Software Engineer Intern

Two in three support teams just shipped an AI Agent. Median resolution: 41%. Salesforce says 66% of support teams shipped an AI agent this year. The median resolution rate is 41%. Here's why the handoff is where the money actually leaks. Salesforce says 66% of customer service organizations now run at least one AI agent. Last year it was 39%. The vendors are on stage taking a victory lap[[1]]. Then you look at the resolution numbers. The enterprise median for tier-1 AI deflection in 2026 is 41.2%, per Zendesk CX Trends and Salesforce State of Service data aggregated by ClarityArc[[2]]. Bottom quartile: 22.4%. That means the typical operator who spun up an "AI customer service agent" this year is deflecting fewer than half the tickets it touches - and a big chunk of them are shipping systems that barely clear 1 in 5. Six in ten conversations still route to a human. And when they do, the handoff is usually broken. That's the story nobody in the launch keynotes is telling. The pitch vs. the math. Every big vendor released a customer-service agent product this year. Salesforce is pushing Agentforce. Zendesk shipped its AI Agents platform. Pylon launched an "Agentic Support Platform" on YouTube to 3.5M views[[3]]. monday.com's "welcome your AI agents to the team" spot crossed 12M[[4]]. The pitch is the same on every stage: cut cost per contact, replace tier-1 headcount, give customers 24/7 answers. The math is more honest than the pitch. * Live-channel cost per contact (phone, chat, email) averages $8.01 according to Gartner benchmarks[[5]]. * Self-service cost per contact is about $0.10. * On paper, the cost delta is 80x. * In practice, IBM measured a 23.5% average reduction in cost per contact from conversational AI deployments[[6]]. * Real operators with clean back-end integrations report 20-40% cost-per-contact reductions over 2-3 years[[7]]. Nobody's getting the 80x. Because 41% deflection means 59% of contacts still land on a human - and the humans now spend more time per ticket, not less, because the ones that escalate are the hard ones. The Lorikeet CX team put realistic ranges on it: 30-50% resolution for early deployments, 50-70% as workflows mature, and 70-85% for deeply integrated, action-taking agents on well-scoped use cases[[8]]. If your team is at 41%, you're average. If you're at 22%, you're paying vendor fees for something worse than a good FAQ page. The Klarna number everyone quotes wrong. Klarna is the canonical case study, and it's getting misremembered in two directions[[9]]. The pro-AI crowd cites the peak: 2.3M chats handled, work of ~700 agents, 25% shorter resolution times, $40M in projected profit improvement in year one. The anti-AI crowd cites the reversal: mid-2025, CEO Sebastian Siemiatkowski admitted quality had slipped and Klarna started rehiring humans. Both are true. Both miss the point. The Klarna story isn't "AI failed." It's that peak deflection ≠ sustained CX quality. They hit the deflection numbers. What they lost was the ability to catch edge-case rage cases before those customers churned. That's not a model problem. That's an escalation-design problem. If you're building an AI service agent for a $5M-$20M business right now, that's the number you need to be modeling - not deflection, not cost savings. The recovery curve on escalated tickets. The handoff is where the money leaks. Here's the metric nobody puts in the deck: cold-transfer handoffs drop CSAT by 12%[[10]]. That's not a small hit. That's the difference between a repeat customer and a chargeback. There are three ways handoffs fail. All three are common in the systems shipping right now: * The cold-context dump. The AI escalates. The human gets no summary - sometimes not even the transcript in a usable place. The customer repeats their whole problem for the third time. * The escalation black hole. The AI says "I'm connecting you to a specialist," and... nothing. No SLA. No queue position. The customer bounces to a competitor. * The undisclosed-AI trap. The customer never knew they were talking to a bot until it messes up. Now trust is gone before the human even joins. Chatbot CSAT typically runs 10-15 points below live-agent CSAT[[11]]. Which is fine - for tier-1 volume - as long as escalated conversations recover to a healthy score. They only do that when context travels with the ticket. Most of what I see in the wild? Context doesn't travel. The AI sits inside one vendor's platform. The humans work in a different tool. The handoff is a URL, not a warm transfer with a summary. What "well-scoped" actually means. The 70-85% resolution rate is real. It just requires design that most operators skip. * Scope narrow. Order status. Refund initiation. Subscription pause. Password reset. Not "answer anything." An agent that can act on 6 workflows will beat an agent that can talk about 60. * Wire the back end first. If your agent can't hit your OMS, your billing system, and your CRM with write access, it's a decision tree, not an agent. Read-only agents defer. Write-capable agents resolve. * Permission-scope every action. Refunds under $50 auto-execute. Above $50, agent proposes and a human clicks. This is the guardrail every real deployment has and every demo skips. * Design the escalation, not just the deflection. When the agent hands off, it should deliver: a one-line summary of the issue, what it already tried, the customer's mood signal, and the recommended next action. That's the difference between recovering CSAT and torching it. * Measure the four numbers. Resolved deflection rate. Cost per contact before and after. Handle time on escalated tickets. Total platform + maintenance cost. If you can't produce all four this quarter, you don't have a program, you have a purchase. Those four numbers come straight from Tommaso Maria Ricci's 2026 CX guide, and they're the right ones[[12]]. Everything else is vanity. The uncomfortable pattern I keep seeing. Two in three service teams shipped an agent this year. Salesforce is calling it success. Most of those deployments are running at median 41% resolution, sitting on top of broken escalation paths, without a real cost-per-contact baseline to prove ROI against. That's not a broken product category. It's operators buying the deflection number and forgetting to fund the plumbing that makes deflection worth having. If your ticket volume is going up quarter-over-quarter and your team is quietly hiring back the humans you thought you'd replace - you're not alone. Klarna did the same thing, publicly. Most companies are doing it privately. The good news: the fix is not another platform. It's narrower scope, real back-end integration, and an escalation path that doesn't burn customer trust when the agent inevitably tags out. If your support stack is showing this pattern - deflection numbers you can't defend, CSAT slipping on escalated tickets, and no honest cost-per-contact math - that's exactly what my audit call is for. Thirty minutes. I'll tell you which of the four numbers is missing and what your version would look like if it were designed to resolve, not just deflect. Sources 12 references * 1 Two in Every Three Customer Service Teams Now Use AI Agents CX Foundationnews AI agent adoption in customer service organizations jumped from 39% (2025) to 66% (2026), citing Salesforce research. | * 2 Customer Service AI Agent Statistics 2026: 120+ Data Points Digital Appliedreport Enterprise median tier-1 deflection is 41.2%; bottom quartile 22.4%. | * 3 Introducing the Agentic Support Platform Pylonvideo Pylon launched an 'Agentic Support Platform' - 3.5M views. | * 4 Welcome your AI agents to the team by monday.com monday.comvideo monday.com's AI agents launch spot crossed 12M views on YouTube. | * 5 Reduce Cost per Contact: What AI Really Saves OMQanalysis Gartner benchmark: live-channel contact ~$8.01, self-service ~$0.10. | * 6 75 AI Customer Service Statistics 2026 NextPhonereport IBM measured 23.5% average cost-per-contact reduction from conversational AI. | * 7 AI in Contact Center: Proven Ways to Lower Costs and Improve Efficiency RITS Centeranalysis Real operators report 20-40% cost-per-contact reduction over 2-3 years. | * 8 What Resolution Rate Can AI Customer Support Achieve (2026 Benchmarks) Lorikeet CXanalysis Realistic ranges: 30-50% early, 50-70% mature, 70-85% action-taking. | * 9 Klarna's AI Reversal: Why It Rehired Human Agents Mirai360analysis Klarna AI handled 2.3M chats, work of ~700 agents; later rehired humans. | * 10 AI-to-Human Handoff in Ecommerce: 7-Step Context Transfer Alhenaanalysis Cold-transfer handoffs drop CSAT by 12%; three handoff failure modes catalogued. | * 11 Human Handoff in AI Customer Support: How Escalation Works (2026) Machaanalysis Chatbot CSAT typically runs 10-15 points below live-agent CSAT. | * 12 AI Customer Service 2026: Cut Cost 45%, Up CSAT 30% Tommaso Maria Riccianalysis Four-number measurement framework for AI customer service programs. ai-agents customer-service cx-strategy ai-adoption operations Get the field notes AI systems, automation & growth tactics. No fluff, no spam - unsubscribe anytime.

Pylon's founders at SaaStr AI Day: how its support team deflected 50% of tickets without increasing headcount. At SaaStr AI Day, Pylon's founders argued that customer support has pursued the wrong AI metrics, demonstrating why agentic AI augmentation outweighs pure automated deflection. Marty Kausas and Advith Chelikani on why deflection rate is the wrong number in CX. And what they're replacing it with, and what you should, too. Pylon closed out the latest SaaStr AI Day with the two co-founders walking through the launch they shipped the week before, which they're calling agentic customer support. Marty Kausas is co-founder and CEO, Advith Chelikani is co-founder and CTO. The company is a little over three years old and around 1,600 customers, almost all B2B. Machine Learning & Artificial Intelligence Their argument is that support bought the wrong AI product for the last two years, and the deflection numbers everyone reports are hiding it. What the session covered: * Why a 50% deflection rate produced zero headcount change * The case for human plus AI over full replacement, and why support ran the other direction * Two live examples of a ticket that arrives already investigated, including one that replaced a manual AWS log query * Beta results from customers: 70% fewer escalations at one, 64.5% faster first response at another 1. The test is whether the job is unrecognizable. Kausas opened with an engineer who joined Pylon earlier this year after a two-year sabbatical. He left as a software engineer with no AI coding tools and came back into a job he described as unrecognizable. Before, you wrote the code, tested it, and reviewed it yourself. Now you're managing agents doing that work. That's the bar Kausas set for support. If a support person joins your company and the job feels like the job they left, nothing has actually changed. He also put up AI spend by department. Coding is the only bar that has meaningfully taken off. Every other function is still at the start. 2. The deflection number that didn't move headcount. Kausas's argument is that the fastest-growing AI companies are running human plus AI augmentation rather than full replacement. He pointed at the model labs themselves, plus Cursor in coding and companies like Harvey in legal, and noted that he uses Claude and Codex daily and neither is close to automating him. Support went the other way. Almost everyone ran at full-resolution agents, which Pylon frames as solving a real but narrow slice of the problem, especially in B2B where tickets carry more context and more relationship. The example: a company of roughly 5,000 people with about 1,000 on the support team deployed a fully automated resolution agent. It deflects around 50% of tickets. Headcount didn't change. The reason is that ticket count and work volume aren't the same measurement. What an agent can fully resolve with no human in the loop is a limited set of questions, and those are the easy ones that consumed the least time to begin with. Deflecting 50% of tickets can remove a much smaller share of the actual work. Everything hard still escalates to a person. Pylon's bet follows from that: automated resolution gets commoditized, and the value sits in helping humans do the escalated work faster. Worth noting this is a vendor drawing a line where their product sits, and the specific 50% example came from one customer conversation rather than a study. The underlying point is still testable in your own numbers. Compare your deflection rate against your support headcount and your median handle time over the same period. 3. Why DIY with Claude hits a wall. Most teams already have people asking Claude for help answering tickets, connected to the help center, feature request logs, the support system, and the code base. Kausas's position is that this gets you real but limited help. Engineering & Technology His example: one of the best ways to answer a new ticket is finding similar past tickets. A skill that has to read through every past ticket to do that is impractical every single time. Pylon's answer is precomputing the context layer instead. Account setup, goals, use cases, interaction history, and sentiment are computed ahead of time, along with related tickets, related knowledge articles, and who the contacts are and what they care about. The claimed result versus rolling your own is three to six times cheaper inference, better quality, faster responses, and team control instead of a group of people trying to co-own one skill. 4. What a pre-investigated ticket looks like. Chelikani demoed Pylon's own support queue. A customer named Ryan asks for a fix to a filter in the analytics part of the product. On the right side, the background agent has already run. What it produced before a human touched the ticket: it interpreted the screenshots and identified the analytics area even though the customer never named it, found two similar past issues from other customers, surfaced a call from the previous week where this customer had been frustrated about the same thing, and checked the code to confirm the backend was implemented but not surfaced on the front end. It concluded this was a UI regression rather than a missing feature, and suggested next steps. Karen, the support engineer, then worked interactively. She asked whether an open feature request existed in Linear (it didn't), pushed the agent deeper into the code base to confirm a suspicion, ran a skill that packaged the customer context and the investigation into a Linear issue tied to the ticket, then asked for a reply drafted in her tone of voice. She reviewed it and sent it. The division of labor is the point. The agent did the gathering, and the human kept the judgment call and the responsibility for the response. 5. The AWS log query that took a minute instead of twenty. The second example was a customer asking when a specific team in their instance had been deleted. The manual version: log into AWS, open CloudWatch, look up the customer's organization ID, write a SQL query against the logs, read the results, and figure out what happened. The agent had already done it, using IDs it already held for that customer plus the new one provided, and it knew where to look and what the query should be because it had solved this type of issue before. It returned the exact timestamps and drafted a response built on the audit trail from the logs. Owen on the support team sent it as written after confirming the results. The customer got a real answer in a minute or two instead of a note saying someone would look into it. 6. The agent learns from people doing their jobs. Asked how much non-technical team members can shape the agent, Chelikani's answer was that going back and forth with it is functionally writing down every step you took to solve the issue. Ask it to check whether a feature request exists, and it learns that in this class of ticket you check feature requests, and where you check them. The next similar ticket arrives with that already done. Same for knowing which Slack channel covers which project. That's a different adoption model from most AI tooling. The people who can't configure anything are still training the system by working in it. 7. Background agents and the Slack surface. Auto-investigating every incoming ticket is the most common background agent, but teams set up others. Pylon's own runs every time a feature request closes: it summarizes what shipped, checks whether it's actually deployed, decides whether it resolves the customer's original ask, and posts that into the AI thread on the relevant ticket. That closes a gap most support teams live with, where engineering ships the fix and nobody tells the person who took the complaint. The Slack surface carries the same knowledge as the app. In the demo, someone forwarded a customer question into a channel, tagged Pylon, and got an answer drawn from the code base, logs, past tickets, and prior customer interactions. Chelikani followed up by asking whether a knowledge base article should be updated to prevent the confusion. The agent found the relevant doc, drafted the update, and tagged him for review before publishing. 8. What beta customers are reporting. Beta customer results report 70% fewer escalations at one company and a 64.5% faster first response time at another, demonstrating the measurable value of human-in-the-loop agentic workflows.