“Your plan has a $2,500 individual deductible, after which coinsurance applies at 20% until you reach your out-of-pocket maximum.” Every word of that is correct, and almost nobody reads it and knows what their MRI will cost.
Benefits language is written to be defensible. It has to be. But a member asking a question in an app at nine in the evening is not looking for the defensible version, and the gap between the two is where health plans lose trust.
The two failure modes
There are only two ways to get this wrong, and they pull in opposite directions.
- Quoting the document. Accurate, useless, and the member calls anyway. All you have done is move the reading to a smaller screen.
- Paraphrasing freely. Readable, friendly, and occasionally wrong about money — which in this domain is not a minor error, it is a compliance event.
The way through is not a better writing style. It is a hard rule about where numbers come from.
Every number is a citation
In our health plan deployments the assistant is not permitted to state a benefit amount that it did not read from the member’s own eligibility and accumulator data. Not from the plan document, not from a summary, and never from the model’s own sense of what a typical deductible looks like.
That means the read path — eligibility, plan design, accumulators year to date, network status of the provider in question — has to be live before the assistant is allowed to say anything about cost at all. Where the read fails, the assistant says it cannot see the current figures and offers a person. It does not estimate.
A member would rather be told “I can’t see that right now” than be told a number that turns out to be wrong.
Then translate, in the member’s terms
Once the numbers are grounded, the language can relax. “You’ve paid $1,840 of your $2,500 deductible this year. Until you reach it, you pay the full negotiated rate for most services — after that, you pay 20% and the plan pays the rest.”
Three things make this land. It uses the member’s actual position rather than the plan’s abstract design. It says what happens next rather than naming the mechanism. And it does not use “coinsurance” without immediately saying what coinsurance does.
Say what you cannot know
Prior authorisation status, whether a particular CPT code will be covered for a particular diagnosis, and anything involving clinical judgement are not eligibility questions dressed up. The assistant should recognise them and route them, rather than answering the adjacent question it can answer and leaving the member to think they have their answer.
What changed when we measured it
The deployments that hold to the citation rule see fewer contacts, but that is not the number we watch. We watch repeat contacts on the same question within seven days. When a benefits answer is grounded and translated, that number falls; when it is merely readable, it does not move at all — the member comes back because they did not quite believe it.
