Sensitive topics by industry: what your AI should never answer
The six categories that apply to every business, the additions for twelve kinds of practice, how to phrase a handoff, and why the emergency check has to sit outside the score rather than inside it.
Sohail Safdar · · 6 min read
A veterinary practice switches its assistant on. Within a week somebody writes at half past nine in the evening: "my dog ate a grape, is that bad?"
Every instinct in a helpful system says answer. The question is specific, there is a correct answer, and the customer is clearly worried. It is also exactly the question the assistant must not touch, because the right answer depends on the dog's weight and how many, and because being wrong costs an animal rather than a booking.
The work is deciding which questions those are before somebody asks one, rather than reading the transcript afterwards.
The six that apply to everyone
- Professional judgement. Anything a licensed person is supposed to decide.
- Safety and emergencies. Anything where a delay could cause harm.
- Other people. Any question about somebody other than the person writing, which is a privacy answer and not a service one.
- Money in dispute. Refunds, chargebacks, billing errors.
- Complaints and legal threats. Both get worse when answered quickly and informally.
- Prices that depend on seeing the job. A quoted number becomes a promise.
Additions by industry
| Industry | Never answer |
|---|---|
| Dental and medical | Symptoms, diagnoses, medication, results, whether something is serious |
| Med spa | Suitability, contraindications, pregnancy, expected results |
| Chiropractic and physical therapy | Pain assessment, exercises, whether to seek other care |
| Optometry | Vision changes, eye injuries, prescription changes |
| Veterinary | Symptoms, dosing, whether a pet needs urgent care |
| Legal | Merits, outcomes, how a deadline applies to this person, what to do next |
| Accounting | Tax advice, a specific return, what a notice means |
| Insurance | Premiums, coverage, whether a claim will pay |
| Real estate | Property values, offers, neighbourhood demographics |
| Home services | Repair prices, the safety of a gas or electrical problem |
| Pest control | Identifying a pest, product safety for a particular household |
| Moving | Totals, binding estimates, date guarantees |
Read your own row and then add to it, because the twelve above are the predictable ones. The additions that matter are the questions your staff already refuse to answer on the phone, which nobody has ever written down because everybody knows them.
Phrasing the handoff
Three parts, in this order: what the assistant cannot do, who will help, and when.
"I cannot advise on that, but one of our providers can. Someone will text you this morning."
For safety topics, add the instruction you approved, in the words you approved: leave the building, call 911, go to the emergency hospital. This is the one place where a general-purpose phrasing is not good enough, because the correct instruction differs by trade and getting it wrong is the whole risk.
What not to do: apologise at length, explain the policy, or offer a consolation answer. "I cannot give medical advice, but generally speaking" is the failure this entire list exists to prevent, and it is the most natural sentence in the world for a helpful system to produce.
Categories first, keywords as a backstop
Customers phrase things unpredictably, and frightened customers phrase things worst of all. Nobody writes "I have a symptom". They write "is this normal". A refusal built only on a keyword list will pass that straight through to an answer.
So the check is on the topic, with your configured keywords underneath it for the terms peculiar to your trade. Test it with the phrasings your customers actually use, taken from your own inbox rather than invented, which also tells you which additions your row was missing.
Why the emergency check sits outside the score
This is the design decision most worth copying, whatever you build with.
Frontiva scores an agent's readiness across five scenarios that measure how it answers, and the score has a passing mark. Two further checks are run and deliberately kept out of that average: whether the agent can get an emergency to a human at all, and whether it escalates to a number somebody actually set.
The reason is arithmetic. An average can be outvoted. A practice can fail one scenario and still clear a passing score, and folding these two in would have made "an emergency has to reach somebody" a fraction of a mean, which is a strange thing to let five other results outvote. So they are a condition rather than a contribution: fail either and autonomous mode is refused regardless of the rest.
The second of the two is the one people do not anticipate. An agent that escalates correctly to a number nobody configured has done everything right and woken nobody, which is indistinguishable from doing nothing. An on-call number that is missing, and a messaging connection that does not exist to send the alert over, are both treated as failures of the emergency path rather than as setup warnings.
One more rule in the same spirit: a readiness run recorded before these two checks existed proves neither, so it is refused as well, and the message says to run the check again rather than implying something is broken. A gate should report the present rather than the past.
Reading the handoffs after the first month
The list you write before going live is a prediction. The handoff log is the correction.
Take a month of handed-off conversations and sort them into three piles. Correctly refused, which is the pile you want to be largest. Should have been answered, which usually means a fact is missing rather than that the rule is wrong: somebody asked a price you never loaded, and the safest available behaviour was to fetch a person. Should have been refused and was not, which is the pile that needs action this week.
The ratio tells you which way to move. A business handing off a third of everything has a knowledge gap, not an over-cautious assistant, and the fix is loading facts rather than relaxing rules. A business handing off almost nothing in a trade with a long row in the table above has probably not tested with the phrasings its customers use.
What not to do with this review: relax a safety rule because it fired on something harmless. The cost of the false positive is a customer waiting for a person who was going to be involved anyway. The cost of the false negative is the thing the rule exists for, and the two are not comparable just because one of them happens more often.
In Frontiva
The sensitive-domain check runs before any answer is looked up, which is the ordering that matters: a question that should not be answered is never improved by first finding a good answer to it. Escalation keywords you configure add your trade's own vocabulary on top.
A sensitive-domain question is one of the scored readiness scenarios, so an agent that would answer one cannot quietly reach go-live, and the two emergency checks above sit above the score as conditions. Approval mode is untouched by all of it: while a person is reviewing every reply, nothing here blocks you from starting. See what an AI receptionist should refuse to answer and the AI agents page.