When integrations fall behind sales
The pattern is familiar to anyone who sells AI software to enterprises. The product demos well, contracts get signed, and then each customer turns out to need something specific: single sign-on through their identity provider, documents pulled from a records system nobody has an API for, model calls routed through their approved region, outputs written back into a ticketing or case-management tool. The solutions team is spread across too many accounts, the product team is building the roadmap, and the customer’s renewal clock is running.
Forward deployment is the response to that gap. An engineer works inside one customer deployment, alongside the customer’s own engineers, until the integration is running and someone on your side can support it. The discipline is less about writing clever code and more about closing the distance between a product that works in your environment and one that works in theirs.
What the work involves
Every deployment is different, but the work runs through the same four areas:
- Access and identity. Getting credentials through the customer’s process, mapping their roles onto your product’s permissions, and making sure the AI component cannot see data the user cannot.
- Data paths. Connecting the customer’s sources, deciding what is copied versus queried in place, and documenting what personal or confidential data crosses which boundary.
- Model behaviour in their context. Checking that outputs hold up on the customer’s own documents and vocabulary, with a small evaluation set agreed with their subject experts, not just your demo data.
- Failure paths. Deciding what happens when a source is down, a model call times out or an output fails validation, and making those failures visible to the people who will support it.
I work in your codebase through your review process, and in the customer’s environment through theirs. Product changes that the deployment needs are raised with your product owner, not merged quietly into a customer branch.
The signature deliverable
You receive a customer-embedded implementation with agreed acceptance criteria: the running integration, the acceptance record, an ownership map and a runbook. Illustrative example of an ownership map extract:
| Integration point | Owner | Credential held by | Failure signal | First responder |
|---|---|---|---|---|
| SSO and role mapping | Customer IT | Customer | Login errors in product audit log | Customer IT, then your support |
| Document sync from records system | Your product | Service account in customer tenant | Sync lag alert above agreed threshold | Your support team |
| Model calls via customer’s approved region | Your product | Your platform | Error rate and latency dashboard | Your platform on-call |
| Write-back to case tool | Shared | Customer-issued API key | Failed write queue | Your support, escalating to customer |
Illustrative example. Not taken from a client engagement.
How acceptance is judged
Acceptance criteria are written before the build starts and signed by you and the customer’s named contact. They usually combine functional checks (the integration completes the agreed workflows end to end), quality checks (outputs meet an agreed standard on the customer’s evaluation set), and operational checks (alerts fire, the runbook procedures work when rehearsed). If a criterion cannot be met for reasons outside the scope, that is recorded with the evidence rather than quietly dropped.
Ownership and handover
Your organisation owns the code, the configuration and the customer relationship throughout. The handover goes to a named person on your side, usually in solutions or support, and includes a walk-through of the runbook and the ownership map. The decision log records what was customised for this customer, so the next deployment can reuse what worked. If the internal team needs a more structured transfer, AI system handover and internal ownership covers that as its own engagement.
When to choose something else
If you want an engineer inside your team under your manager, see hire a forward-deployed AI engineer. If you are a software vendor looking for a longer-term delivery partner across many customer deployments, see AI deployment partner for software vendors. If the question is whether to integrate a product at all, start with build, buy or integrate.