Professional services
Cited knowledge assistant
Read-only retrieval across engagement archives and methodology, answering with citations and declining outside its evidence. Deployed in weeks, with no write path to defend.
AI Engineering · 04
Autonomy is a cost, not a feature.
Most organisations do not need an autonomous agent. They need a well-scoped assistant with reliable retrieval, or a copilot embedded where the work already happens. Every increment of autonomy adds oversight machinery, evaluation surface, and incident exposure — so it should be bought deliberately, step by step, against a stated benefit.
Four levels, four different engagements
- Scripted bot
- Deterministic flows over a fixed intent set. No model in the decision path. Ideal for status lookups, form intake, and routing where the branch logic is genuinely known.
- Retrieval assistant
- Answers questions over your corpus with citations and refuses outside its evidence. Read-only by construction, which is why it is usually the correct first deployment.
- Copilot
- Embedded in the operator's existing tool, proposing drafts and actions a person accepts. Autonomy stays at zero while the throughput gain is immediate.
- Agent
- Executes multi-step work against real systems under policy. Justified when volume is high, the process is well-defined, and exceptions can be routed rather than guessed.
The autonomy–oversight matrix
Oversight requirement tracks autonomy almost linearly, and blast radius grows faster than either. The useful reading of this chart is that nothing sits in the lower-right quadrant: there is no configuration in which a system acts independently against production systems while requiring little supervision.
Figure 1
Required oversight against granted autonomy
Bubble size indicates blast radius — the scope of what a single unattended failure can affect. Systems are positioned by the level of independent action they are granted over systems of record.
Illustrative reference data — engagement figures vary by environment.
Build effort is only half the commitment
Selection conversations tend to focus on delivery. The ongoing operating load — evaluation suites, drift review, exception queues, incident response, and access recertification — scales alongside it and is the cost most frequently underestimated at approval.
Figure 2
Indexed build and ongoing operating effort by system type
Indexed to an autonomous agent at 100. Operating effort is annualised and assumes the system stays in production with an owner and an evaluation cadence.
Illustrative reference data — engagement figures vary by environment.
Use cases
Insurance
Adjuster copilot
Drafts the claim summary, the coverage position, and the correspondence for an adjuster to accept or amend. Decision authority stays with the adjuster; the assembly work does not.
Shared services
Supervised back-office agent
Executes a defined, high-volume process end to end with approval gates on every mutation and an exception queue for anything outside tolerance.