Evaluating AI Email Triage: Can LLMs Safely Draft Client Replies?

Email triage is the sorting of incoming messages by urgency, intent, and required action. A draft-reply system adds an LLM that proposes text for a human to approve or send. The safety question is not whether the model writes fluent English; it is whether it avoids unauthorized commitments, misstates pricing or availability, leaks private data, and misses escalation triggers. For a small business, one bad reply can cost a job, a deposit, or a client relationship. Homeowners evaluating contractor-facing tools face the same issue: a fast reply is worthless if it confirms a date or price that no one can honor. Define your critical-error categories before you sign a contract, because those categories determine review depth, vendor requirements, and whether the system is worth paying for.

Use a risk-adjusted monthly value formula, not a simple time-saved estimate:

Net monthly value = H × [A × (T_manual − T_review) + (1−A) × (T_manual − T_review − T_fix)] ÷ 60 × R − C_tool − H × P_crit × C_error

Variable	Meaning	Example baseline
H	Client emails per month	480
T_manual	Minutes to draft manually	6
T_review	Minutes to review/send AI draft	2
T_fix	Extra minutes to correct rejected draft	4
A	Draft acceptance rate	75%
R	Loaded admin hourly rate	$32
C_tool	Monthly software and API cost	$90
P_crit	Critical error rate per email	0.25%
C_error	Expected cost per critical error	$300

Track acceptance, critical-error rate, escalation recall, review minutes, and privacy incidents. A draft is accepted only if it needs no substantive change to price, scope, timing, or legal terms.

Assume a six-person landscaping company receives 480 client emails per month. Manual drafting takes 6 minutes; reviewing an AI draft takes 2 minutes. The model is accepted on 75% of emails. Rejected drafts need 4 extra minutes, so they save 6 − 2 − 4 = 0 minutes. Accepted emails save 4 minutes each: 480 × 0.75 × 4 = 1,440 minutes, or 24 hours. At $32/hour, gross value is $768. Subtract $90 tool cost and expected critical-error cost: 480 × 0.0025 × $300 = $360. Net value is $768 − $90 − $360 = $318/month. Sensitivity matters: at a 0.5% critical-error rate, expected error cost rises to $720 and net value becomes −$42. At 0.1%, net value is $618. The decision turns on error cost and review discipline, not on the model’s writing style.

Set go/no-go thresholds before the pilot. Use a comparison matrix:

Metric	Green	Yellow	Red
Draft acceptance	≥75%	60–74%	<60%
Critical-error rate	≤0.25%	0.26–0.75%	>0.75%
Escalation recall	≥95%	90–94%	<90%
Privacy incidents	0	0 with near miss	Any incident
Net monthly value	>$500	$0–$500	<$0
Review time per draft	≤2 min	2–4 min	>4 min

Green means expand only with human approval. Yellow means restrict to low-risk categories such as scheduling acknowledgments and FAQ replies. Red means stop or redesign. Never auto-send pricing, scope changes, legal terms, refunds, or availability promises until the system has passed a shadow-mode test and a supervised pilot.

Five expensive mistakes recur. First, teams measure fluency instead of factual accuracy, so polished drafts hide wrong prices or dates. Second, the model quotes availability without a live calendar or inventory source. Third, there is no escalation taxonomy, so complaints, legal threats, safety issues, and refund requests get routine replies. Fourth, reviewers rubber-stamp drafts because the text looks confident; sample rejected and accepted drafts weekly. Fifth, privacy controls are ignored: client addresses, payment details, and contract terms may be retained or used for vendor training. Each mistake shifts cost from labor to liability. A single wrong scope commitment can erase months of savings, and a privacy incident can damage trust faster than faster replies can rebuild it.

1. Define critical errors and escalation rules in writing: pricing, scope, timing, safety, legal, refunds, and privacy. 2. Collect 150–300 historical emails with the correct outcome and label them by category. 3. Build a test set and score accuracy, escalation recall, and critical-error rate before any live use. 4. Run shadow mode for two weeks: the LLM drafts, but staff send manually and record edits. 5. Start a supervised pilot on two low-risk categories with 100% human approval. 6. Review metrics weekly against the matrix; expand only when green for two consecutive weeks. 7. Add source-of-truth integrations for calendar, pricing, and CRM before allowing any auto-send. 8. Document review ownership, data retention, and incident response.

• Critical-error categories and escalation rules are written and approved.
• Test set includes at least 150 real emails with known correct outcomes.
• Draft acceptance, critical-error rate, and escalation recall meet green thresholds.
• Pricing, availability, and scope answers come from approved sources.
• Human approval is mandatory for high-risk categories.
• Privacy, retention, and vendor-training settings are reviewed.
• Review time per draft is measured, not assumed.
• Net monthly value remains positive after expected error cost.
• A named owner reviews metrics weekly and can stop the rollout.