In this article

FDE or SaaS: how to scope an AI integration engagement

SaaS supplies a product; forward deployed engineering adapts a solution to business operations. They can work together. A useful engagement defines the process, deliverables, acceptance criteria, operating ownership and exit arrangements.

An engineer and an operations manager working on a laptop in a logistics workshop
Editorial illustration

Decide what you need to buy

Forward deployed engineering brings engineering work close to the client’s operating context: observing the process, adapting a solution, integrating software and supporting real usage. SaaS provides a product whose capabilities and terms are set by its vendor.

These approaches can be combined. An organization can retain its CRM, use an existing AI capability and bring in engineers to connect the operations. First identify the precise gap between what the product supports and the work your teams need to complete.

Why the topic is receiving attention

In What FDE means at OpenAI, published on October 22, 2026, OpenAI describes closing the gap between model capabilities and reliable customer use. It also discusses replacing custom implementations with shared components. This is a vendor’s operating practice, not a contract standard for every engagement.

For a buyer, the implication is useful: a good engagement can lead to more reuse of existing software. The amount of custom code produced is an incomplete measure of value.

A decision matrix for SaaS, FDE and hybrid delivery

Your situation First option to examine Point to verify
Common need already covered by a product Configured SaaS Adoption, data access and export
Cross-system process with specific business rules Integration with an FDE approach Staff availability and API access
Suitable AI feature but missing business connection SaaS plus targeted development Responsibilities between vendor and integrator
Unstable process or poorly understood data Discovery before implementation Avoid automating undefined rules

A distributor might already have inventory software. If customer enquiries require information from three sources, the missing capability may be an interface and controlled reconciliation. Replacing the entire platform would create a much larger project. This example illustrates the decision rather than reporting a completed customer engagement.

Scope the work around four acceptable deliverables

1. A process map and a baseline

The initial deliverable describes the trigger, steps, owners, exceptions and software. The business owner validates it. Measure an appropriate baseline: processing time, repeated entry, incomplete cases or time spent searching. Without one, the final assessment may remain subjective.

2. Technical scope and action rules

The technical brief specifies sources, connectors, environments and permissions. It distinguishes what AI proposes from what the system executes. Preparing invoice fields and changing an approved accounting record carry different responsibilities.

If MCP is used, ask who manages identities and authorizations. The MCP authorization specification addresses an access mechanism; the business rule allowing a credit note or blocking a price change still needs to be implemented in your system.

3. A pilot evaluated on representative cases

The pilot should cover ordinary cases and exceptions. Agree in advance what an accepted case means, which errors block release and who resolves disagreements.

A short weekly review can help decisions move forward. This is a suggested working cadence, not a commitment to production delivery in four weeks. Vendor access and internal approvals may determine the critical path.

4. Operations the client can take over

Handover covers configuration, relevant source code, evaluation cases, alerts and a procedure for returning to manual work. Test the operating guide with someone who did not build the solution. If that person cannot identify an incident and reach the right owner, handover is incomplete.

Assign responsibilities before starting

The client appoints a business owner and makes process experts available. The supplier defines its development, integration and support responsibilities. System administrators approve permissions. Appropriate specialists review data processing and applicable obligations.

Include these contributions in the schedule. Engineers cannot independently resolve an ambiguous commercial policy or create access that your software vendor does not permit.

Also define the end of the engagement: which accounts belong to the client, which components require licenses, which exports are available and what support response has been agreed? An explicit dependency can be managed; a dependency discovered during exit is harder to resolve.

Buy the ability to keep improving

Acceptance should cover a usable process and a team able to maintain it. Assess outcomes against the baseline, correction work and recurring operating costs. A useful pilot may justify expanding the scope; it may also demonstrate that configured SaaS is sufficient.

MTECH is a product and AI studio in Paris. Our workflow automation and integration work connects software with operations. Our article on AI agent projects and the FDE approach explains the role of discovery and adoption.

To scope an engagement, show us one cross-system process, its owner and the software involved. We can then discuss deliverables that allow your team to evaluate the outcome.

FAQ

Frequently asked questions

What is a forward deployed engineering engagement?

It is engineering work carried out alongside a client’s teams to understand a process, build its integration, deploy it and support real use. Scope and responsibilities should be defined in the engagement.

Do we need to replace our SaaS to work with an FDE?

No. An FDE engagement can retain the SaaS and build missing connections, rules or interfaces. The decision depends on the gap between the available product and the process it needs to serve.

Can an FDE engagement be completed in four weeks?

A limited initial scope can sometimes be tested quickly, but four weeks is not a standard duration. Access, data, approvals and operating requirements determine the schedule. Ask for stages with explicit acceptance criteria.