# AI Engineer vs ML Engineer vs Software Engineer

> What AI engineers, ML engineers and software engineers own, from live postings at the same employers, and how to tell which job a posting describes.

- Source: https://www.zarifautomates.com/blog/ai-engineer-vs-ml-engineer-vs-software-engineer
- Published: 2026-09-26
- Updated: 2026-09-26
- Pillar: AI Careers
- Tags: ai-engineering, machine-learning, software-engineering, careers
- Author: Zarif

---

An AI engineer builds products on top of models. An ML engineer builds and ships the models. A software engineer builds the systems both of them plug into. That is the short version, and like most short versions it breaks on contact with real job titles, so the rest of this guide checks it against postings.

The postings compared here were opened on September 26, 2026. Three of them come from one company, Ramp, which was hiring for all three roles that day. Comparing them removes the question of whether a difference is about the role or about the employer. Roles close and ranges change, so open the link before relying on any detail.

## The three jobs side by side

| | AI engineer | ML engineer | Software engineer |
| --- | --- | --- | --- |
| Owns | A product feature or workflow whose core step is a model call | A model: its data, training, evaluation and deployment | A product surface or system: its data model, APIs and reliability |
| Main question | Does the product behave correctly when the model is uncertain? | Does the model predict well enough, and keep doing so? | Does the system do what it should, at scale, without breaking? |
| Typical work | Retrieval, tool use, prompts and output contracts, evals, agent harnesses | Features, training runs, offline metrics, model serving, monitoring drift | Backend services, data modeling, integrations, performance |
| Uses models by | Calling a hosted or open model and shaping its context | Training or fine-tuning one | Integrating a model service another team owns |
| Main artifact | A feature that works, plus the eval set that proves it | A model in production, plus the metrics that justify it | A service in production, plus the tests and runbooks around it |

These are defaults drawn from the postings below, not laws. The distinction has a published history too. swyx's 2023 essay [The Rise of the AI Engineer](https://www.latent.space/p/ai-engineer) described a software engineer building with foundation models rather than training them. Chip Huyen's [AI Engineering](https://github.com/chiphuyen/aie-book) (O'Reilly, 2025) separates building applications on foundation models from traditional ML engineering's feature engineering and model training.

## AI engineer: builds the product around the model

[Ramp's Applied AI Engineer](https://jobs.ashbyhq.com/ramp/d204e136-2749-42de-82b4-88a0dd352090) sits on its Applied AI team and is asked to "ship full-stack AI projects end to end." It wants "a track record of working on full-stack AI projects, particularly those involving production use cases of LLMs." The posting also mentions infrastructure "supporting production-level inference and fine-tuning," so the line to ML work is not absolute.

[Brex's AI Engineer, Product](https://www.brex.com/careers/8606845002?gh_jid=8606845002) builds its Audit Agent, and the posting explains what the job is not. "The agent itself reasons; the surrounding product harness is what makes that reasoning useful, trustworthy, and operable for real customers." It adds that "the majority of your work will be backend: system design, data modeling, API shape."

[Clay's Software Engineer, Applied AI](https://jobs.ashbyhq.com/claylabs/5e07db20-d96a-4dff-b7d3-3bf1cdde6fc1) puts evaluation at the center: "Build and run evals that measure whether an agent actually completed the task correctly - not just whether the output looked plausible." [Samsara's AI Engineer, Customer Success](https://www.samsara.com/company/careers/roles/8024110?gh_jid=8024110) asks the person to "define success criteria and evaluation frameworks before you write the first line of code" and to "own the deployment, versioning, and rollback plan for every agent."

So the AI engineer is judged on whether a model-powered feature works for users, measured by evals the engineer designed. The model is usually someone else's. The product, the harness and the measurement are theirs.

## ML engineer: builds the model

[Ramp's Machine Learning Engineer](https://jobs.ashbyhq.com/ramp/2888b101-b1da-4e53-a02e-1bb9b1b5a951) sits in the Data department and works on fraud. The posting asks the person to "employ statistical and machine learning techniques on large datasets to discover patterns of fraud, platform abuse, and identity theft" and to "prototype and productionize machine learning models and rules-based systems." It requires "prior experience deploying Machine Learning models to production."

[Glean's Machine Learning Engineer, Search Quality](https://job-boards.greenhouse.io/gleanwork/jobs/4738120005) works on language models too, but at the model layer. Its list includes "train a model to capture interactions between signals in our ranking system" and "design smarter ways to domain-adapt language models to each customer's corpus." It asks for "proven ability to design, build, and ship production-ready models."

The verb that separates the two roles is train. An ML engineer changes the model's weights or builds a new one, and is judged on its metrics in production. Glean's posting shows the overlap: its team also works on "natural language question-answering, evaluation, and experimentation," which an AI engineer would recognize. Where the posting says train, rank or domain-adapt, expect an ML interview.

## Software engineer: builds the system everyone depends on

[Ramp's Software Engineer, Core Product](https://jobs.ashbyhq.com/ramp/5fe4c64e-9336-4384-9e6f-ff32eeb3fdae) works on cards, approvals and spend controls. The posting says the team's systems "handle billions of dollars in transactions and integrate deeply with Ramp's AI platform." One responsibility is to "integrate with Ramp's internal AI platform to automate spend policy enforcement and anomaly detection."

That line is the point. Software engineers increasingly call models, but the job is still judged on the system: correctness, performance and resilience of the transaction path. The AI platform is a dependency, owned elsewhere. Moving from this role to an AI engineer role means taking ownership of the model-powered part and the uncertainty that comes with it.

## Where the titles blur

Four patterns showed up often enough to watch for.

- Applied AI Engineer as a customer-facing role. [Anthropic's Applied AI Engineer, Enterprise Tech](https://job-boards.greenhouse.io/anthropic/jobs/5057647008) is "a trusted technical advisor to Enterprise Technology companies adopting the Claude API." [OpenAI's Applied AI Engineer, Startups](https://jobs.ashbyhq.com/openai/71e7252f-abb1-4b74-8e69-318413042357) sits within Go To Market and partners "with the Sales team across startup accounts." [Cognition's Applied AI Engineer](https://jobs.ashbyhq.com/cognition/811c3f5a-b26d-4162-b49b-93890a91794d) puts it plainly: "You don't demo Devin - you deploy it." These are closer to a [forward deployed engineer](/blog/forward-deployed-engineer-vs-solutions-architect-vs-consultant) than to a product engineer.
- AI engineer inside another function. [Stripe's AI Engineer](https://stripe.com/careers/listing/ai-engineer/8044460?gh_jid=8044460) sits within Solutions Architecture and builds tools for that team: "You'll discover where SAs lose time and build high-impact solutions." [Gusto's Enterprise Application AI Engineer](https://job-boards.greenhouse.io/gusto/jobs/7369003) sits in IT and connects systems such as NetSuite, Zuora and Jira to agents. The users are colleagues, not customers.
- GenAI engineer who also builds models. [Datadog's Staff GenAI Engineer](https://careers.datadoghq.com/detail/7974511/?gh_jid=7974511) in APM is asked to "build and benchmark GenAI/ML models using state-of-the-art techniques." That is ML engineering vocabulary under a GenAI title.
- Software engineer as the AI title. Clay's agent role is titled Software Engineer, Applied AI. Many product teams use a plain software engineer title for work that another company would call AI engineering.

None of these is a bad job. They are different jobs, and a candidate who prepares for the title prepares for the wrong interview.

## How to tell which one a posting really is

Count which column wins. Most postings mix a little of each.

| Signal in the posting | Leans AI engineer | Leans ML engineer | Leans software engineer |
| --- | --- | --- | --- |
| Main verbs | Ship, integrate, evaluate, orchestrate | Train, fine-tune, rank, benchmark | Design, scale, build, migrate |
| Named skills | LLM APIs, retrieval, evals, agent frameworks, MCP | PyTorch, statistics, feature pipelines, experimentation | Languages, databases, distributed systems |
| What is measured | Task completion, correctness, abstention, cost per request | Precision, recall, ranking quality, drift | Latency, uptime, correctness of the transaction |
| Who uses the output | Product users, or colleagues through an internal tool | A product surface that consumes predictions | Every other service and product team |
| Department | Engineering, Applied AI, sometimes a business function | Data, ML, research-adjacent teams | Engineering product or platform teams |

If the department is Go To Market or the posting mentions sales accounts, read it as a customer-facing role whatever the title says.

## Posted ranges at one employer

Ramp posted all three roles on September 26, 2026. The ranges are not labeled the same way, which is itself worth noticing.

| Ramp role | Posted range | What the posting says it covers |
| --- | --- | --- |
| Applied AI Engineer | $204.4K – $352K | Compensation with equity offered; final pay depends on location and level. Not labeled as base salary |
| Machine Learning Engineer | $200K – $330K | "Estimated Base Salary," with equity offered |
| Software Engineer, Core Product | $168K – $275K | "SF/NY Target Base Salary," with equity offered |

Each range spans several levels, and one of the three does not say it is base salary. The table shows how one employer posts pay on one day. It does not show that any of these roles pays more than another at the same level. The [AI engineer career path](/careers/ai-engineer) lists more AI engineer ranges from other employers, each with its scope.

## Which one fits your background

If you are a backend or product engineer, the AI engineer role is the nearest step. The gap is designing for output that is sometimes wrong and building the eval set that shows how often. The [how to become an AI engineer guide](/blog/how-to-become-an-ai-engineer) orders that work.

If you are a data scientist, look at both AI and ML engineering. You already think in datasets and error rates. ML engineering adds training and serving, while AI engineering adds running a product service.

If you are an ML engineer, you can move to AI engineering quickly, but expect to be asked about product scope, latency and cost more than model architecture.

## Next steps

- Follow the [AI engineer career path](/careers/ai-engineer): reading order, employer-posted pay and practice prompts.
- Read the [AI engineer career guide](/blog/ai-engineer-career-guide) for a portfolio project with test cases.
- Use the [AI engineer interview guide](/blog/ai-engineer-interview-guide) once you know which job you are applying for.
- If the agent side of the work appeals most, read the [agent engineer vs AI engineer vs ML engineer comparison](/blog/agent-engineer-vs-ai-engineer-vs-ml-engineer).