A prototype appears quickly. A month later, staff still recheck every field, the integration relies on someone's personal credentials, and nobody owns support.
A forward deployed AI engineer (FDE) works directly with customers to understand a problem, build software, and deploy it in their environment. Anthropic's role description includes discovery and production engineering. The exact remit varies by engagement.
Before hiring, agree five deliverables: a workflow baseline, a credible business case, a measured pilot, approved access, and an operating handover. Here is what to ask for.
1. One workflow with an accepted output
“Roll out AI” is too broad to commission. Pick a recurring process with a named owner and a result they can judge.
Consider supplier document intake. A useful initial scope could be:
- Receive an eligible document pack.
- Propose required fields with source references.
- Flag missing or conflicting information.
- Let an authorised reviewer correct and accept the result.
- Save the record and review history.
Payments and contract acceptance remain outside this scope.
Ask the engineer to observe the current process, including exceptions. The first deliverable should state volume, handling time, eligibility, quality requirements, and who resolves ambiguity. If those facts are unknown, baseline measurement is the first project.
2. A business case that includes review
The numbers below are hypothetical planning assumptions, not customer results.
| Monthly workload and cost | Assumption |
|---|---|
| Eligible document packs | 450 |
| Current handling time per pack | 20 minutes |
| Assisted handling, including routine review | 6 minutes |
| Extra work on 20% of packs | 20 minutes each |
| Loaded labour rate | £40/hour |
| Ongoing platform and support cost | £1,200/month |
| Initial implementation and training | £12,000 |
Assisted work averages 6 + (20% × 20) = 10 minutes per pack. The improvement releases 75 hours a month, valued at £3,000 at the assumed labour rate.
Across twelve full operating months: £36,000 capacity value − £14,400 running costs − £12,000 implementation = £9,600 net capacity value. An actual plan must also account for rollout delay and transition costs.
That is not automatically cash saved. State how the hours will be used: less overtime, an avoided external cost, or a larger queue handled by the same team. Do not count the same hours twice.
Test the downside: at 150 eligible packs, capacity value falls to £1,000 a month—below ongoing costs before recovering implementation spending. Ask what volume or performance would justify stopping.
3. Pilot evidence that supports a decision
Agree the evaluation before tuning the system. Include ordinary documents, poor scans, missing pages, contradictions, and receiving-system failures.
Measure the whole workflow:
- Quality: correct fields and supporting source references.
- Effort: review time, corrections, and exception handling.
- Operations: delays, failed writes, recovery, and support cost.
- Use: whether staff can complete real work with it.
Anthropic's evaluation guidance explains tasks, trials, and graders, including why repeated trials matter. Keep separate evaluation cases for release decisions.
Shadow processing can test proposals alongside the existing process. An assisted pilot tests the review experience. Choose the mode for the risk and keep users involved early; shadow results alone do not prove adoption or time savings.
Our document AI evaluation guide goes deeper on source evidence and reviewer corrections.
4. Approved access and a tool choice you can maintain
Bring IT a specific data flow: sources, identities, external processing, retention, writes, and revocation. Enforce permissions and business rules outside the model.
The NCSC secure AI guidelines cover design through ongoing operation. Put monitoring, incident handling, and recovery in the delivery budget.
Ask why the proposed approach fits your environment. An approved platform extension, a small integration, or a custom pipeline may each be appropriate. Compare total cost, controls, exportability, and support. A polished demo does not settle that choice.
5. A handover your team can operate
Name the business owner, access approver, reviewer, and operating owner. Several roles may sit with one person, but none should be left implicit.
Internal AI champions can help colleagues and reveal workflow problems. Microsoft's champion guidance distinguishes that role from security approvals, IT support, and business accountability.
Before sign-off, request:
- Agreed code, configuration, documentation, and access transfer.
- Monitoring, escalation, recovery, and stop procedures.
- Staff training and a demonstration using a synthetic failure.
- Known limitations, ongoing costs, and support terms.
Have the designated operator investigate that failure using the handover materials. This reveals gaps a document review can miss.
The hiring question that matters
Ask: “What will our team receive, and what evidence decides whether we expand, revise, or stop?”
A credible proposal connects discovery, pilot, production readiness, and handover to those decisions. One experienced engineer may deliver a bounded project with your domain and IT owners. Broader transformation needs the relevant team and authority.
If you need help scoping the engagement, our AI strategy and roadmapping service starts with workflow priorities and feasibility. Bring one real process, representative inputs, and its owner.
