AI Careers: FDE, AI Integration, Voice Agents, and the Work Behind Them

AI careers now include forward deployed engineering, AI integration, voice agent development, agent orchestration, evaluation, retrieval, security, and AI search visibility. Some are newer specializations. Others are established professions taking on different tools and responsibilities. The useful question is what work each person actually owns.

As the founder of Suede AI, I keep coming back to the same question: who owns the work after the demo?

A voice agent has to handle an interruption. A workflow has to recover when an API fails. A knowledge assistant has to respect access permissions. Someone has to know whether the system completed the job, how much it cost, and what happens when it gets something wrong.

That's where I see the career opportunity: closing the gap between what a model can do and what a business can reliably use.

Here are nine directions I'd take seriously, with the skills and portfolio evidence that make each one concrete.

1. Forward deployed engineer: take AI into the customer's operation

A forward deployed engineer, or FDE, works closely with customers to turn an operational problem into deployed software. The role combines engineering, discovery, integration, and delivery.

Forward deployment is an established model. Palantir describes its forward deployed engineers as responsible for customers' technical and operational outcomes. OpenAI's current FDE description applies that model to AI, covering scoping, system design, implementation, rollout, and adoption. Palantir's role overview, OpenAI's FDE description.

Think of a distributor whose staff manually classify incoming requests. An FDE would examine the process, build the connections, define exceptions, and help the team operate the resulting system.

Skills to build: production software, APIs, data modeling, debugging, clear communication, and the judgment to narrow a messy problem.

First portfolio project: take one fictional business workflow from discovery notes through a working deployment. Include the requirements, architecture, test results, and operating guide. Show how someone besides you would use it.

2. AI integration engineer: connect models to useful systems

An AI integration engineer connects AI capabilities to business applications and data. The title overlaps with application engineering, integration engineering, and solutions engineering. Read the responsibilities in a job description carefully.

The work might include connecting a support system to a knowledge base, passing validated data into a CRM, or exposing a narrowly scoped tool to an agent.

Model Context Protocol, or MCP, is one interface worth understanding. Its architecture connects AI applications with servers that expose capabilities such as tools and resources. It's one part of an integration, alongside authentication, permissions, validation, and monitoring. MCP architecture documentation.

Skills to build: Python or TypeScript, HTTP, webhooks, SQL, authentication, structured outputs, and retry handling.

First portfolio project: build a support-triage integration using synthetic tickets. Classify each request, look up an approved answer, and draft a response for review. Demonstrate that a retried webhook cannot create duplicate tickets.

The failure handling tells me more than the happy-path screenshot.

3. Voice agent developer: make conversation work in real time

A voice agent developer builds systems that listen, respond, and use tools during spoken conversations. This combines software engineering with speech technology and conversation design.

Voice creates a very visible reliability problem. People pause, correct themselves, talk over the agent, or ask for a person. LiveKit's documentation treats turn detection and interruption handling as explicit parts of voice-agent implementation. LiveKit's turn-handling guide.

A useful deliverable could be an appointment assistant that checks availability, confirms the requested time, and hands off when it cannot resolve the request.

Skills to build: real-time audio, speech recognition, speech synthesis or speech-to-speech systems, telephony, tool calls, latency measurement, and conversational recovery.

First portfolio project: build a voice booking demo against a test calendar. Test background noise, interruptions, changed dates, unavailable slots, and a request for human help. Report task completion and response delay.

Use a voice you have permission to use. Build disclosure, consent, and data-handling requirements into the project from the start.

4. Agent workflow engineer: control how work moves

An agent workflow or orchestration engineer designs how models, tools, rules, and people work together. These labels describe a developing specialization, with substantial overlap across AI and backend engineering.

Anthropic distinguishes predefined workflows from agents that choose their next steps dynamically. That distinction matters when deciding how much freedom a system needs. Anthropic's guide to effective agents.

A purchase-request workflow might gather information, check a policy, prepare a recommendation, and stop for approval. The engineering includes state, limits, retries, and the record of what happened.

Skills to build: workflow design, queues, state machines, tool contracts, observability, and human approval steps.

First portfolio project: build a research-to-brief workflow with source retrieval, citation checking, an approval checkpoint, and a spending limit. Interrupt a run midway and demonstrate how it resumes safely.

Start with the simplest design that completes the job. Every additional agent introduces another interaction you have to understand and test.

5. AI evaluation and reliability engineer: prove the system works

AI evaluation engineers define success, build test sets, investigate failures, and measure regressions. Reliability work also covers production monitoring, incident response, and recovery.

This builds on testing, quality engineering, and machine-learning evaluation. Agent systems add a particular challenge: an impressive response can conceal an incorrect action. Anthropic's evaluation guidance separates the agent's transcript from the actual outcome and describes code-based, model-based, and human graders. Anthropic's agent evaluation guide.

Skills to build: test design, Python, basic statistics, domain-specific scoring, trace inspection, and clear failure reporting.

First portfolio project: create an evaluation suite for a support assistant using synthetic data. Test correct answers, unsupported questions, attempted policy violations, and failed tool calls. Repeat trials and disclose the sample size.

Publish the failures too. A useful report explains what broke, what changed, and whether the fix damaged something that previously worked.

6. AI knowledge and RAG engineer: make the right information available

A retrieval-augmented generation, or RAG, engineer builds the system that finds relevant information for a model to use. This specialization draws on search, data engineering, and information architecture.

Microsoft's RAG guidance describes practical challenges including query understanding, multiple data sources, response time, and access control. Those are the problems behind a seemingly simple “ask your documents” interface. Microsoft's RAG overview.

The work includes preparing documents, preserving metadata, designing retrieval, handling updates, and checking whether answers actually follow the retrieved evidence.

Skills to build: ingestion pipelines, keyword and vector search, chunking, ranking, permissions, and retrieval evaluation.

First portfolio project: build an assistant over public documentation. Show citations, test questions that have no answer, and update a source to demonstrate freshness. Add a synthetic restricted document and prove that an unauthorized user cannot retrieve it.

A larger document collection is only useful if the right material reaches the right person.

7. AI security and governance: define boundaries and accountability

AI security and AI governance are related disciplines with different responsibilities. Security focuses on threats and technical defenses. Governance establishes ownership, review, acceptable use, and risk-management processes.

These professions already exist. AI adds systems that interpret untrusted content, generate outputs, and sometimes take actions through connected tools. OWASP documents excessive agency as a risk arising from too much functionality, permission, or autonomy. NIST's AI Risk Management Framework provides a voluntary foundation for managing AI risks. OWASP on excessive agency, NIST AI Risk Management Framework.

Skills to build: threat modeling, least-privilege access, data handling, audit trails, risk assessment, and communication across technical and business teams.

First portfolio project: threat-model a small document assistant. Demonstrate an attempted prompt injection in a safe test environment, limit tool permissions, and document escalation procedures.

Be precise about what your controls prove. A checklist alone cannot establish that a system is secure or legally compliant.

8. AI search visibility specialist: help people find trustworthy answers

An AI search visibility specialist works on how a business is discovered and represented in search and AI-generated answers. Related terms include search engine optimization, answer engine optimization, and generative engine optimization: SEO, AEO, and GEO.

The labels overlap. The deliverables should be specific: crawlable pages, clear service information, accurate author and company identities, useful answers, credible sources, and measurement.

Google says its existing SEO best practices remain relevant to AI Overviews and AI Mode. It does not require special AI markup or new AI text files for inclusion. Google's guidance on AI features.

Skills to build: technical SEO, search intent research, editing, analytics, structured data, and source verification.

First portfolio project: audit a site you control, improve several pages, and record a fixed set of questions across search experiences over time. Separate being cited, being mentioned, and receiving a qualified visit. Record dates and repeated observations.

The work is to strengthen source quality, topical relevance, and clear answers, then measure visibility across the engines that matter to the business.

9. AI implementation lead: make adoption somebody's responsibility

An AI implementation lead or AI product manager connects the use case, users, engineering work, and operating process. Product and implementation management are established careers with AI-specific decisions to make.

Which task is worth changing? What is the current baseline? Who reviews uncertain results? Who supports the system after launch? When should the team stop the rollout?

Skills to build: process mapping, user research, requirements, basic data analysis, change management, and enough technical fluency to question an architecture.

First portfolio project: prepare a rollout plan for a fictional business. Include the existing process, a narrow pilot, success criteria, total operating costs, training, and a rollback plan.

This can be a practical direction for someone coming from operations or product. Technical literacy still matters. You need to understand what you're asking people to rely on.

How to choose an AI career path

Start with the work you already understand.

• Software developers can explore integration, forward deployment, orchestration, voice, or retrieval.
• QA engineers and analysts can explore evaluation and reliability.
• Security and risk professionals can develop an AI specialization.
• Marketers and editors can build technical SEO and AI visibility skills.
• Operators and product managers can focus on implementation and adoption.

Then build one small system around a recognizable problem. Use public or synthetic data. Write down its limits. Measure the outcome. Let another person test it without you narrating every step.

That's the practical thread connecting my work at Suede AI, from Agent Studio to AI engineering and search visibility: a capability becomes useful when somebody can understand it, operate it, and hold it accountable.

The title may change. The responsibility remains.

About the author: Jason Colapietro is the founder of Suede AI. His work spans AI integration, agent workflows, search visibility, and products for creators.