Free Inbox Triage + Reply Drafter with Cloudflare Workers AI
Most "free AI for your business" guides tell you to put a chatbot on your website. That's backwards. Here's what actually holds up: paste one Worker script into Cloudflare, deploy in under ten minutes, and you get a private, always-on endpoint that sorts inbound customer email into urgent / sales / support / billing / spam and drafts a first reply for the ones that need one. You review before you send, so a bad draft costs you a re-write β not a customer.
The limit, stated up front: the free tier is capped at 10,000 neurons per day 2026-08-20 β roughly 900β1,200 short drafted replies on the 8B model, or ~110 on the 70B β and Cloudflare retires model names on a schedule, so the exact handles below are date-stamped and should be re-checked before you deploy.
Pick your chore β get the right free model
What this replaces
Same job, three paid ways vs. the free tier. Figures are 2026-08-20 and will drift β re-check before quoting.
| Route | Typical cost | With this workflow |
|---|---|---|
| Hosted AI chatbot (Tidio + Lyro AI) | $79β$150/mo | $0.00 |
| You, typing the same 10 replies (20/wk Γ 5 min) | ~87 hrs/yr | ~2 hrs/yr reviewing |
| OpenAI API, gpt-4o-mini ($0.15/M in Β· $0.60/M out) | ~$0.20/mo | $0.00 |
| Cloudflare paid tier (only if you exceed free) | $5/mo + $0.011/1k neurons | $0.00 under 10k/day |
Will your volume fit in 10,000 neurons/day?
Neuron rates are from the official pricing table 2026-08-20. Formula: neurons = in_tokens Γ rate_in + out_tokens Γ rate_out.
Build it in ~10 minutes
What breaks
A public bot gets unbounded traffic β strangers, spammers, the whole internet β so your 10,000 neurons die by mid-morning. And an 8B model answering strangers in public gives worse answers than a canned FAQ, publicly. The free tier's actual sweet spot is your own bounded, repetitive, high-value chores (your inbox), where you review every output before anyone else sees it. Skip this distinction and you will burn the daily quota before your first real customer writes back.
Model handles expire
Cloudflare deprecates models on a published schedule β the May 30, 2026 batch retired @cf/google/gemma-3-12b-it and @cf/meta/llama-3.1-70b-instruct, among others. A guide you read last quarter may point at a dead handle. Pin yours, re-check quarterly.
Small models truncate long input
8B models silently cut off long emails β the script slices input to 4,000 characters for a reason. Pasting a whole contract and expecting it read is how you get a reply about the first paragraph only.
Neuron math drifts
Rates change when models move tiers or get replaced. Every number on this page is tagged 2026-08-20; the durable part is the formula and the model-picking tree, not the exact per-million rates.
10 ms CPU per request (free plan)
Workers Free gives you 100,000 requests/day but only 10 ms CPU each. The AI calls themselves are I/O waits (they run on Cloudflare's GPUs), but loops and heavy string work in your code eat the budget β keep your Worker to one or two AI calls.
Not for regulated data
This is a third-party edge network. Don't send PHI, PCI card data, or anything under NDA through it. Triage the structure of messages, not secrets.
Three Things you can ship this week
@cf/openai/whisper and post audio instead of text. Whisper costs 41.14 neurons per audio minute, so a 2-minute voicemail β 82 neurons β over 100 transcriptions a day on the free tier.@cf/baai/bge-m3 (1,075 neurons per million tokens) and have the 8B model answer only from those chunks. This kills the "the model made up a refund policy" failure mode.Deploy your free triager today
Open dash.cloudflare.com, create a free account (no card), paste the script from Step 3, and hit Deploy. If you exceed 10,000 neurons/day, Cloudflare stops β it never surprise-bills.