feedback can feed directly into the next change.
One Pass Through the Loop
An AI engineer who is close to a product adds a new feature that uses a model, then notices users are confused by one kind of response. How does a single pass through the iteration loop handle this?
- 1Try
The engineer puts the feature into the product so it can be tested with real use, not only on a separate evaluation.
- 2Observe
The engineer watches how the feature behaves and notices the confusing responses.
- 3Adjust
Because the engineer is involved in how the model is used in the product, they can change that usage directly and try again.
The feedback from one pass feeds straight into the next change. Repeating this cycle quickly is how the workflow rewards the engineer, and how the product improves.
Mistakes Learners Make About These Workflows
Common mistake Assuming the difference between the two workflows is the technology used
Why it failsThe source states that the main difference is how the work is organized and who participates in which decisions.
Do this insteadCompare the teams by who makes product decisions and where work is handed off.
Common mistake Assuming an ML engineer in a traditional team is responsible for the product direction
Why it failsUnder the traditional pattern, ML engineers are rarely involved in product decisions at many organizations.
Do this insteadRecognize that the ML engineer usually receives a product direction and builds a model to serve it.
Common mistake Treating iteration as random trial and error with no judgment
Why it failsThe source defines iteration as trying an idea in the product, seeing how it behaves, and adjusting quickly, with each feedback cycle feeding the next change.
Do this insteadThink of iteration as a structured loop of try, observe, and adjust.
Common mistake Believing foundation models only change which model is chosen, when they also change how the engineer participates in the product
Identifying a Team's Workflow Pattern
A team describes its process this way: a product lead writes the requirements, an engineer trains a model to match them, and the finished model is then handed to a separate group that builds the user-facing product. Which workflow pattern is this team following, and which roles are involved in product decisions?
Hints
- Ask whether model development and product development happen as separate stages or as one connected activity.
- Look for who sets the product direction and who receives it.
- The traditional pattern is characterized by handoffs between model work and product work.
Key Takeaways
- In traditional ML engineering, model development and product development are often disjointed, with the ML engineer frequently receiving a product direction rather than shaping it.
- With foundation models, AI engineers tend to be much more involved in building the product, including deciding how the model is used inside it.
- The AI engineering workflow rewards engineers who can iterate fast, meaning they try an idea in the product, observe how it behaves, and adjust quickly.
- The main difference between the two workflows is how the work is organized and who participates in which decisions, not the technology used.