runix

Serial entrepreneur: 2 exits 😎, 2 fails. Building no-code AI chatbot (Chattermate) and Cheerpop (Share special moments) 🚀. 💻😊. Seed Investor

[runix @runix_b]

The counterweight: a bot that bails at the first hint of difficulty deletes the reason you built it.

The goal isn't escalating constantly. It's escalating at the right moment, so it reads as the bot doing the customer a favour rather than giving up.

[runix @runix_b]

Most bots escalate too late. They wait until the customer types "GIVE ME A HUMAN" in caps.

By then you've lost the argument. The customer had to fight your software to reach a person, and they'll remember that longer than whatever the problem was.

Four earlier signals:

[runix @runix_b]

3. Explicit request. Someone asks for a human, connect them. Don't make them prove the bot can't help first.

4. Topic. Cancellations, refunds over a threshold, anything legal, safety or a distressed customer. Route on sight. Don't let the bot freelance.

[runix @runix_b]

A customer spends four minutes explaining a billing problem to your bot. Account number, the charge, the date, all of it.

The bot gets stuck and transfers.

The agent opens with "Hi, how can I help you today?"

That's where you lost them. https://t.co/cldOklsqpr

[runix @runix_b]

ChatterMate just crossed 100 stars on GitHub.

⭐ 100 stars
🍴 44 forks
👥 2 paying customers
📚 24+ free guides
bootstrapped · open source · self-hostable

Last week: how you'd know the AI worked. This week: the moment it stops working.

[runix @runix_b]

Week 37, one idea: the number on your AI support dashboard was defined by the person selling you the AI.

Track your own reopen rate instead. One query, brutally honest.

Next week: what happens when the bot gives up and hands you to a human. https://t.co/cldOklsqpr

[runix @runix_b]

@killix Right, and almost no stack stores that link. The workaround is joining on customer ID + 7-day window + intent tag rather than thread ID — same person, same intent, twice. Fuzzier than a true causal link, but it catches the deflection that already banked the win.

[runix @runix_b]

The reason support AI metrics drift from reality: a wrong-but-confident answer scores identically to a right one.

Nothing in the metric penalises invention. So the dashboard looks fine while trust quietly leaks.

[runix @runix_b]

Week 4 — list every question the bot couldn't answer from your docs. Rank by frequency.

That list is your content roadmap, and working it moves your numbers more than any prompt tuning will.

Free, honest, and nobody can sell it to you. https://t.co/cldOklsqpr

[runix @runix_b]

How to find out whether your support AI is actually working, in 30 days, with a spreadsheet.

No new tools. Works whatever you're running.

[runix @runix_b]

Week 3 — tag 100 random conversations by intent. Order status, password reset, billing dispute, complaint. Score CSAT per tag, not blended.

You'll find automation is winning in two categories and quietly losing in two others.

[runix @runix_b]

Which is why ChatterMate cites the doc behind every answer — not as a setting, as the default.

A cited answer is checkable. An uncited one is a claim. You can't audit a number built out of claims.

Open source, first 300 chats free. https://t.co/cldOklsqpr

[runix @runix_b]

Unpopular opinion: containment rate is a vanity metric.

It counts conversations that ended without a human. It cannot tell the difference between a customer who got their answer and one who read a confident wrong answer and closed the tab.

Tell me I'm wrong.

[runix @runix_b]

Gartner asked the human whether the issue was resolved. Different denominator, different judge, different answer.

Wrote up why both numbers are true, and the scorecard that's harder to game:

https://t.co/s43W1DXHpN

[runix @runix_b]

Two credible sources, four months apart, on the same question.

Intercom: 76% of support conversations resolved by AI, across 7,000+ teams.
Gartner: 14% of issues fully resolved in self-service, from 5,728 customers.

Sixty-two points apart. Neither is lying.

[runix @runix_b]

Containment counts conversations that started in the bot and ended in the bot. That's the whole definition.

It doesn't ask whether the customer got what they came for. Someone who reads a confident wrong answer and closes the tab is a perfectly contained conversation.

[runix @runix_b]

4. Coverage gaps.

Gartner: 43% of failed self-service was customers not finding relevant content. Not a model problem, a docs problem.

Every question your bot couldn't answer is a ranked to-do list for your help centre. Work that list. https://t.co/cldOklsqpr

[runix @runix_b]

Containment rate is the easiest support metric to game.

Bury the "talk to a human" button. Require two failed AI attempts before escalation. Lower the confidence threshold so the bot always says something.

All three move the number up. All three are bad ideas.

[runix @runix_b]

2. Escalation quality, not volume.

A handoff where the AI already has the order number, verified the account and summarised the issue beats a "resolution" nobody cared about.

3. CSAT split by intent. Blended CSAT hides everything worth knowing.