# Before You Apply for a Forward Deployed Engineer Job, Try This Customer Call

> Source: <https://dev.to/emilywoodsnyc/before-you-apply-for-a-forward-deployed-engineer-job-try-this-customer-call-2i7m>
> Published: 2026-11-16 06:10:17+00:00

Imagine joining a customer call where someone wants an AI agent by Friday, their data lives in three different systems, and nobody agrees on what the agent should actually do.

You know how to write the code. You’re less sure which part of this conversation is supposed to become code.

That second problem is worth paying attention to if forward deployed engineer keeps appearing in your job search.

A recent Reddit discussion about the role asked a wonderfully reasonable question about how it differs from software engineering. Honestly, the title alone doesn’t help much. It sounds like someone sent your laptop on a business trip.

I prefer to investigate the work behind it. What do you build, whose problem are you solving, and what happens after the demo?

Let’s walk through a fictional customer call and turn it into something you could actually practice. This is my exercise, not a reported interview question or an employer’s assessment.

Forward deployed engineering commonly puts engineering work close to a customer’s environment and workflow. The mix can include discovery, implementation, integration, deployment, and explaining technical decisions to people who have different priorities.

For a concrete example, Anthropic’s Staff Forward Deployed Engineer role, it describes customer discovery, architecture, production builds, and handing over systems customers can operate. It also includes shaping scope and feeding repeatable deployment patterns back into product and engineering.

That is one employer’s Staff role, not a universal FDE job description. Don’t assume its seniority, responsibilities, or travel expectations apply to every opening.

Plenty of software engineers already work with customers. Plenty of solutions engineers write code. The useful distinction is the responsibility this particular team expects you to own, rather than a tidy boundary between job titles.

So let’s give ourselves a customer and see where the responsibility takes us.

Suppose a support manager wants AI to handle incoming tickets automatically. Their agents use an order system, a policy wiki, and a help desk. A polished demo is expected on Friday.

Jumping straight to retrieval, tool calls, and model selection is tempting because those are decisions you can immediately start making.

Instead, ask the manager to walk you through one recent ticket from arrival to resolution. Follow the work closely enough to notice where someone searches, waits, copies information, or asks another person for permission.

In our fictional example, the biggest delay is finding the correct order and the current return policy. Writing the reply is relatively quick. Refund approval remains a separate human decision.

Now you have a more useful first target. Help an agent assemble the right evidence and prepare a reply, while leaving approval and sending under human control.

Before leaving the conversation, clarify three things with the customer.

You haven’t avoided engineering. You’ve made the engineering problem less imaginary.

If you’re exploring this career move, try these [FDE role-fit and client-work questions](https://prachub.com/interview-questions/forward-deployed-engineer-screen-role-fit-client-work-and-databricks-projects?utm_source=devto&utm_campaign=v1012). Start with the role-understanding and customer-experience prompts, then answer using something you genuinely built. Keep this fictional exercise separate from your work history.

For the first pilot, choose one ticket category, a small approved group of support agents, and a controlled set of data. If access approval is still pending, use synthetic examples for the demo and make that limitation visible.

The proposed flow takes an eligible ticket, retrieves only information the user is permitted to see, and produces a draft with its supporting sources. An agent reviews and edits the reply before sending it. Missing or conflicting evidence sends the case back to the normal workflow.

Automatic refunds are outside this pilot. So are unrelated ticket categories and broad access to customer records.

That boundary changes the technical work. You need identity and access checks, reliable order matching, current policy references, a review interface, and a way to handle missing data. A model producing fluent paragraphs does not settle any of those questions.

Human review also needs a usable design. Show the evidence beside the draft, make uncertainty visible, and let the reviewer reject it easily. Adding an approve button to a confusing screen is a rather optimistic definition of oversight.

Notice how much of the challenge involves ordinary software engineering. Integrations can time out. Permissions can change. A source document can be outdated. Someone still needs to make those cases behave sensibly.

Now imagine the manager insists that the demo must show automatic sending because that is what leadership expects.

This is the part where communication becomes a design skill.

Explain what can be demonstrated safely with synthetic examples, what can be piloted with human review, and what remains unverified before live automation. Connect each option to the customer’s objective instead of burying the decision in technical vocabulary.

For our example, a demo can illustrate the intended end to end experience without touching live tickets. A limited pilot can test whether assisted drafting saves time. Unattended sending would require a separate decision supported by evidence and appropriate controls.

Those are three different deliverables. Write down which one Friday means, who accepts it, and what would stop it from going ahead.

I like this exercise because it makes scope visible without turning every disagreement into a dramatic refusal. You’re helping the customer choose a commitment they can understand.

A draft appearing in two seconds looks lovely. It helps much less if reviewing it takes longer than writing the original reply.

Compare similar ticket categories and include the entire handling process. Record lookup time, drafting time, review time, and the need for corrections or escalation. Keep changes in ticket difficulty in mind when comparing results.

For this pilot, useful evidence could include whether replies are supported by current policy, whether access restrictions hold, and whether agents continue using the tool after the initial novelty wears off.

Do not make one attractive average stand in for every outcome. A time saving workflow can still be unacceptable if it exposes information to the wrong user or confidently uses an outdated rule.

Agree on acceptance criteria before reviewing the results, including cases that block expansion. In a practice project, describe the criteria as proposed criteria. Don’t invent successful deployment numbers to make the write-up sound impressive.

You can show good judgment with a modest result, or even with a well-supported decision that the pilot should not expand yet.

The handoff is part of the engineering problem, so include it before declaring the project finished.

Someone needs to own credentials, policy updates, monitoring, user support, and the fallback when a dependency is unavailable. That person also needs enough documentation and access to operate the system.

In our exercise, ask the receiving team to disable assisted drafting and handle a ticket through the normal workflow. Then ask them to identify a failed retrieval and find its cause without needing the original builder on a call.

If they cannot do either, the handoff still has work left.

This also gives you better questions for an actual FDE recruiter conversation. Ask what the last hire shipped, how much time is spent building versus advising, and who owns customer-specific code after launch. Ask about travel and support expectations explicitly rather than trying to decode them from the title.

Give yourself 45 minutes with the fictional support scenario. Spend ten minutes writing discovery questions, fifteen outlining the smallest useful pilot, ten defining evidence and failure cases, and five recording a plain-language update for the customer.

Your output can be one page and a rough diagram. Include the original request, the problem you uncovered, the proposed boundary, the evidence you need, and the next decision’s owner.

Then look at the exercise honestly. Did you enjoy turning a messy conversation into a buildable plan? Did you want to investigate the integration details? Could you explain a limitation without sounding dismissive or promising something you hadn’t established?

Those questions won’t certify you for a role. They can help you decide what to investigate next and which real examples to bring into an interview.

Your next job does not need a more exciting title to be a better fit. It needs work you want to get good at, and forward deployed engineering deserves that same practical inspection.

*~ Emily*
