Customer operations
Three tiers per ticket
Fast-tier intent tagging on arrival, balanced-tier summarisation and drafting, reasoning-tier adjudication only on contested escalations — with a fallback tier declared for each step.
AI Engineering · 03
There is no best model. There is a correct tier for each step.
Model families differ along capability, latency, context window, modality, and cost by more than an order of magnitude. Treating a workflow as one model call forfeits that spread. Decomposing it and routing each step to the cheapest tier that clears its quality bar is the single largest lever on both economics and responsiveness.
The five tiers we route between
- Reasoning
- Extended deliberation for multi-constraint analysis, contested decisions, and planning. Highest quality, highest latency, highest cost — reserved for the few steps that need it.
- Balanced
- The workhorse tier. Summarisation, extraction, drafting, and most tool-calling steps run here at a fraction of reasoning-tier cost.
- Fast
- Sub-second classification, routing, tagging, and guardrail checks. Deployed wherever a human is waiting on a response in real time.
- Long-context
- Whole-corpus review — contract sets, case files, log windows — where retrieval alone loses cross-document relationships.
- Multimodal
- Scanned documents, photographs, diagrams, and audio. Used at the ingest boundary, then handed to a text tier for downstream reasoning.
- Small & self-hosted
- Fine-tuned open models inside your own network for residency-constrained or extremely high-volume steps with a narrow task definition.
Capability against cost
Capability rises quickly at the low end and then flattens while cost continues to climb. Most workflow steps sit comfortably on the flat portion of the curve; sending all of them to the reasoning tier buys little accuracy and multiplies spend.
Figure 1
Relative cost against composite capability by tier
Cost is indexed to the reasoning tier at 100. Bubble size reflects usable context window. Positions are directional and shift with each provider release.
Illustrative reference data — engagement figures vary by environment.
Latency budgets decide interactivity
Latency is a product constraint before it is an engineering one. A reviewer waiting on a screen tolerates several hundred milliseconds; a nightly batch tolerates minutes. Routing respects the declared budget first and optimises quality within it.
Figure 2
Typical end-to-end response time by tier
Median time to a complete response for a representative workload. Streaming changes perceived latency but not the budget a downstream step must plan against.
Illustrative reference data — engagement figures vary by environment.
Trade-off profile
Comparing tiers on a single score hides the trade. Plotted across six axes, the shape of each tier makes the routing decision obvious: the fast tier wins decisively on latency and cost efficiency and loses on structured reasoning, which is precisely why workflows are decomposed rather than assigned wholesale.
Figure 3
Tier profile across six evaluation axes
Normalised 0-100 composite scores. Higher is better on every axis, including latency and cost efficiency, which are inverted so that the outer edge is always preferable.
Illustrative reference data — engagement figures vary by environment.
Use cases
Legal & contracts
Whole-portfolio clause review
Long-context review across an entire agreement set to catch cross-referenced obligations, then balanced-tier drafting of the redline summary for counsel.
Field operations
Image-first intake
Multimodal capture of photographed forms and equipment plates at the edge of the workflow, normalised to text records that cheaper tiers then process at volume.