Contents
Ask an operations director whether they'd let AI answer their email and you'll get the same answer every time: some of it, obviously, and absolutely not the rest.
They're right, and the interesting engineering question is where the line goes. Draw it too conservatively and you've built a very expensive way to produce drafts nobody trusts. Draw it too aggressively and you find out, from a customer, that an AI told them something untrue about their claim.
Our email agent draws the line with a confidence score and three tiers. Here's exactly how.
Three tiers
Every inbound email produces a recommendation, and there are only three:
Auto-approve. The agent acts. The reply sends, the task updates, the documents attach. A record of what it did lands on the case.
Review recommended. The agent has done the work and is confident enough to show it, but a human clicks send. In practice this is the tier most operations live in for the first few months, and a lot of them stay there because a handler approving a good draft in four seconds is already the win.
Review required. Nothing happens without a person. The draft exists, the reasoning is attached, and it waits.
The default thresholds are 0.85 and 0.70. Above 0.85, auto-approve. Between 0.70 and 0.85, review recommended. Below 0.70, review required. Those are configurable per installation, and the system refuses a configuration where the auto-approve threshold sits at or below the review threshold, because we've watched people try.
The eleven factors
The score is a weighted average of eleven signals in four groups. Not all of them are about the email.
What kind of email is this? Scenario confidence, at 15%, is how sure the classifier is that it has understood the category at all. Scenario complexity and urgency are each 10%, and both are inverted: simpler and less urgent score higher. An urgent email isn't harder to understand, but it's a worse candidate for an automated reply, because urgency is where the cost of being wrong goes up.
What did we get out of it? Extraction confidence and information completeness are 15% each. Ambiguity is 10%, inverted. If the model pulled six clean fields out of the message, that's a good sign. If it pulled six fields and flagged three of them as ambiguous, the score drops and the reasoning names which ones.
How good is the response? Relevance and completeness at 10% each, template match at 5%. Whether the draft actually addresses what was asked, and whether an existing approved template fits.
Who are we talking to? Customer history and prior interactions, 5% each. Someone we've exchanged five messages with this month is better understood than someone writing for the first time. Escalation history pulls the score down: a customer who has escalated before is a customer whose next email deserves a human.
Every score comes with its reasoning as plain text, because a number by itself tells a reviewing handler nothing they can use. "Extracted 7 fields (confidence 91%). Ambiguous fields: policy_number. Information completeness: 86%." A handler can read that in a second and know where to look.
The four hard overrides
Here's the part that matters more than the weights, and it's the part most vendors leave out of the demo.
Four conditions force review required regardless of what the score says:
- The scenario is classified as one that requires human review
- The category is a complaint or an escalation
- Sentiment is negative
- Urgency is critical
These aren't thresholds. They're gates, and the score cannot open them. An impeccably clear, perfectly extracted, high-confidence complaint scores beautifully on all eleven factors and still goes to a person, because "the AI understood the complaint very well" is not a defence anyone wants to run at the FOS.
We think of this as the difference between a system that's confident and a system that's safe. Confidence is about the model's grasp of the email. Safety is about what happens if the grasp is wrong, and that depends on the subject matter, not the score.
Complaints are the clearest case. Under DISP, a complaint starts a clock and a set of obligations the moment it arrives. An automated reply that fails to recognise something as a complaint doesn't just answer badly, it starts the clock late. So complaints never auto-approve, in any configuration, at any score.
What the agent actually does
Between reading the email and producing a recommendation, the pipeline runs a fixed sequence. Classify the scenario. Enrich it with context from the case file, so the model knows what's already happened. Decide on an action. Read attachments, in two passes, because a first pass to work out what a document is makes the second pass at what it says far more reliable. Score the result. Then execute, queue, or hold.
The actions available are deliberately small: draft a reply, read documents, attach documents, update a task, create a subtask, create a new task, escalate, or do nothing. "Do nothing" is a real decision the agent makes, and it's more common than people expect. Plenty of inbound email needs to be filed and understood without anybody replying to it.
The number that decides your rollout
The threshold configuration is the most consequential setting in the whole system, and it's worth being deliberate rather than accepting a default.
Start with auto-approve effectively switched off, by setting the threshold to something nothing will reach. Everything routes to a handler with a draft attached. Run it for a month. You now have a body of evidence about what the agent recommended and what your team did with it, on your email, in your business.
Then look at the disagreements. If the drafts your handlers approve unchanged cluster in one or two scenario categories, those are your candidates for automation. If your handlers rewrite most drafts of a particular kind, the fix is a better template rather than a lower threshold.
We'd rather a customer spend the first month generating that evidence than accept our defaults, which came out of our data and describe our assumptions.
The honest summary
If someone tells you their AI answers email autonomously, ask them three questions. What happens with a complaint. What happens when sentiment is negative. What the reviewing handler sees when they open the queue.
Automation that can't answer those isn't automation. It's an unreviewed liability with good marketing.
Further reading:
- Why our AI never agrees a total-loss value: the same argument where money is involved
- Claims outcome prediction: recommending without deciding, applied to claims
- Human in the loop: the approval design across every channel
- Email automation: what the email agent does end to end