The Register a Bot Should Use for a Difficult Question

TL;DR

• One tone across every HR question is the commonest configuration mistake. The voice that works for a holiday balance reads as insulting on a bereavement.
• There are three registers worth defining, not one: transactional for facts, careful for money and entitlement, and plain for anything involving distress, where the correct behaviour is to stop answering and fetch a person.
• Warmth is not kindness. Enthusiasm, exclamation marks and sympathetic openers on a serious question signal that the system has not understood what was asked.
• Hedging is the other failure. Four qualifications on a simple factual answer reads as evasion, and employees conclude the system knows nothing.
• Length carries tone. A six-paragraph answer to a yes or no question is its own kind of rudeness, and a one-line answer to a question about pay feels dismissive.
• Specify the register per topic in writing before launch, then test it by asking the hard questions yourself and reading the answers as the person who asked.

Great Question

An employee asked an HR chatbot what happened to their pay if they needed time off after a death in the family. The reply opened with "Great question!" and included an emoji.

Everything after that was accurate. It covered compassionate leave, it noted the discretionary element, it pointed at the right policy and it offered to connect the person to HR. The content was better than what most companies manage. None of that mattered, because the first two words had already told the employee that the thing on the other end did not understand what they were asking about, and the rest was read through that.

The configuration was not careless. Somebody had deliberately set a friendly, approachable tone, because the alternative seemed cold and the brief for the whole project had been to make HR feel more human. That instruction, applied uniformly, produced the least human possible response to the one question in the set that needed care.

Recommended Tools

Best tools for AI HR Tools

B BambooHR HR management system L Lattice Performance management D Deel Global payroll & EOR

View all recommended tools

So the problem is not warmth. The problem is one tone, applied to a question set that spans annual leave balances and bereavement. No single voice is correct across that range, and most products let you set exactly one.

The Three Registers

Stop thinking about tone as a slider between formal and friendly. Think about it as three separate modes, selected by what the question is about.

Transactional. For facts that have an answer. When is payday, how many days do I have, who approves expenses over a threshold, what is the notice period. The right voice is short, direct and unadorned. No preamble, no sympathy, no enthusiasm. The answer, then the source, then stop. This covers the majority of volume and it is where brevity reads as competence.

Careful. For questions about money, entitlement, contract terms and anything where being wrong has consequences for the person. What happens to my pay if I go on sick leave, do I accrue holiday while on parental leave, does this deduction mean what I think it means. The right voice is still direct, and it adds two things: an explicit statement of what the answer does not cover, and a named route to a person. Not hedging. Scoping.

Plain. For distress. A bereavement, a grievance, a disclosure, a mental health question, anything about a dispute or somebody's personal circumstances. The right voice is short, unembellished, and oriented entirely towards getting a person involved. It should not attempt sympathy, it should not deliver a policy summary, and it should not ask qualifying questions. One or two sentences and a route to a human being.

The third row is the one almost nobody configures, and it is the one every employee will remember.

Why Warmth Reads as an Insult

This is worth being precise about, because the instinct to be warm is a decent instinct and the advice to drop it sounds cold.

Warmth in writing works by signalling attention. When a colleague opens a reply with a sympathetic line, the warmth carries information: they read what you wrote, they understood that it was difficult, and they are responding to you specifically. That is why it helps.

An automated system cannot signal attention, because it did not attend. When it produces a sympathetic opener it is producing a template, and on an easy question nobody minds. On a hard question the employee knows perfectly well that nothing read their situation, so the sympathy is revealed as a format, and a format pretending to be sympathy is worse than no sympathy at all. It is the difference between a condolence note and a condolence-shaped mailshot.

There is a second failure underneath it. Enthusiasm implies the question was welcome and routine. "Great question!" on a bereavement query says, accurately, that this message has been classified as one of several hundred this week. Employees are very good at detecting that, and the conclusion they draw is not about tone. It is that the system did not know what they were asking, which makes them doubt the answer as well.

And warmth takes up the first line, which is the line that should carry either the answer or the route to a person.

The Four Things to Specify

Most answering products expose a tone or persona instruction. The useful move is to write that instruction per topic rather than once, and to specify four things rather than a vibe.

Opening behaviour. What the first sentence does. On transactional questions it states the answer. On careful questions it states the answer. On plain questions it acknowledges in one short line and names the next step. The default behaviour of most models is to open with a preamble restating the question, and that costs you the most valuable line in the reply.

Length ceiling, by register. Transactional answers should be one to three sentences. Careful answers can run to a short paragraph plus a scoping line. Plain answers should be two sentences. Without a ceiling the length is set by how much the corpus contains, which has nothing to do with what the question needed.

Forbidden constructions. This is where most of the gain is, and it is a short list: no exclamation marks, no emoji, no enthusiasm about the question itself, no sympathy templates, no rhetorical questions back at the asker, and no closing offers of further help on serious topics. Each of those is a specific thing you can forbid and test for.

Refusal wording. Exactly what it says when the corpus does not cover the question, when the topic is out of scope, and when the question needs a person. Write these as fixed sentences rather than leaving them to be generated, because they are the answers that matter most and the ones a model is most likely to pad.

Where the Bot Should Decline to Be Warm

There is a category of question where the correct behaviour is not a better answer. It is a shorter one.

A grievance, a safeguarding concern, a disclosure about health, a dispute about pay with a named manager in it, a question that mentions bullying or harassment, anything describing a personal crisis. For every one of these, any answer the system produces is the wrong answer, and the quality of the prose is irrelevant.

What it should do is three sentences at most: acknowledge in one line without attempting to interpret, state that this needs a person, and name the route with enough specificity to be actionable. A name or role, a channel, and when they will respond.

What it must not do is the set of things that feel helpful and are not. It must not summarise the grievance policy, because the person asking is not researching procedure, they are starting one. It must not ask clarifying questions, because every question is a request to type the difficult thing again. It must not offer to help further. It must not express sympathy at length. And it must never record and route the content silently without saying so, which is a privacy problem dressed as efficiency.

The hard version of this rule: it should decline even when the corpus technically contains an answer. A well-stocked assistant can produce a decent explanation of the grievance process. Doing so is still wrong, because it has answered the procedural question the person did not ask instead of responding to the situation they described.

What People Actually Need From the Voice

Why does the friendly tone annoy people?

Because it slows them down on the questions they ask most. The majority of inbound is somebody wanting one fact while doing something else, and three sentences of pleasantness before the fact is three sentences of friction repeated forty times. Friendliness has a cost that nobody measures and people feel it as the system being slow to the point.

Why does the formal tone fail too?

Because on the hard questions it sounds like a policy document, and the person asking already found the policy document unhelpful. Formality is fine for facts and inadequate for distress, where what is needed is not a register at all but a handover. The mistake is thinking the choice is between two tones when the real answer is a different behaviour.

Why do people distrust hedged answers?

Because hedging is indistinguishable from not knowing. An answer carrying four qualifications, two references to individual circumstances and a recommendation to check with HR has told the employee nothing, and they will go and ask a person, which is the outcome the whole deployment was meant to avoid. Scope the answer once and state the rest plainly.

Why does answer length matter so much?

Because length is read as a signal about the question. Six paragraphs in response to a yes or no question says the system did not understand. One line in response to a question about pay says the system did not take it seriously. People calibrate effort against the thing they asked, and a mismatch in either direction registers before the content does.

What do people actually want when they ask something difficult?

To reach somebody, quickly, without explaining it twice. That is the whole requirement. Everything else the system could offer on that question is in the way, and the measure of success is how few words stand between the question and a named person.

The Three Approaches

One tone everywhere

Set a single persona and apply it to everything. This is what most deployments do, usually because it is what the configuration screen offers.

When this is defensible. When the corpus is narrow and genuinely contains only transactional content: a benefits enrolment assistant, an IT how-to agent, anything where no question in scope can be distressing. Narrow the scope and one tone is correct.

When it fails. The moment the surface accepts open text about anything HR-related. You cannot stop somebody typing a bereavement into a box that answers holiday questions, and when they do, the single tone will be wrong.

Topic-routed registers

Classify the question first, then answer in the register that topic calls for. Most products can do a version of this through instructions, and better ones support explicit routing.

When this is right. Almost always, for any assistant covering a general HR corpus. It is more configuration work and it is the difference between a system that is adequate on hard questions and one that is embarrassing.

When it fails. When the classification is wrong, which it sometimes will be. The mitigation is asymmetry: make the classifier eager to treat something as sensitive. Routing an easy question to the careful register costs a slightly over-serious answer. Routing a hard one to the transactional register costs a great deal more.

Narrow scope, no sensitive topics at all

Restrict the corpus so the assistant answers only process and policy facts, and have it refuse everything else by default.

When this is right. As a first deployment, for a team with no capacity to design and test three registers. It is the honest minimum: a system that does one thing properly and says so.

When it fails. When the refusal is badly worded, which converts a reasonable limitation into a wall. A refusal that names a person and a route is a service. A refusal that says it cannot help with that is a dead end, and people remember dead ends.

Five Questions That Decide Your Configuration

Can somebody type a distressing message into this surface? If yes, you need the plain register, regardless of how narrow the intended scope is. Intent does not constrain input.

What does the first sentence do today? Go and ask it five real questions and look only at the first sentence of each. If any of them restates your question back at you, you have a preamble problem that is costing you the most important line of every answer.

Who wrote the refusal sentences? If the answer is nobody and the model generates them, they are the least controlled part of a system where they are the most consequential part. Write them.

Is there a named route to a person? Not an address. A role or a name, a channel, and a response expectation. An escalation that says contact HR is not a route, because the person asking contacted HR, and this is what answered.

Has anybody read the answers as the person who asked? Reading your own bot's output as an evaluator is a different activity from reading it as somebody in difficulty. The second one is the test that matters and it takes twenty minutes.

The Options

Matram, named below, is owned by the same people who publish HROpsLab. Its limitations are stated as plainly as everybody else's.

Matram

A document-trained answering agent with a strict mode that declines rather than guessing when the corpus does not cover a question, which is the behaviour the plain register depends on. Pricing is flat at $29, $69 and $199 a month with unlimited seats.

Relevant here: strict mode gives you a controllable refusal, and the tone instruction applies across the agent rather than per topic, so topic-routed registers take deliberate prompt work rather than arriving as a feature. Its cons stand: it has no free tier, it answers rather than actioning tickets, and it is not an HR or ITSM workflow platform, so the named route it offers has to point at something outside the product.

Freshservice

Service management with Freddy AI over a knowledge base, at $19, $49 and $99 per agent per month plus a $29 per agent per month AI add-on.

Relevant here: because it sits on a ticketing engine, the handover from a hard question to a person is a real route rather than a sentence, which is the strongest argument for a platform in this specific area. The answering tone controls are adequate rather than fine-grained.

Zendesk

Mature support tooling with an AI layer. Suite Team is $55 per agent per month, Copilot is $50 per agent per month, so $105 per agent per month together. At twelve agents that is $1,260 a month.

Relevant here: it has the most mature escalation behaviour of anything on this list, built for exactly the case where a conversation becomes a person's problem. Its tone defaults are customer-service shaped, which is close to the wrong register for HR distress and is adjustable.

Leena AI and Moveworks

Enterprise answering and actioning platforms. Neither publishes a price, both sell with implementation attached.

Relevant here: both are built around knowing who the asker is and routing to the right human, which is what the plain register needs. The sensitive-topic handling is typically part of the implementation conversation, which is genuinely useful and is also why the evaluation takes months.

Crisp

A messaging platform with a chatbot layer, priced per workspace at a free tier and then $45, $95 and $295 a month.

Relevant here: it is a conversation tool first, so a human taking over mid-conversation is native, which matters more for this topic than knowledge depth does. The knowledge layer is lighter than the dedicated answering products.

SiteGPT

A document and site-trained answering agent at $59 and $129 a month billed monthly, or $468 and $948 per year billed yearly. The annual plans work out near $39 and $79 a month billed yearly and should not be quoted without that qualifier.

Relevant here: it is designed for public-facing surfaces, where the distress case is rarer and the handover options are thinner. For an internal HR corpus that assumption works against you.

Intercom Fin

Priced per resolution at $0.99 rather than per seat, which is an unusual model worth understanding before comparing it to anything else.

Relevant here: per-resolution pricing means a declined question costs nothing, which aligns the commercial model with the behaviour you want on sensitive topics. It is built for customer support, so the register defaults need work for HR use.

The Comparison

The second and third columns decide this category. The fourth is where most evaluations start.

The Decision Table

Writing the Three Registers Down, Which Nobody Does

Every deployment discussed above assumes somebody has written the registers as instructions. In practice the tone lives in one sentence in a configuration field, usually something like be friendly and helpful, and the rest is whatever the model does by default.

So here is the artefact. One page, three blocks, written before launch.

The transactional block. State the answer in the first sentence. One to three sentences total. No opening pleasantry and no restating the question. Name the document the answer came from. Do not offer further help; if the person has another question they will ask it. Where a figure or a date is involved, give it rather than describing where to find it.

The careful block. State the answer in the first sentence, same as above. Then one sentence naming what the answer does not cover: the populations it excludes, the circumstances where it differs, or the fact that a contract takes precedence. Then a named route to a person, every time, regardless of how complete the answer looks. Four to six sentences total. No distributed hedging, which means no individual circumstances may vary attached to each separate statement; the scoping sentence does that work once.

The plain block. Two sentences. The first acknowledges what the person said without interpreting it, restating nothing. The second names a person or role, a channel, and when they respond. No policy content, no clarifying question, no offer of further help, no expression of sympathy beyond the acknowledgement. Then stop, and say plainly whether anything has been recorded or passed on, because an employee in difficulty is entitled to know that before they decide how much more to type.

Two things make this page work rather than decorate a folder. Each block names the forbidden constructions explicitly, because a positive instruction to be concise does not stop an exclamation mark and a negative one does. And the refusal sentences are written out verbatim rather than described, so the three most consequential answers in the system are fixed text somebody chose instead of generated text somebody hopes about.

Then put a name and a date on the page. When somebody adjusts a prompt in four months because answers feel too curt, this is the document that says the curtness was deliberate and which register it applied to.

Testing Tone, Which Takes Twenty Minutes

Tone is the easiest thing in this category to test and the thing least often tested, because testing it feels subjective. It is not. Write the list, ask the questions, read the answers.

Build a fixed set of fifteen questions. Five transactional, five careful, five plain. The plain five should be genuinely hard: a bereavement, a grievance about a named manager, a mental health disclosure, a question about a pay dispute, and something ambiguous that could be either a policy question or a crisis. Write them in the words a person would use, with the typos and the incomplete sentences.

Read only the first sentence of all fifteen. This catches preamble, enthusiasm and misrouted registers faster than reading full answers does. If the first sentence of the bereavement answer is not an acknowledgement and a route, nothing further down fixes it.

Then read the plain five in full, as the person who asked. Not as somebody checking a deployment. This is the step that produces the uncomfortable findings, and it is why it should be done by somebody who was not involved in the configuration.

Count the things on the forbidden list. Exclamation marks, emoji, offers of further help, sympathy templates, rhetorical questions. These are countable, which means this part of the test is not a matter of taste and can be re-run by anybody.

Re-run the same fifteen after every prompt change. Tone regresses silently when somebody adjusts an instruction for an unrelated reason, and the plain five are where the regression will be invisible until it reaches a real person.

The Ambiguous Middle, Where Routing Actually Breaks

The three registers are easy to apply at the edges. Payday is transactional. A bereavement is plain. The interesting failures live between them, and they are worth naming because they are where a sensible configuration still gets it wrong.

The policy question that is actually a crisis. Somebody asks what the sickness absence policy says about long-term absence. That is a transactional question on its face and it is sometimes a person working out whether they are about to lose their job. The register cannot be chosen from the words alone, which argues for the careful register as the default on anything touching sickness, pay or termination, with the named route attached whether or not the question appears to need it.

The crisis question that is phrased as admin. My manager has asked me to sign something, what is the process. Entirely procedural language, potentially a dispute. The tell is usually a named individual and a request for process in the same message, and a classifier that treats the combination as sensitive will over-trigger slightly and be right about the thing that matters.

The question that is genuinely both. How much compassionate leave am I entitled to, asked by somebody whose parent is dying. The transactional answer is wanted: they need the number. The plain register is also correct: they need a person. The resolution is the careful register with a plain opening, which means give the number, then name the route, and leave out every other sentence that would normally pad a careful answer.

The second attempt. Any question asked twice in a conversation has already failed once, and repeating a better-phrased version of the same answer is the response people find most frustrating in this entire category. A second attempt on the same topic should change register rather than wording, moving to a route regardless of what the question is about.

The pattern across all four is the same. Route on topic sensitivity rather than on apparent emotional content, because the words people use to describe a crisis are often flat and procedural, and the words that sound dramatic are often about a parking space.

When the Employee Says That Is Not What I Was Told

One more register that nobody writes, and it comes up weekly once a system has real users.

An employee reads a correct answer and disputes it, because a manager told them something different, or because the practice in their team diverges from the policy, or because they are remembering an older version of the rule. The model's default behaviour here is the worst available: it apologises, softens, and either restates the same answer more tentatively or starts agreeing with the person.

Both of those are wrong. Restating the same thing more tentatively converts a correct answer into an uncertain one for no reason. Agreeing is worse, because the system has no basis for it and has just confirmed something that may be false.

The correct behaviour is three sentences and no apology. Repeat the answer once, unchanged, and say which document it comes from and when that document was last reviewed. Acknowledge the conflict as a fact rather than as a fault: the policy says this, and if a manager has said otherwise, that difference needs resolving by a person. Then name the route.

That is not stubbornness, it is the only useful thing the system can do. A divergence between written policy and local practice is a real organisational problem, and an assistant that quietly folds in the face of pushback hides it, while one that states the document and escalates surfaces it to somebody who can fix either the practice or the policy.

So specify this alongside the three registers. It is the only case where holding a position firmly is the kinder behaviour.

What Getting This Wrong Costs

The cost is not a bad answer. It is the end of the deployment, delivered by one screenshot.

A sympathetic-sounding template on somebody's bereavement gets shared. It goes to a manager, or into a team channel, or to whoever is sceptical about the project, and it does more damage in one image than six months of correct answers repair. Nobody can defend it, because it is indefensible, and the argument about the other 9,000 correct answers is not an argument anybody can win in that moment.

There is a slower cost too. Hedged, padded answers on ordinary questions teach people that the assistant is a slower route to a worse answer than asking a colleague. Adoption falls off without anybody filing a complaint, and the usage graph looks like a product problem rather than a tone problem, which sends the next round of work in the wrong direction.

And there is the cost to the person who asked. That is the one worth leading with internally. Somebody in difficulty reached out to their employer and got a template with an emoji on it. The project cost and the adoption curve are secondary to that, and framing it that way is also the most effective way to get the configuration work funded.

When You're Ready to Go Further Than Tone

Once the three registers are written, tested and holding, two things are worth doing next and neither is about voice.

Make the handover real. A plain-register answer that names a person is only as good as what happens when somebody follows it. If the named route is a shared inbox with no owner, the register is honest and the outcome is the nine-day disappearance that the whole thing was meant to prevent. Fix the route before refining the sentence that points at it.

Review the sensitive-topic log, carefully and with restraint. You want to know how often the plain register fires and whether it fired correctly, which means somebody has to look. That same log contains the most private things anybody has typed into a company system, so decide who may read it, how long it is kept, and what gets recorded, before you start reviewing it rather than after. Counting that the register fired is usually enough; reading the content rarely is.

Frequently Asked Questions

HROpsLab takes no vendor money and publishes no paid placements.