Choose the right tool
Not everything needs a model.
Most automation projects fail because they reach for judgement where a rule would do — or for a rule where the input is genuinely ambiguous. We split the workflow before we build it.
When the rule is knowable
Repetitive, high-volume work with a defined outcome. Cheaper, faster and fully auditable — the default choice.
- Rules, thresholds and state machines
- Predictable cost per run
- Replayable, diff-able, testable
- No prompt to regress
When context has to be read
Unstructured input, fuzzy classification, drafting and summarising — wrapped in guardrails that keep the failure mode boring.
- Context from your own data, not the open web
- Scoped tools with permission boundaries
- Confidence thresholds and human escalation
- Evaluated on your cases before launch
Where it pays back first
Four steps, starting with an audit
- 01
Workflow audit
Two weeks mapping volume, handling time and error cost per step. You get a ranked list of what to automate and what to leave alone.
- 02
Scope & guardrails
Before any build: what the agent may touch, what needs approval, what happens on low confidence, and how we’ll measure correctness.
- 03
Build & validate
Shadow mode first: the flow runs alongside the humans until it agrees with them often enough. Then a staged cutover by segment.
- 04
Monitor & improve
Dashboards for accuracy, escalation rate and cost per run — plus a monthly review where we tighten the cases that still need a human.
Safeguards
An agent is only as good as its limits.
These are not optional extras — they’re in the first sprint, and they’re what makes the system reviewable by risk and compliance.
- Approval gatesAny action above a value or risk threshold waits for a named human.
- Full audit loggingInput, retrieved context, decision, tool call and outcome — queryable per case.
- Fallbacks and retriesModel unavailable, tool failing, output malformed — each has a defined path.
- Human handoffEscalation carries the full context, so the person doesn’t start from zero.
- Evaluations in CIA graded case set runs on every prompt or model change, so regressions are caught before release.