The forward-deployed engineer works for the vendor. Know whose goals the advice serves.

A forward-deployed engineer (FDE) from an AI vendor is often the fastest way to get that vendor’s product working in your company. An FDE is not a neutral advisor. The vendor pays the engineer, and the vendor earns money when you use more of its product and renew the contract. Use vendor engineers to implement a decision that you already made. Use an independent person to make the decision.

In 2026, several large AI and software vendors expanded teams that work inside client companies to deploy AI. One short video compared this to a car-buying advisor who works for the dealership. The comparison is a little unfair to many good engineers. But the point about incentives is correct.

I am an independent consultant. I have a reason to prefer the independent option. I explain below where vendor engineers are the better choice.

What is a forward-deployed engineer?

A forward-deployed engineer is an engineer from a vendor who works with your team. The engineer often works inside your systems. The goal is to make the vendor’s product work for your use case. The idea is old. Some data and software companies used it for many years. AI vendors use it now because many companies cannot deploy AI systems alone.

Typical FDE work:

• Connect the vendor’s model or platform to your data and systems.
• Design prompts, agents and workflows for your use case.
• Fix problems in the first deployment.
• Train your team on the vendor’s tools.
• Find new use cases for the vendor’s product.

The last item is important. An FDE’s success is often measured by how much the client uses the product.

How do the incentives differ?

Neither incentive is bad. They are different. An FDE wants your project to succeed, because a successful project leads to more usage. But an FDE will rarely recommend a different vendor, a smaller model from a competitor or no AI at all.

Where is a forward-deployed engineer strong?

• Deep product knowledge. The FDE knows the vendor’s product better than anybody outside the vendor. The FDE knows the limits, the hidden settings and the next release.
• Direct access to the vendor. The FDE can escalate bugs, get quota increases and talk to the product team.
• Often low or no direct cost. The vendor sometimes includes FDE time in a large contract.
• Speed. For a use case that fits the product well, an FDE can deploy very quickly.

Where is a forward-deployed engineer weak?

• One product. The FDE solves each problem with the vendor’s product, even when another tool is better or cheaper.
• Cost growth. The FDE has no reason to reduce your usage. The FDE may select the largest model, because it gives the best demo.
• Lock-in. The design uses the vendor’s special features. Moving to another vendor later becomes expensive.
• Short stay. The FDE moves to the next client. Knowledge leaves unless your team learns it.
• Limited view of your business. The FDE knows the product well and your business less well.

What questions should I ask a vendor before I accept FDE help?

1. How is the FDE’s performance measured? Usage, revenue, customer satisfaction or project success?
2. Who owns the code, prompts and configuration that the FDE builds? They must be yours.
3. Where does the work live? In your repositories and accounts, or in the vendor’s?
4. Does the design use features that only this vendor has? Which ones, and why?
5. What is the cost per task at full volume? And what is the plan to reduce it?
6. What happens when the FDE leaves? Who on your team will own the system?
7. Can we test another model against the same eval set? A confident vendor says yes.

When is a forward-deployed engineer the correct choice?

Use an FDE in these cases:

• You already selected the vendor with an independent test, and the decision is final.
• The use case fits the product closely. For example, a document workflow that the product was built for.
• You have an internal owner who works with the FDE each day and learns the system.
• You have an eval set and a cost limit that you control.

In these cases, the FDE implements your decision fast, and you keep control.

When do I need independent advice?

Get independent advice in these cases:

• You have not selected a vendor yet. Do not let a vendor’s engineer run the selection.
• You must decide build or buy. An FDE always prefers buy, with its own product.
• The cost is growing fast. An independent person can test smaller models and other vendors.
• You want a plan to change vendors later. An FDE has no reason to make this easy.
• You need someone to say “do not do this.” An independent advisor can recommend no AI.

How do I use both safely?

Many companies get the best result from a combination. The independent person owns the decisions. The FDE implements them.

The independent owner can be your CTO, a senior engineer with AI experience or a fractional AI CTO. The important property is that this person does not get paid by the vendor.

How do I protect against lock-in?

Lock-in is not always bad. Sometimes the best product is worth the dependency. But make the choice with open eyes. Use these protections:

• Keep the eval set yourself. It lets you test any model or vendor against the same examples.
• Keep prompts and configuration in your repository. Not only inside the vendor’s console.
• Use a thin layer between your code and the vendor’s API. It makes a later change easier.
• Write down each vendor-only feature that the design uses. With the reason and the cost to replace it.
• Test one alternative each year. Run your eval set against another model or vendor. Know your options before the renewal.

The harness as a service article explains why the layer around the model, not the model itself, is the asset to own.

How do I compare the total cost?

FDE time can look free. It is usually part of a larger contract. Compare the total cost over three years, not the cost of the first project.

In many cases, the independent owner pays for itself through lower usage costs and a lower cost to change later. In some cases it does not. For a small, simple deployment that fits the product well, the FDE alone can be the cheapest option. Calculate both for your case.

What does an independent review of FDE work check?

If an FDE already built a system for you, an independent review can take one to two weeks. It checks these items:

• Model choice per step. Is the largest model necessary for each step? What does the eval set say about smaller models?
• Cost per task. At the current volume and at full volume.
• Vendor-only features. A list, with the cost to replace each one.
• Data flow. Which data goes to the vendor, where it is stored and for how long.
• Ownership. Are the code, prompts and configuration in your accounts?
• Evals and monitoring. Do they exist, and does your team control them?
• Handover. Can your team run and change the system without the FDE?

The result is a short list of changes, ordered by value. Most of these changes are small. The FDE can often make them.

What about vendor-neutral claims?

Some consulting firms and agencies say they are vendor neutral. Check this claim:

• Does the firm have partner status with a vendor? Partner status often includes sales targets, discounts or referral fees.
• Do all of the firm’s case studies use the same vendor?
• Does the firm resell licenses?
• Will the firm write in the contract that it receives no money from the vendors it recommends?

A firm can have partnerships and still give good advice. But it must tell you. The guide to fake AI consultants has more checks.

A hypothetical example

A mid-size insurance company wanted to automate the intake of claim documents. Its cloud vendor offered FDE support as part of a large contract. In four weeks, the FDE built a working system on the vendor’s AI platform. It used the vendor’s largest model for all steps.

Before the launch to all claims, the company asked an independent person to review the design. The review found three things. First, two of the five steps worked equally well with a much smaller model, at a fraction of the cost per claim. Second, the system used a proprietary document format that only this vendor supported. Third, there was no eval set that the company controlled.

The company did not replace the FDE. It asked the FDE to change the two steps to the smaller model and to store the extracted data in an open format. The independent reviewer built an eval set from 300 real claims. The FDE ran the changes against it. The system launched a few weeks later at a lower cost, with an eval set that the company could use to test any future vendor.

Next step

If a vendor offers FDE support, ask the seven questions above before you accept. If you have not selected a vendor yet, get independent advice first. For more on the choice between an independent person and an agency, read independent AI consultant or agency? For my own approach, read how my fractional AI CTO engagement works.