How to choose an omnichannel inbox: the nine questions that decide it
Every tool in this category demos well. These are the nine questions that separate them — and what a bad answer to each one sounds like.
Every tool in this category demos well. They all show one queue with messages from four channels, and at that level of detail they are identical.
The differences show up in month three, and they are always in the same nine places. Here they are, with what a bad answer sounds like — because the answer you get is more informative than the feature list.
We build one of these, so read accordingly. The questions are the ones we would ask, and a few of them are ones we would rather not be asked.
1. What happens when the same person messages on two channels?
Why it decides things: this is the difference between an inbox and a list. If Instagram-them and WhatsApp-them are two records, your team will ask a customer to repeat something they already said, which undoes the whole point.
A bad answer: “You can search by name.” Searching is not merging.
What good looks like: one contact, one history, automatic where identifiers match and manual where they do not.
2. What states can a conversation be in, and who sets them?
Why: a queue without states is a longer list. You need at least open, in progress, resolved — and abandoned, which is the one most tools skip and the one that measures lost demand.
A bad answer: “Read and unread.”
What good looks like: states that are the same words on every channel, set automatically where possible, and reportable.
3. Where do automated answers come from?
Why: this is the single biggest determinant of whether the thing is safe to point at customers. An answer assembled by a model from general knowledge will be confident and occasionally wrong about your prices.
A bad answer: anything involving “the AI figures it out” or a demo that answers a question you never gave it information about. That is the failure, being presented as the feature.
What good looks like: answers from sources you maintain, and an explicit, demonstrable refusal at the edge of them.
4. Show me it failing.
Why: every vendor can show you a success. How a tool behaves when it does not know is the thing you are actually buying, because that is most of the hard cases.
A bad answer: a reluctance to do it, or a fallback that invents something plausible.
What good looks like: they volunteer it, and the failure is a clean handover with the conversation attached.
5. What does escalation carry with it?
Why: an escalation that arrives as “customer needs help” has saved nobody any time.
A bad answer: “It notifies your team.”
What good looks like: the thread, the detected intent, the contact details, and a route to a specific person or team rather than a shared pile.
6. Can you route by channel and number, or only by conversation?
Why: hand-routing works at two channels and collapses at five. This is the question that decides whether the tool survives you growing.
A bad answer: “Anyone on the team can pick up anything.”
What good looks like: rules — this number goes to this team, this channel to those people — set once.
7. What can I measure without exporting anything?
Why: a metric that needs a monthly CSV is a metric nobody looks at twice.
A bad answer: message volume, and a promise that reporting is on the roadmap.
What good looks like: response and resolution times, abandoned conversations, share handled without a person, and what people actually asked — in the product, per channel, over a date range.
8. What happens to my data if I leave?
Why: you are putting your customer conversation history into this. Ask on day one, not on the day you want to go.
A bad answer: hesitation, or a support ticket.
What good looks like: a self-serve export of conversations and contacts, in a format something else can read, without asking anyone.
9. What does this cost at three times my current volume?
Why: pricing in this category is usually per conversation, per seat, or both, and the curve matters more than the headline. Messaging volume is also not yours to control — it grows with your marketing.
A bad answer: a single monthly figure with no unit attached.
What good looks like: a clear unit, a worked example at your volume and at triple it, and a straight answer about what the platforms themselves charge on top. For WhatsApp in particular, Meta bills per conversation independently of your vendor, and any quote that does not mention it is incomplete.
The test that beats all nine
Take fifty real conversations out of your current inbox — messy ones, not the tidy examples — and put them through each shortlisted tool. Then count three things:
- Answered correctly from your own information
- Escalated — and whether those were the right ones
- Answered confidently and wrongly
The third number decides it. A tool that scores brilliantly on the first and above zero on the third is not ready to talk to your customers, because you will not find those answers until a customer does.
Two hours, and it tells you more than a month of demos.
Fellix answers those nine the way this post says to. See how the Front Desk works, or run your own numbers.