What does AI email triage do for a small business?
It gets the right message in front of the right person, so nobody has to read the whole shared inbox every morning. In a small business, the "info@" or "contact@" inbox mixes everything: a prospect asking for a price, a customer whose machine has stopped, a supplier invoice, three sales pitches and a fake delivery notice. Triage answers no one. It sorts, it flags, and it sets up the next step.
Four uses come up most often:
- Quote requests: spotted and passed to whoever prices the job, with the key details already pulled out. This is where our page on handling quote requests with an AI agent starts.
- Customer support: kept apart from sales, sorted by product or by severity (annoyance, broken with a workaround, fully blocked).
- Spam and cold pitches: set aside, never deleted automatically.
- Urgent issues: a blocked customer or a close deadline, flagged to someone within the hour instead of read the next day.
Your email client's rules already do part of this without AI: a word in the subject line, a sender address. Their limit is well known. A rule on "quote" misses the customer who writes "could you price this for me". If your emails already arrive well labelled (a web form, a dedicated address), a rule is enough and AI adds nothing.
What is OpenAI's Decisions API?
It is an OpenAI service, in public beta since October 6, 2026, that writes nothing: it answers closed questions. You send it an email (or an image, such as a product photo) with one or more questions, and it sends back a typed answer. The model is called gpt-6-luna, and it is the only one available. OpenAI says answers come back up to ten times faster than through its usual API with the same model.
| Question type | What it returns | Example for a shared inbox |
|---|---|---|
| Yes or no ("predicate") | The probability, from 0 to 1, that the statement is true | "Is this email a complaint?": 0.92 |
| Choice ("choice") | One value from your list, with a confidence level | Quote, support, supplier invoice, spam, other |
| Score ("score") | A position on a scale you define | Severity of a fault: annoyance, workaround available, fully blocked |
Several questions can be asked about the same email in a single request. The service can also refuse to answer a question, so your setup needs to say where that email goes (to a person, of course).
How much does AI email triage cost with the Decisions API?
OpenAI only bills what the model reads: $0.10 per million input tokens, with no output or caching charges. A token is a piece of a word, and the email, your instructions and your list of categories all count as input. The table below is a worked example, not a measurement: it assumes 500 tokens read per email, instructions included, and 30 days a month.
| Emails per day | Tokens read per month | Monthly cost (standard rate) | With regional processing (+10%) |
|---|---|---|---|
| 50 | 750,000 | about $0.08 | about $0.08 |
| 200 | 3 million | about $0.30 | about $0.33 |
| 1,000 | 15 million | about $1.50 | about $1.65 |
Two surcharges apply: keeping processing in a specific region (Europe or the United States) costs 10% more, and a very long input switches to the long-context rate, which is double. An ordinary email never comes close to the second. For comparison, Simon Willison, the developer who released a tool to use it on day one, notes that a competing service, Jev, charges $0.042 per million tokens.
The real cost is elsewhere: connecting the inbox, writing categories that don't overlap, testing on your own emails, deciding what happens when the model hesitates. It is the same point we make in our article on what stays expensive as AI prices fall: you pay for a well-scoped task, not for tokens.
What the top-ranking pages leave out
Method: on October 8, 2026, we took the Google results for "ai email triage small business" (United States) and "ai email sorting" (United Kingdom) and read the 10 guides and product pages among them in full.
None of the 10 gives a cost per email; the ones that talk about money quote tool subscriptions. None mentions a confidence threshold, the level of certainty below which the triage should hand over to a person. That is exactly what a business owner needs to know: how much, and what slips through when the sorting is wrong. The next two sections cover it.
What happens when the triage gets it wrong?
It will get things wrong, and you need to decide in advance what that costs. A spam email filed as a quote request wastes two minutes. A complaint filed as spam can lose you a customer. That is why the Decisions API returns a probability rather than a verdict: you set the threshold.
OpenAI's own documentation says so: set thresholds from labelled examples taken from your business, based on what a false positive and a false negative cost you. In practice:
- take a few hundred real emails from recent months, already sorted by hand;
- run them through the triage, compare, and look hardest at mistakes in the sensitive categories;
- set one threshold per category: the costlier a category is to miss, the stricter its threshold, and anything below it goes to a person;
- keep a visible review queue and check it daily for the first few weeks.
This is how we build request triage: every classification carries a confidence level, and anything uncertain goes back to a person. The questions to ask before connecting any tool to a mailbox (access, data, permissions) are in our article on four questions to ask before connecting an AI agent to your tools.
Limits to know before you start
It is a beta. OpenAI expects general availability "in the coming weeks". Response formats, prices or terms may change before then, so don't build a critical process on it without a way back.
Quality on your emails is unmeasured. OpenAI's published examples are short and generic. We found no public measurement on real small-business inboxes, in English or in other languages, and we have not been able to test the service ourselves yet. Your own test, on your own emails, is the only measurement that counts.
Your emails leave your mail system. Every email you triage is sent to OpenAI. The service offers data processing in Europe (EEA and Switzerland) or the United States, and a zero data retention option for eligible customers, subject to conditions and signed agreements. Check those terms before connecting an inbox that receives customer data, and only send to triage what actually needs it.
Two categories are never sorted automatically. A request about personal data (access, correction, deletion) must be answered within one month if you serve customers in the EU or the UK (GDPR, Article 12), and it commits the business: it goes to a person. So does a complaint, because a botched reply costs more than the time saved. Triage can spot and flag them; it never files them in a folder nobody opens.
For an overview of what we automate in a small business, and what we leave to people, see our page on custom automation.
FAQ
How much does AI email triage cost?
With OpenAI's Decisions API, the model costs $0.10 per million tokens read. For 200 emails a day at about 500 tokens each, instructions included, a worked example comes to about $0.30 a month. Setting it up (connecting the inbox, defining categories, testing on your emails) costs far more than the model.
Aren't my email rules enough?
Often, yes. When emails already arrive labelled (a form, a dedicated address, known senders), a filter rule is simpler and never guesses. AI helps when the same kind of request is phrased a hundred different ways.
Where do my emails go?
To OpenAI. The service offers processing in Europe or the United States at a 10% premium, and a zero data retention option for eligible customers. The conditions are on OpenAI's "data controls" page; check them before connecting an inbox that holds customer data.
Can AI answer my emails for me?
The Decisions API writes nothing: it sorts. Other tools draft replies; we recommend that a person reads every reply before it goes out, and always for a complaint or a personal data request.
Sources: OpenAI, "Decisions" guide (developers.openai.com, accessed October 8, 2026: question types, beta status, pricing, regional processing, zero data retention) · OpenAI API pricing (developers.openai.com/api/docs/pricing, accessed October 8, 2026: 10% regional processing uplift, gpt-6-luna long-context rate) · Public beta announcement on the OpenAI developer forum, October 6, 2026 (community.openai.com) · Simon Willison, "Release: llm-openai-decisions 0.1a0", October 6, 2026 (simonwillison.net) · Regulation (EU) 2016/679 (GDPR), Article 12(3) (eur-lex.europa.eu) · Reading of 10 Google results pages (US and UK), October 8, 2026. Costs are worked examples, not measurements. No client is named.