Build, buy or work with an FDE team: compare scope and ongoing responsibility

Start an AI proposal comparison with one shared scope of work. A software licence and a proposal covering data preparation, integration and operations do not buy the same thing. Their totals alone cannot establish which offers better value.

Choosing a product or a custom system is a product decision. Assigning delivery to an internal, external or joint team is a staffing and responsibility decision. These choices can combine: an internal team can integrate SaaS, and external engineers can help deploy an existing product. Forward deployed engineering (FDE) does not inherently mean self-hosting, training a model from scratch or outsourcing all development.

Define the work every option must complete

Consider a hypothetical task: helping support staff find answers in approved documents. Identify users, sources, access rights, review steps and who handles unanswered questions. Then evaluate every option against the same normal, exception and refusal scenarios. This is an assessment example, with no associated customer result.

The MAP function in NIST's AI RMF calls for understanding context, objectives, constraints and risk. MANAGE 2.1 also includes viable alternatives that do not use AI. This establishes a starting point for comparison without prescribing a sourcing arrangement.

Include existing search, process changes or manual work when they can complete the task. Keeping the current approach as an option makes the required improvement explicit.

Use data sensitivity to define what needs checking

Sensitive data calls for specific checks: processing locations, access, retention, deletion, subcontractors and incident arrangements. Establish which requirements come from organizational restrictions and which need regulatory or legal assessment, then request evidence from each option. The labels "SaaS" and "self-hosted" cannot establish suitability on their own.

External engineers also need appropriate access, agreed uses and a process for removing access when their work ends. GOVERN 6 and MANAGE 3 in the NIST AI RMF cover third-party software, data and ongoing monitoring. External delivery does not remove the organization's own responsibilities.

Test product and delivery arrangements against exceptions

This comparison worksheet is an editorial framework. Each row needs evidence; no score automatically identifies a best option.

For unusual workflows, compare configuration and extension options, additional development and the organization's willingness to change its process. Numerous exceptions do not automatically rule out SaaS. Where internal capacity is limited, compare training, hiring, ongoing support and joint delivery. A short engineering engagement cannot substitute for an operating arrangement that nobody has agreed to provide.

Compare costs over the same period and workload

Specify the period, currency, usage and service scope. Separate actual spending, quotes and estimates. Check licence and usage charges, data preparation, integration, customer effort, testing, human review, operations support and migration or shutdown work. Mark unquoted items as unresolved.

FinOps for AI describes spending that can span SaaS, model services, cloud and data centres, with more complex allocation and forecasting. Comparing model prices alone omits other inputs.

FinOps unit economics guidance connects technology spending to business work. For the support example, one comparison unit could be a case that meets the agreed acceptance criteria. State which costs, human tasks and failed attempts are included. This is a proposed method, not an observed customer result.

Saved working time may create available capacity. Whether it becomes cash savings depends on staffing decisions, and it does not establish realized revenue. Without comparable inputs, there is no basis to claim that internal development, SaaS or FDE is generally cheaper or faster.

Keep a decision record that others can review

Record viable options, exclusion reasons, evidence obtained, unanswered questions, responsible people and conditions that would trigger reassessment. If permissions, a critical integration or operations arrangements remain unclear, the next step can be gathering information or running a bounded evaluation. A production build need not be the immediate decision.

For team arrangements, read FDE versus an internal engineering team. For receiving responsibilities, see the AI project handover and acceptance guide. To discuss an actual scope, start with how we work and bring both the available evidence and the missing inputs.