Almost nobody tells you support was the reason they left. They just take the next option they’re offered. We build the helpdesk, the content and the team that stop that — ten-plus implementations across the EU, UK and US. Fixed price, fixed scope, never an hourly rate.
Before the sale, an unanswered question is an order that never happens. After the sale, a badly handled one is a customer who stays just long enough to find an alternative. Both are invisible in the dashboard your helpdesk ships with, because neither leaves a complaint behind.
“will this hold my weight? i’m about 240 lbs and i don’t want to order the wrong one”
The weight rating was published in the help centre the whole time — on a different system to the chat widget on the storefront. Two front doors, one lost order, and nothing in the queue to show it happened.
Never becomes a ticket“this is the third time i’m asking about the refund”
The refund arrived in the end, the ticket closed as resolved, and the CSAT survey went unanswered. They never ordered again, and they never said why. Renewal season simply came up short.
Closes as resolvedThis is the whole argument for treating support as a revenue function rather than a cost centre — and the reason every engagement here starts by measuring what your setup can answer today, not by counting how many tickets it closed.
Zendesk and Intercom only. Two platforms done properly beats a logo list of twenty done once each.
A fixed-scope examination of how your support actually performs — not just how it is configured. You get scored findings, a prioritised fix list with the revenue attached to each, and a ninety-day plan your team could run without us.
Zendesk or Intercom stood up the way it should have been the first time — or an inherited setup rescued from whoever left. Configured on your tenancy, documented for your team, handed over with no dependency on us.
The tools and the team. For founders carrying support themselves, or a first support hire with no one above them to learn from. We design the function, help you staff it, and stay close until it runs without us.
Every assistant we assess gets the same five scenarios, typed by hand, scored the same way. Identical inputs are the only honest way to compare your bot in March against your bot in June — or against your competitor’s.
| Scenario | What gets sent | What it exposes |
|---|---|---|
| Refund | “arrived yesterday, i want to send it back” | Whether current policy is retrievable — or whether the bot quotes a window you changed last year. |
| Ambiguous | “is this one right for me?” | Whether it asks a clarifying question or confidently guesses. Guessing is the expensive answer. |
| Misspelled | “wheres my ordr, orderd tuseday” | Retrieval robustness against how customers actually type, rather than how your articles are titled. |
| Two-part | “do you ship to Canada and how long does it take?” | Whether both halves get answered. Dropping the second half is the most common silent failure. |
| Escalation | “i just want to talk to a person” | Whether handoff works at all, how long it takes, and whether the agent arrives with any context. |
We run these by hand, as a customer would. No scripted bots talking to your bot — that wastes your agents’ time, trips abuse detection, and produces worse data than five careful conversations.
Everything below is read from what your help centre already publishes. No logins, no credentials, no access to your tickets. If your setup is in good shape, the scorecard will say so and we’ll leave you alone.
Twelve automated checks on what your setup already publishes, written up as one page of findings with a measured deflection ceiling. Yours to act on with or without us.
The deep version. Five scripted scenarios against your live assistant, full content and config scoring, every revenue dead-end named alongside the article that should have caught it, and a costed list of what to fix first.
Helpdesk setup, migration, the AI layer, or the whole support function including the people — scoped from the review, so the number is real before you commit to it. Configured on your tenancy and documented for your team.
Monthly, against real conversations rather than a launch-day benchmark. Failed answers reviewed, new intents captured, content gaps reported to whoever writes your docs — and, where we built the team, QA and coaching alongside it.
I’m Aman Bhatia. Thirteen years in software, and more than ten helpdesk implementations behind me — for brands selling across Europe, the UK and North America. Global rollouts have a particular shape: several languages, regional policies that quietly contradict each other, and a support team spread across time zones who all have to answer from the same source of truth. Most of what goes wrong in a support setup goes wrong at exactly those seams.
The part that matters most is that a good chunk of those years was spent inside customer support and customer success rather than advising it. I have owned an Intercom implementation end to end and lived with the consequences of my own configuration decisions afterwards — which is why the migration checklist on this site is opinionated about unglamorous things like orphaned tags and macros nobody has fired since 2022.
The last two years have been retrieval systems and AI agents full time, most recently on a Fortune 500 pharmaceutical programme where field teams use what I build every day, and where a rework of how the system called its models took roughly a million dollars a year out of the run cost. That one stays unnamed — it sits under NDA.
What you get here is the person who does the work. No account manager, no hand-off to someone you meet once at kickoff. Two engagements at a time, which is the honest constraint and the reason the scorecard exists: it stops me pitching people who don’t need me.
One page back within a business day — whether or not there is work in it for us.