Forward-deployed engineer vs AI consultant
A consultant can tell you what to build. Only someone inside your systems finds out whether it works.
The short answer
An AI consultant tells you what to do and why. A forward-deployed engineer does the work in your systems and leaves a working result. Many engagements need both, in that order.
Decision table
| Question | Forward-deployed engineer | AI consultant |
|---|---|---|
| What it is | An engineer who writes and runs code in the client's environment | An adviser who assesses, recommends and plans |
| Who owns the outcome | Shared with the client's named owner | The client, who decides on the recommendation |
| Typical duration | Weeks to months, on one use case | Days to weeks for an assessment |
| What you get at the end | A running system, runbook and handover | A written assessment, options and a plan |
| Main risk | Building the wrong thing if the use case is unclear | A plan that nobody builds |
| Best for | A chosen use case that must reach production | Deciding where to start, and whether to start |
Use cases for a forward-deployed engineer
- A bank's document checks. The operations team has chosen the KYC checker. An engineer builds it in the bank's own case system, and an officer approves every flag before a customer sees it.
- Dealer service replies for a manufacturer. Requests arrive by email in several formats. The engineer connects the agent to the parts catalogue the service team already uses.
- Claim file summaries for an insurer. The engineer builds the summary step and the approval screen, then hands both to a named claims lead with a runbook.
- An engineering assistant over private repositories. The engineer sets the permissions, so the agent sees only the repositories the team owns.
Use cases for an AI consultant
- Choosing between three candidate use cases. Leadership has budget for one. A consultant scores each on value, data access and risk, and writes the recommendation.
- A board asks whether an agent is worth building. An independent assessment, written so the board can decide without a sales conversation.
- A readiness check before any build. Two weeks on one workflow, to find out whether the data, the approvals and the owner exist.
- A roadmap across several business units. The consultant maps which units share data and which need separate controls.
Trade-offs
Forward-deployed engineer
Advantages. The agent is tested on your real data inside your systems. The engagement leaves a working system and a handover, not only a report. If the handover is done properly, the knowledge stays with your team.
Disadvantages. The engineer needs access to your systems, which needs a security review first. The first working system takes longer than a plan. A weak handover leaves the system dependent on one person.
AI consultant
Advantages. The recommendation is not tied to a build, so it can be honest about whether to build at all. It is quicker and cheaper to start than a build. It is useful when several options compete and the choice needs a written comparison.
Disadvantages. A plan does not prove that the data or the systems will cooperate. The recommendation often waits for a build team that has not been hired. A report can be filed without anyone building the first slice.
How each option runs
Forward-deployed engineer, week by week
A typical shape, not a fixed plan:
- Week 1. Map one workflow from trigger to outcome, sample real records, and list the personal data each step reads.
- Weeks 2 and 3. Build the first working slice inside the client's environment. The named owner reviews each change as it lands.
- Week 4 onward. Run the evaluation set, fix what fails, and prepare the handover pack.
AI consultant, week by week
- Week 1. Talk to the owners of the candidate workflows and collect the evidence each one already has.
- Week 2. Score each candidate on value, data access and risk.
- Week 3. Write the recommendation, the options rejected and the risks, with a named owner for each decision.
What drives the effort
- How many systems the agent must read from or write to.
- How clean the data is, and who controls access to it.
- How many actions need a person to approve them before they run.
- How quickly the owner can make decisions.
This site does not publish prices. The engagements page describes what each shape includes.
A checklist before you choose
- Do you have a named owner for the decision the agent will support?
- Is the workflow already chosen, or are you still choosing between several?
- Can your team give test access to real records within a week?
- Will someone on your side run the system after launch?
If you answered no to the first two, start with the consultant. If you answered yes to all four, the forward-deployed engineer is the usual next step.
What people get wrong
- They ask for a strategy and expect code. A strategy needs a build step after it.
- They build before the readiness question is answered. The data, the approvals and the owner are then discovered late.
- They treat the consultant's report as the finish line. The report is the start of the build.
A hybrid that often works
Start with a short readiness review that ends in a written decision to proceed, reframe or pause. If the decision is to proceed, the same team, or a different one, builds the first working slice and hands it over. Make sure one person is accountable from the review to the handover.
Common questions
Is a readiness review a consultant's job or an engineer's?
It is both in practice. It needs someone who understands the business question and someone who can test the data and the architecture. A single person with both skills is rare, so many teams use two people for it.
Can a forward-deployed engineer give a strategy?
Yes, within the scope of the use case. A strategy for the whole organisation is a different piece of work.
What should the output of an AI consultant be?
A written recommendation, the reasons for it, the options rejected and the risks, with a named owner for each decision.
How do I know the build is working?
Ask for the evaluation set, the results against it, and the record of who approved each action the system takes.