When is building in-house the right choice?
If your organisation has a standing AI engineering team, an eighteen-month horizon and a highly specific need, building remains defensible. We would rather say so.
Comparison
Build your governance stack, or deploy one?
Both approaches are legitimate. They do not solve the same problem of calendar, maintenance and control. This page helps you identify which one is yours.
What internal development brings
-
Full control of code and architecture
Every technical choice stays in-house: languages, hosting, security policy, release cycles.
-
Exact fit to existing systems
Connectors and workflows can match the detail of applications already in production.
-
No vendor dependency
No licence contract, no third-party roadmap, no product end-of-support risk.
-
Skill acquired and kept in-house
The team that builds the stack remains the reference. The knowledge stays in the organisation.
What NEXA does differently
-
A first process in production in weeks, not quarters
The observed time to a first industrialised process is on the order of 6 to 12 weeks, versus 9 to 18 months for a stack built from scratch.
-
Stack maintenance is included
LLM engines, patches, connector changes: absorbed by the platform, not by a permanent in-house team.
-
Deliverable code under contract
Code control is not total. Reversibility and code delivery can be provided for contractually.
-
Three-year cost falling per process
Each extra process reuses the stack already deployed. Unit cost falls; in-house, stack maintenance remains a fixed load.
What you have to build
-
Multi-engine orchestration
Chain several LLMs and tools in one execution, with a technical receipt per run.
-
Per-execution traceability
Link each assertion to its sources and a confidence score.
-
Evidence panel
Present sources, confidence and validation status in an exportable view.
-
Validation chain
Draft → Reviewed → Approved, with identities recorded at each stage.
-
Audit log
Keep who saw, changed, validated, and when.
-
Execution replay
Replay a past run (versions, parameters, inputs) under the same conditions.
-
Personal-data detection
Identify and handle PII before it enters a model.
-
Data residency
Control where data and logs are stored (EU, private, on-premise).
-
Versioning
Freeze prompts, policies and models used for a given run.
-
Cost steering
Track cost per execution, not only per seat.
-
Identity management
SSO, roles and entitlements on the validation chain.
-
Capitalisation of validated patterns
Feed back what was approved to lower the cost of later runs.
Each of these elements is feasible. The issue is cumulative lead time and maintenance over the years.
Two models, side by side
| Criterion | Internal development | NEXA |
|---|---|---|
| Time to first process in production | 9 to 18 months | 6 to 12 weeks |
| Maintenance | Permanent dedicated team | Included |
| LLM engine evolution | On you | Absorbed by the platform |
| Code control | Total | Partial, code deliverable under contract |
| Upfront cost | High, spread over time | Moderate |
| Cost at three years | Rising | Falling per process |
-
Time to first process in production
- Internal development
- 9 to 18 months
- NEXA
- 6 to 12 weeks
-
Maintenance
- Internal development
- Permanent dedicated team
- NEXA
- Included
-
LLM engine evolution
- Internal development
- On you
- NEXA
- Absorbed by the platform
-
Code control
- Internal development
- Total
- NEXA
- Partial, code deliverable under contract
-
Upfront cost
- Internal development
- High, spread over time
- NEXA
- Moderate
-
Cost at three years
- Internal development
- Rising
- NEXA
- Falling per process
How to choose
What do you have to build for an AI governance stack?
Multi-engine orchestration, per-execution traceability, evidence panel, validation chain, audit log, execution replay, personal-data detection, data residency, versioning, cost steering, identity management, capitalisation of validated patterns. Each of these is feasible. The issue is cumulative lead time and maintenance over the years.
How long until a first process in production?
For an in-house stack, 9 to 18 months is a common order of magnitude before the first process in production. On NEXA, the first industrialised process typically sits between 6 and 12 weeks. These are observed durations, not a universal contractual commitment.
If your organisation has a standing AI engineering team, an eighteen-month horizon and a highly specific need, building remains defensible. We would rather say so.
See how a process becomes a defensible dossier
A walkthrough on a process your teams already run. Then a framing session to size the first deployment.