Claude

TEMRIK editorial assessment · Sources checked 25 September 2026

Our recommendation

Claude deserves a place on an SME shortlist when the work involves reading, drafting or developing software. The useful question is whether it delivers an acceptable result with less total effort—including checking and correcting it.

For a quantity surveyor, a sensible first use is organising a specification into a review checklist. A poor first use is letting it certify quantities or approve a variation without checking the drawings, contract and arithmetic.

Key functions and usefulness matrix

These are editorial suitability ratings, not measured accuracy scores. Strong fit means worth piloting with human review; conditional fit requires additional controls; poor fit means do not use it as the sole decision-maker.

Choosing a cheaper model

API costs depend on the model, input and output tokens, and optional features. Providers measure processed content in tokens; the number varies with the content and model. Anthropic lists different model tiers, caching rates and batch rates. Compare the live pricing table before committing.

Our recommendation: test a cheaper model such as Haiku for bounded classification or extraction; escalate difficult exceptions to a more capable model. Keep a fixed set of representative cases and compare missed fields, unsupported claims, latency and reviewer corrections. A lower token bill is not a saving if it doubles checking time.

Budget total cost as model usage + integration operation + human review + rework. Use shorter relevant context, sensible output limits and eligible caching or batch processing. Model changes require regression checks; do not silently swap engines on high-consequence work.

Alternatives worth comparing

For dependable totals or fixed approvals, the best alternative may be ordinary software: a spreadsheet, rules engine or specialist system. Use an LLM only where language interpretation adds value.

A useful first pilot

1. Choose one recurring document task and prepare authorised, representative examples with known answers.
2. Ask Claude for a structured draft with supporting passages and an explicit “not found” option. Never reward it for filling every field.
3. Compare the result with a human baseline and at least one alternative model. Record errors, review time and cost per accepted output.
4. Approve wider use only after the workflow meets your quality threshold. Keep a manual fallback and repeat the check when models or prompts change.

Watch for confident omissions, invented references and instructions embedded in uploaded documents. Limit tool permissions and require approval before sending messages, changing records or committing money.