Autonomous AI Customer Support: When to Send vs. Approve

Klarna built a fully autonomous AI support system, handled a large share of their contact volume, and then quietly started rehiring human agents. The lesson is not that AI failed. It is that fully autonomous replies, across all query types, at scale, created quality problems they could not catch fast enough. You are not Klarna. But the failure mode is identical at 200 emails a month.

The short version

Autonomous AI works for a predictable slice of your volume. Draft-and-approve covers the rest. You do not have to pick one mode and stick with it. The practical move is to build a short safe list of query types you are comfortable automating fully, put everything else in draft mode, and expand the safe list over time as you see how the AI actually performs. That is it. Everything below is the reasoning and the mechanics.

This is a decision framework, not a philosophical debate about whether AI is good or bad at support. It is good at some things and bad at others. The job is to figure out which is which for your specific inbox.

What Klarna actually did (and what small teams get wrong about it)

Klarna deployed an AI agent that handled a large share of their customer contacts autonomously. No human in the loop, no draft review. For a while, the numbers looked great. Then quality problems started showing up in places they were not catching in real time. They started bringing human agents back.

Most teams I have talked to hear this story and conclude one of two things: either AI is overhyped and they should wait, or Klarna just did it wrong and they will do it better. Both miss the point.

The actual lesson is simpler. Fully autonomous, across all query types, means you are not seeing the bad replies before they go out. At Klarna's volume, even a 1% error rate is thousands of bad replies a day. At your volume, 50-500 emails a month, a 1% error rate is basically nothing numerically. But one bad autonomous reply to the wrong customer at the wrong moment - someone threatening a chargeback, someone about to cancel a big account, someone who is already furious - costs more than the time you saved that week.

Two modes every AI support tool should offer

Autonomous mode: the AI reads the incoming message, generates a reply, and sends it. No human sees it before it goes out. Fast, zero friction, and genuinely useful for the right queries.

Draft-and-approve mode: the AI reads the message, generates a reply, and parks it as a draft. You open it, read it, make any edits, and hit send. One extra step, but you catch mistakes before they ship.

Most tools default to autonomous because it looks more impressive in demos. 'The AI just replied for you' is a better demo than 'the AI wrote a draft and you approved it.' But for most small teams, autonomous-first is the wrong default. You have not seen the AI perform yet. You do not know where it gets things wrong. Sending first and reviewing later means your customers are the ones finding the errors.

In Trigli, draft-and-approve is the default. The AI drafts inside Gmail, you approve before anything goes out. Autonomous send rules exist, but they are opt-in and you set them per query type. That is a deliberate choice, not a missing feature.

The queries where autonomous is almost always safe

Here is a useful mental test. If a 22-year-old intern could answer the question correctly on day one, with no context about the customer, just a copy of your FAQ doc, autonomous is probably fine. The answer is deterministic. There is no judgment call. Getting it wrong is unlikely, and even if the AI is slightly off, the stakes are low.

Query types that usually pass this test include order status or tracking requests where the answer is a direct data lookup, hours and location questions with a single correct answer, password reset instructions or how-to links that do not vary by customer, refund policy questions where the answer is 'here is the policy' and not 'here is your refund,' pricing page requests or feature comparison questions you have documented, and shipping time estimates for standard orders.

In my experience, 40-60% of ticket volume fits this bucket. That is a real number. If you are handling 300 emails a month and 150 of them are genuinely answerable with no judgment, autonomous on those 150 saves you meaningful time without meaningful risk.

The queries where you should always require approval

Some queries are not about information. They are about a situation. The AI can generate a technically accurate reply and still make things worse. These should always go to draft mode: any message containing the words lawyer, lawsuit, attorney, BBB, chargeback, or fraud; angry or emotionally escalated messages where even a rough sentiment signal should route to draft; refund or compensation requests above a dollar threshold you set (a reasonable default is anything over $50 gets a human); messages from customers you have flagged as high-value, enterprise, or VIP; any query the AI flags as low-confidence or outside its knowledge base; and first contact from a brand-new customer where tone-setting matters more than speed.

The grey zone: queries that look simple but are not

'Can I cancel my subscription?' looks like a simple FAQ. It is actually a retention moment. The right answer might be 'here is how to cancel' or it might be 'let me find out why you want to cancel and see if there is something we can fix.' An autonomous reply that just sends the cancellation link skips the conversation entirely.

'Is this compatible with X?' looks like a product question. But if the AI gets it wrong and the customer buys based on that answer, you have a return, a frustrated customer, and a trust problem. The stakes are higher than they look.

Politely worded complaints are the hardest one. A customer who writes 'I just wanted to let you know that my order arrived damaged, no big deal, just thought you should know' is not actually fine. Sentiment tools miss this constantly. The message is calm but the situation is not.

The rule for the grey zone: default to draft mode for anything not explicitly on your safe list. Not the other way around. If you are not sure, draft.

A practical trigger-rule template for a 1-10 person team

Here is how to actually set this up. It should take under an hour.

1. Start with a safe list of 5-10 query types you are genuinely comfortable automating. Write them down explicitly. 'Hours and location,' 'password reset link,' 'pricing page link,' 'shipping time estimate.' Specific beats vague.
2. Build a block list of keywords that always force draft mode regardless of query type. Start with: lawyer, lawsuit, attorney, chargeback, fraud, BBB, refund, cancel, angry, furious, disgusted, legal.
3. Set a confidence threshold below which the AI always drafts instead of sends. Most tools expose this as a setting. If yours does not, that is a red flag.
4. For the first month, review every autonomous send weekly. Look for anything that went out and made you wince. Add those query types to draft mode.
5. After 90 days, if your autonomous sends look clean, expand the safe list. Earn your way to more automation, do not assume it.

Concrete example: a 3-person SaaS team I know runs autonomous on exactly four things - business hours, a link to their pricing page, password reset instructions, and links to their documentation. Everything else goes to draft. That setup took them 45 minutes to configure and has not caused a problem in six months.

For more on building this kind of setup, the post at /blog/automate-customer-support-without-losing-human-touch walks through the two-column method in detail.

What draft-and-approve actually costs you in time

Let us be honest about the math. If 50% of your volume is autonomous, you are reviewing the other 50%. At 200 emails a month, that is 100 drafts to review. At 2 minutes each, that is about 3-4 hours a month.

A full-time CX rep runs $3,000-$4,000 a month fully loaded when you include salary, benefits, and management overhead. The post at /blog/cost-of-support-rep-vs-ai breaks down that number in detail if you want to pressure-test it. Three to four hours of draft review versus $3,000-$4,000 a month is not a close comparison.

There is also a benefit to draft review that people underestimate. Every draft you review is a feedback loop. You see what the AI got right, what it got slightly wrong, and what it missed entirely. That is how you improve your knowledge base over time. Fully autonomous means you lose that feedback signal.

Where fully autonomous AI support makes sense eventually

After 3-6 months of data on your autonomous send accuracy, you can expand the safe list with real confidence instead of guessing. You will know which query types the AI handles cleanly and which ones it fumbles.

High-volume, low-complexity products get there faster. A SaaS product with a narrow feature set and a well-documented knowledge base can often automate 60-70% of volume cleanly within a few months. A service business with lots of custom situations might stay at 30-40% autonomous indefinitely. Both are fine.

The goal is not 100% autonomous. The goal is the right percentage autonomous for your specific query mix. Anyone selling you 100% autonomous as the target is selling you a demo, not a real support operation.

What Trigli does here, and what it does not

Draft-and-approve is the default mode in Trigli. The AI drafts inside Gmail, you approve before anything goes out. Autonomous send rules are available but opt-in, and you set them per query type or keyword. The AI learns your reply style from past sent emails, so drafts sound like you wrote them rather than like a generic support bot.

Trigli also includes a chat widget bundled with the email product. Same knowledge base, same confidence-based routing, same draft-or-send logic. You are not paying extra for a separate chat tool.

No helpdesk workflow routing or SLA tracking. If you need ticket assignment rules, escalation queues, and a reporting dashboard, Trigli is not that. No phone or SMS support. No Outlook yet - Trigli currently supports Gmail via OAuth 2.0, with Outlook support planned for a future release. No native Shopify order actions, so if you need the AI to actually pull order data and process returns inside Shopify, you will need a different tool or a custom integration. No enterprise reporting.

Be honest with yourself about what you actually need. If you are a 3-person team running on Gmail with 200-400 support emails a month, Trigli fits. If you need a full helpdesk with ticket routing, Trigli is not the right call.

Where other tools land on this spectrum

I am not going to pretend the other options are bad. They are not. But they make different default choices.

Intercom Fin defaults to autonomous resolution. It charges per resolution - roughly $0.99 per Fin resolution at last check. Draft mode is not the primary UX. That is a real incentive to understand. The AI gets credit for closing tickets, not for flagging ones it should not touch. That is not necessarily bad, but it is the dynamic you are buying into.

Zendesk AI is autonomous-first and built for teams that already have Zendesk workflows. If you are already running Zendesk and want AI layered in, it is a natural fit. If you are not already on Zendesk, the setup cost is real.

Where Intercom Fin wins

Fin is genuinely strong if you are already inside the Intercom ecosystem. The workflow integrations are deep, the handoff to human agents is smooth, and the per-resolution pricing can actually be cheaper than a flat monthly fee if your volume is low and your resolution rate is high. If you are running a high-volume product support operation and already paying for Intercom, Fin is a natural fit. It also has more mature reporting than Trigli does today.

Where Zendesk AI wins

Zendesk AI is the right call if you need enterprise-grade SLA tracking, ticket assignment rules, escalation queues, and a reporting dashboard that your whole support org can work from. If you have a team of 10 or more agents and compliance requirements around response times, Zendesk is built for that. Trigli is not. Zendesk also has a much larger app ecosystem if you need integrations beyond Gmail.

Trigli's flat pricing - $49/month for Starter, $149/month for Growth, $349/month for Pro - means there is no financial pressure on the AI to close tickets autonomously. The pricing does not change based on what the AI resolves. See the full breakdown at /pricing.

For a longer comparison of the autoresponder landscape, the post at /blog/ai-email-autoresponders-small-business covers several tools side by side.

An honest summary: the right default is draft mode

Start in draft mode. Earn your way to autonomous on specific query types. Autonomous is a privilege you grant after you have seen the AI perform, not a setting you turn on day one.

The Klarna lesson scales down. Even at 50 emails a month, one bad autonomous reply to the wrong customer costs more than the time you saved that week. The customer who gets a chipper AI reply when they are threatening a chargeback does not forget it.

Draft mode is not a limitation. It is the safety net that lets you actually trust the system over time. Once you trust it, you expand. That is how you end up with a support setup you actually trust, rather than one you are quietly nervous about.

How to get started today

Pull up last month's support emails. Tag them by query type. You are looking for your safe-list candidates - the ones where the answer is deterministic and the stakes of a wrong reply are low. That audit alone will tell you what percentage of your volume is genuinely automatable right now.

Pick a tool that lets you run both modes simultaneously. Not all do. Some tools are autonomous-only with no draft option. If you cannot run draft-and-approve on the queries you are not sure about, you are flying blind.

Start on a free tier or trial before committing to anything. Trigli's free tier is 50 emails, 25 chats, and 10 tickets per month with no credit card required. That is enough to run a real test on a week of your actual inbox. If the drafts look good and the autonomous sends on your safe list are clean, you have your answer. If they are not, no hard feelings - you found out for free.

For SaaS teams specifically, the setup at /use-cases/saas shows how teams with narrow feature sets tend to configure the safe list and where they typically land after 90 days.