AI agents in customer support cut cost-per-ticket by 83 to 92 percent against a human-handled interaction. That part of the pitch holds up.
What gets skipped is the real deflection rate. It runs far below the 60 to 80 percent figures shown in most demos.
That gap is exactly what our AI agent readiness checklist is built to catch before you sign anything.
This guide works from named, sourced case studies, not sales decks. Inside:
- The pricing-model math vendors rarely walk you through
- A deflection-rate reconciliation across 10 named case studies
- The escalation mechanics that decide whether savings actually stick
Quick answer: AI agents in customer support cut cost-per-ticket by 83-92%, but the number that decides your budget is your real deflection rate, not a vendor’s demo number.
- AI-resolved tickets cost $0.50-$2.00 versus $6.00-$13.50 for human-handled tickets
- Real deflection averages 41.2% (median), not the 60-80% shown in demos
- Simple intents clear 70%+; nuanced complaints rarely break 25%
- Median payback runs about 5.4 months
- The savings are real only for the ticket types the agent is actually scoped to own
Key Takeaways
- AI-resolved tickets cost $0.50-$2.00 versus $6.00-$13.50 for human-handled tickets, an 83-92% reduction per interaction.
- Real deflection runs far below vendor claims: a median of 41.2%, a top quartile of 58.7%, across enterprise CX programs.
- Klarna’s SEC filing tells a different story than its press release: 80% of chats handled and $39M in 2024 profit contribution, followed by rehiring humans for complex cases in 2025.
- A Stanford/MIT study of 5,179 support agents found AI assistance raised resolution throughput 14% on average, and 34% for the newest agents.
- Pricing model changes your real cost more than the headline price. At 100,000 monthly conversations and a 60% resolution rate, per-outcome pricing runs roughly $59,400/month against $200,000/month for per-conversation pricing.
- Median payback across deployments is about 5.4 months. Enterprise rollouts can break even in under a month.
What “AI Agent” Actually Means in a Customer Support Context
An AI agent reasons over a request, decides what needs to happen, and takes the action itself.
That could mean a refund, an order update, or an escalation, completed without a human scripting every step first.
That’s different from a keyword chatbot matching phrases to scripts, or an IVR that only routes calls.
Our breakdown of AI agent vs. chatbot vs. virtual assistant covers the distinction in full. Every cost figure in this guide only applies to real agents.
Why This Distinction Changes Your Cost Model
The difference shows up directly in what a ticket costs to close:
- A scripted chatbot only closes a ticket when the phrasing matches a pre-built script
- A real agent reasons over account and order data and finishes multi-step actions itself
- Only the second category produces the savings covered in this guide
A bot that guesses wrong just relabels the ticket “the bot didn’t understand me.” That costs more to resolve than if a human had taken it first.
Confusing the two is the most common reason a deployment underperforms. Our comparison of a real AI agent vs. no-code automation covers how to tell them apart before you buy.
Containment vs. Deflection: A Distinction Vendors Blur on Purpose
Deflection means the customer never reached a human. Containment means the AI actually resolved the issue.
- Deflected: the customer left the chat, resolved or not
- Contained: the issue is genuinely closed
- A ticket can be deflected and still unresolved if the customer simply gave up
Ask any vendor which number their case study reports. If they can’t answer cleanly, treat their headline figure as a deflection number until proven otherwise.
The Real Cost Numbers: AI Agents vs. Human Agents Per Ticket
Cost-per-ticket is the cleanest number in this category. It doesn’t depend on how a vendor defines “resolved.”
Cost Per Resolution by Channel and Intent
| Channel / Intent Type | Human Agent Cost | AI Agent Cost | Reduction |
|---|---|---|---|
| Chat ticket, general queue | $6.00-$13.50 | $0.50-$2.00 | 83-92% |
| Voice call | $7.40 average | $1.18 average | ~84% |
| Simple, high-volume intent | $6.00-$7.00 | $0.50-$0.99 | 85-92% |
| Complex, multi-step case | $10.00-$13.50+ | Often still human-routed, or lower net savings once review time is counted | Varies |
Intercom’s Fin AI Agent prices at $0.99 per resolution, an outcome-based model.
Use that as a sanity check on any vendor you evaluate. Do you pay for an attempt, or for an outcome?
Why the Pricing Model Matters More Than the Headline Price
Two vendors quoting $0.99 and $2.00 per resolution sound comparable. They aren’t, once you look at how each one actually bills.
At 100,000 monthly conversations and a 60% resolution rate, here’s the real math:
| Pricing Model | Monthly Cost | Effective Cost Per Actual Resolution |
|---|---|---|
| Per-outcome ($0.99, bills only on resolution) | ~$59,400 | $0.99 |
| Per-conversation ($2.00, bills on every attempt) | ~$200,000 | $3.33 |
| Per-seat (flat platform fee, variable usage) | $38,250+ | Depends entirely on volume |
A per-conversation vendor at a lower headline price can cost more than three times as much per real resolution. Every unresolved attempt gets billed the same as a successful one.
Read the pricing page before the case study. Our overview of core AI solutions for business covers how we scope this exact trade-off before recommending a platform to a client.
The ROI Timeline and Payback Period by Company Size
Cost per ticket shows the unit economics. Payback period shows when the deployment clears its own cost.
That number depends heavily on volume.
Payback Period by Deployment Size
| Company Size | Ticket Volume | Typical Payback Period |
|---|---|---|
| Enterprise (500+ employees) | 50,000+ tickets/month | As fast as 27 days |
| Mid-market (50-500 employees) | ~5,000 tickets/month | About 2 months |
| Small business (1-50 employees) | ~500 tickets/month | 5-7 months |
| Median across all sizes | Blended | ~5.4 months |
Higher volume compounds the same resolution rate into savings faster. A small team isn’t doing anything wrong at 5 to 7 months. There simply isn’t enough ticket volume yet to shorten the curve.
Starting narrow and proving the number before scaling is the fastest path through that curve. Our phased AI rollout framework covers exactly why.
The Compounding ROI Curve
Reported ROI trajectories average $3.50 returned per $1 spent, with leading organizations reaching 8x. The curve compounds across three years:
- Year 1: ~41% ROI
- Year 2: ~87% ROI
- Year 3: 124%+ ROI
That compounding needs a clean baseline and defined escalation rules from day one. Our guide to measuring AI agent productivity gains covers how to build that baseline before you launch.
10 Real AI Agent Customer Support Case Studies
Case studies are the most misused evidence in this category. A single best-case number gets repeated forever, stripped of the context that produced it.
Here are ten, with sources.
| Company | Industry | Result |
|---|---|---|
| Klarna | Fintech | 80% of chats handled, $39M profit contribution in 2024 (per SEC filing) |
| NIB Health Insurance | Health insurance | $22M saved, 60% lower service cost, 15% fewer calls |
| Medtronic | Medical devices | $6M saved in one year via better call routing |
| Unity | Gaming / software | 8,000 fewer tickets, $1.3M in savings |
| ReserveBar | E-commerce | 850 agent hours saved, 93% CSAT maintained |
| Good Eggs | Grocery delivery | 40% lower handle time via AI copilot, no customer-facing bot |
| Rocket Money | Fintech | $1M annual ROI, 68% resolution rate |
| Anthropic | Software / AI | 1,700+ hours saved in the first month, 58% resolution rate |
| Dollar Shave Club | E-commerce | 6x containment growth, 3x ticket coverage, targeting 70% containment |
| Intercom Fin customers | Multiple (SaaS, fintech) | Real resolution rates of 42-70%, well below marketing headlines |
The Klarna Timeline Most Articles Skip
Most coverage of Klarna quotes a single February 2024 number: 2.3 million chats in the first month, 67% automation, work equal to 700 agents.
Klarna’s 2025 Form 20-F filing with the SEC tells a fuller story:
- 80% of customer-service chats handled across all of 2024
- $39 million contributed to profit
- Average resolution time cut from 12 minutes to 2
- Repeat inquiries down 25%
By late 2025, the AI was doing the work of roughly 853 agents. Klarna also began rehiring humans for complex cases that May, after finding full automation had hurt quality on harder tickets.
That’s not a failure story. It’s the clearest public example of automation compounding fast on simple volume, then needing a human safety net everywhere else.
Rehiring after a rollout is a course correction, not a crisis. Our AI change management checklist plans for that adjustment before launch, not a year in.
Why Intercom’s Real Numbers Matter More Than Its Marketing Page
Intercom’s own case studies report Fin resolution rates of 42% at Linktree and 50% at Robin. Both sit well under the 60 to 80% range common in general marketing.
Fundrise crossed 50% within three months. Sharesies reached 70% after 12 weeks of active tuning, the highest publicly reported figure and clearly the exception.
Low resolution rates in the first few weeks usually mean the agent is guessing instead of escalating. That’s the same failure mode covered in our guide to why AI agents hallucinate in production.
That gap between “possible with tuning” and “true out of the box” is exactly why the next section exists.
The Deflection Rate Reality Check
“Deflection rate” is the most gamed metric in this category. Understanding why explains most of the gap between what you’re pitched and what you’ll actually book.
1. Claimed Deflection vs. Real Deflection, by Intent Type
| Intent Type | Vendor Demo Range | Real-World Reported Range |
|---|---|---|
| Password reset, order status, simple FAQ | Often quoted 80%+ | 70%+ (usually accurate) |
| General mixed ticket queue | 60-80% | Median 41.2%, top quartile 58.7% |
| Nuanced complaints, billing disputes | Rarely disclosed | Under 25% |
| Regulated industries (finance, healthcare) | Same 60-80% figures used | Often 25-45% in year one |
Whenever a specific, named number is published, it lands below the round figures used in marketing. Ask a vendor to break their blended number out by intent before you trust it.
2. The Deflection Reality Ladder
Use this to translate any number a vendor gives you into what it means for your budget:
- Vendor demo number: best-case scenario, cherry-picked intents, often the vendor’s single best customer.
- Published case-study number: real, but usually from a customer with clean data and months of tuning behind it.
- Your queue’s blended number: typically 15-35 points lower than the case-study number once your full ticket mix hits the agent.
- Net cost-adjusted number: your blended rate, minus the time your team spends correcting what the agent got wrong. This is the only number for a board deck.
Most vendor conversations stop at rung one. Internal planning should start at rung four.
3. Why “Deflected” Doesn’t Always Mean “Resolved”
A ticket can be logged as deflected for reasons that have nothing to do with getting helped:
- The customer gave up
- They rephrased the same question three times
- They abandoned the chat, frustrated
None of that shows up in a deflection percentage. It shows up in your recontact rate and your CSAT score.
Ask for the 7-day recontact rate on tickets marked resolved before trusting the headline number attached to any platform. It’s exactly the number our team pulls before recommending one to a client.
The Metric That Predicts ROI Better Than Deflection Rate
Deflection rate measures activity. It doesn’t measure whether that activity actually removed work from a human’s plate, the only thing that shows up on a P&L.
The Net Minutes Removed Formula
Net Minutes Removed = (Tickets Fully Resolved by the Agent × Average Human Handle Time) − (Human Review and Correction Time on AI-Touched Tickets)
If the agent closes a ticket in two minutes but a human then spends eight minutes fixing it, the work moved. It didn’t disappear.
Correction time climbs fastest when the agent’s grounding is weak. Our comparison of RAG vs. fine-tuning for AI agents covers how to close that gap.
A Worked Example
| Line Item | Value |
|---|---|
| Tickets fully resolved by the agent per month | 4,000 |
| Average human handle time for the same ticket type | 6 minutes |
| Gross minutes removed | 24,000 minutes |
| Tickets “resolved” but a human still had to fix | 900 |
| Average correction time per touched ticket | 5 minutes |
| Minutes re-added by correction work | 4,500 minutes |
| Net minutes removed | 19,500 minutes ≈ 325 hours/month |
Priced at your team’s fully loaded hourly cost, that 325 hours is the real savings figure.
Our guide to measuring AI agent productivity gains applies the same discipline across a full AI rollout, not just support.
Why the Same AI Agent Performs Differently at Every Company
Two companies can run the identical platform and land 20+ points apart on deflection rate.
The gap is rarely the model. It’s almost always what the agent was allowed to touch.
1. The Ticket-Tiering Model
Sort every ticket type into one of three tiers before configuring anything.
| Tier | Examples | Automation Level |
|---|---|---|
| Safe to Automate | Password resets, order status, shipping ETAs | Full autonomy, no review needed |
| Safe to Assist | Refund evaluation, account changes, billing questions | AI drafts, human approves |
| Never Automate Without a Human Gate | Legal or PR risk, financial advice, vulnerable customers | Human owns it; AI summarizes context only |
Companies with the strongest published numbers, Sharesies’ 70% among them, scoped the agent to a narrow tier first. They expanded only after it earned trust in production.
2. Knowledge Base Quality Is the Hidden Variable
An agent can only resolve what it can confidently ground in accurate source material.
Outdated policy docs and conflicting FAQ answers cap the resolution rate, regardless of which model powers the agent.
If your team has never audited what the agent actually pulls answers from, that’s the first fix. Not a model upgrade.
3. Hallucination Risk Rises Where Grounding Is Weakest
When an agent can’t find a confident answer, it doesn’t always say so. Sometimes it generates a plausible, wrong one instead.
That costs more to clean up than an honest escalation would have.
Our guide to why AI agents hallucinate in production covers what ungrounded answers have already cost real companies. Read it before expanding past Tier 1.
The Handoff Problem: Where Cost Savings Quietly Disappear
Even a well-scoped agent creates a new cost center if the escalation to a human loses context.
This is the single most common complaint about AI-first support.
Why a Clean Escalation Matters More Than the Bot’s Answer
A customer who explains an issue for five minutes, gets escalated, then repeats it all to a human hasn’t experienced automation.
They’ve experienced a slower queue with an extra step bolted on.
That pattern shows up directly in CSAT. Tickets that touch the AI first and then escalate score lower, even with an identical resolution.
The Handoff Context Checklist
Confirm the escalation package includes all of the following, every time:
- The customer’s original question, in their own words
- Every step the agent already attempted
- Any account or order data the agent already pulled
- A stated reason for the escalation, not a generic “transferring you now”
- The customer’s frustration signal, if your platform tracks sentiment
None of this needs a bigger model. It needs the agent’s memory layer to actually persist what it already knows.
Our guide to how AI agents handle memory and context covers that mechanism in technical depth.
A lost handoff is what turns a savings story into a public complaint. It’s exactly the risk we design against for clients.
Agent-Assist vs. Full Autonomy: Which Pays Back Faster
Not every deployment needs to face the customer directly to produce savings. The evidence points toward agent-assist as the faster, lower-risk starting point for most teams.
The Evidence for Starting With Agent-Assist
- A Stanford/MIT study of 5,179 support agents found a 14% average productivity gain from AI assistance, rising to 34% for the newest agents, without any customer touching the AI directly.
- Good Eggs cut handle time 40% with an AI copilot that only human agents used.
- Anthropic saved 1,700+ hours in its first month at a 58% resolution rate.
- Rocket Money reported $1M in annual ROI at a 68% resolution rate.
When Full Autonomy Is Worth the Risk
Full autonomy earns its ceiling once Tier 1 volume is proven reliable and your knowledge base has been audited. It rarely earns that ceiling on day one.
Voice and chat carry different risk profiles here. A misrouted chat is annoying. A misrouted phone call is harder to correct mid-conversation.
Our comparison of voice AI agents vs. chat AI agents covers which channel fits which use case.
Some teams eventually split the workload across specialized agents once volume outgrows one general-purpose system. Our guide to single-agent vs. multi-agent systems covers when that complexity actually pays off.
Regulated Industries: Why Lower Deflection Is Actually Fine
Finance, healthcare, and insurance teams consistently report lower deflection than retail or SaaS. Treating that as a failure is the wrong read on the data.
The Real Numbers in Regulated Support Queues
Realistic early-stage deflection in regulated industries runs 25 to 45%, below the 60 to 80% figures common in general marketing.
More of the ticket volume legitimately belongs in the Safe to Assist or Never Automate tiers above.
NIB Health’s $22 million in savings still proves regulated support can produce a strong number. It’s just built from a narrower base of automatable tickets, not a retail company’s blended rate.
What Matters More Than the Deflection Number Here
For regulated teams, the sharper questions aren’t about deflection at all:
- Where is customer data stored and processed
- Can every AI answer cite its source
- What does the audit trail look like after a human reviews a decision
Our comparison of on-premise vs. cloud-hosted AI agents and our guide to giving an AI agent access to customer data safely both cover that groundwork.
Both matter more to a compliance review than a headline deflection number ever will, and it’s the first thing we scope on any regulated-industry engagement.
Build, Buy, or Blend: How to Actually Deploy This
Once the tiering and ROI math make sense, the remaining decision is who builds and maintains the system.
What You Gain and Lose With Off-the-Shelf Platforms
An off-the-shelf platform gets you to a working agent faster.
You get less control over exactly how the knowledge base is grounded and how escalations are scoped. Our comparison of a real AI agent vs. no-code automation covers that trade-off from a technical angle.
When In-House or an Agency Pays Off Faster
A custom build gives you full control, at the cost of owning the data plumbing behind it.
Most support teams under a few hundred agents reach a measurable deployment faster with a partner who has already built the tiering and escalation logic once.
Our breakdown of AI agency vs. in-house AI team covers that staffing and cost trade-off directly.
If your last vendor call only produced a blended deflection number and no tiering plan, that’s the exact gap GVM Technologies AI scopes before proposing a platform.
Rolling this out in stages protects every number in this guide. Our phased AI rollout framework covers why starting narrow beats a full-queue launch.
How to Calculate Your Own ROI Before You Sign a Contract
Run this before evaluating a single vendor, so you have a number to hold theirs against. It’s the same math we walk through with clients considering an AI automation agency versus building in-house.
The Formula
AI Agent Customer Support ROI (%) = [(Net Minutes Removed × Fully Loaded Hourly Support Cost) − Total Annual Cost of the Agent] ÷ Total Annual Cost of the Agent × 100
Total annual cost of the agent has four line items, not one:
- Platform fee
- Integration work
- Knowledge base cleanup
- Ongoing review time from the Handoff Context Checklist above
Skip that last line, and the number won’t survive the first hard question in the room.
A Worked Example
A 30-person support team resolves 12,000 tickets a month at $28/hour fully loaded cost.
Model an agent that fully resolves 35% of volume with a 12% review-touch rate. That’s a far more defensible business case than “the vendor says 70% deflection.”
Build the version specific to your own queue before you buy anything. Our AI agent readiness checklist scores whether your data and ownership are ready to produce that number honestly.
Common Mistakes That Turn Real Savings Into Real Problems
- Buying against a blended deflection number instead of a per-intent breakdown. The 41.2% median hides both 70%+ intents and sub-25% intents in one figure.
- Skipping the Ticket-Tiering Model. Letting the agent attempt everything is the single biggest cause of a disappointing first quarter.
- Never auditing the knowledge base. A resolution rate can’t outrun the accuracy of its source material.
- Treating deflection as proof of success without checking recontact rate or CSAT. A customer who gave up isn’t a customer who got helped.
- Losing context on every escalation. This is the fastest way to turn a cost-saving deployment into a customer-experience liability.
- Skipping the pilot in regulated industries because the ROI math looked good elsewhere. Compliance requirements, including where the agent’s data actually lives, don’t scale down just because a case study looked impressive.
- Comparing headline per-resolution prices without checking the billing model. A lower price per attempt can cost more per real resolution.
- Never re-measuring after the first 90 days. Klarna’s own numbers changed materially between month one and month twelve, and yours will too.
Every mistake on this list shows up in a first vendor call. GVM Technologies AI runs through all eight before recommending a platform, not after.
FAQs About AI Agents in Customer Support
1. How much do AI agents actually save on customer support costs?
AI-resolved tickets typically cost $0.50-$2.00 versus $6.00-$13.50 for a human-handled ticket, an 83-92% reduction per interaction. Total savings depend on the share of volume the agent genuinely resolves.
2. What’s a realistic AI agent deflection rate, not a vendor’s demo number?
Industry-wide, the median sits around 41.2%, with a top quartile of 58.7%. Simple intents often clear 70%+. Nuanced complaints rarely break 25%.
3. How long until an AI customer support agent pays for itself?
Median payback is around 5.4 months. Enterprise deployments with high volume can break even in under a month; smaller teams often take five to seven months.
4. Is Klarna’s AI customer service case study still accurate?
Directionally, yes, but the full picture matters. Klarna’s 2025 SEC filing reports 80% of chats handled and $39M in profit contribution for 2024.
It also rehired humans for complex cases in 2025, after finding full automation had hurt quality on harder tickets.
5. Should we start with a customer-facing agent or an agent-assist tool?
For most teams, agent-assist is the faster, lower-risk starting point. The Stanford/MIT study of 5,179 agents found a 14% productivity gain from assistance alone, rising to 34% for newer agents.
6. Do AI agents work for regulated industries like finance or healthcare?
Yes, but expect 25-45% deflection rather than 60-80%, since more ticket types legitimately require a human. NIB Health still reported $22 million in savings from a narrower automatable base.
7. Why do our numbers look worse than the case studies we read about?
Almost always because the ticket-tiering step got skipped. Sharesies’ 70% resolution rate came from scoping the agent narrowly first, then expanding gradually.
8. What’s the biggest hidden cost in AI customer support ROI math?
Human review and correction time on tickets the agent “resolved” but got partially wrong. The Net Minutes Removed formula above is built specifically to catch that line item.
9. Can a small business realistically deploy an AI customer support agent?
Yes. The tiering and ROI math scale down cleanly. Payback simply runs longer, five to seven months, because lower ticket volume takes longer to accumulate into a measurable dollar figure.
10. What happens when an AI agent escalates a ticket badly?
The customer repeats information already given, the human starts from zero context, and CSAT on that ticket drops below what a human-only interaction would have scored. That’s a design issue, not a model-quality issue.
The Bottom Line
AI agents in customer support produce real, measurable savings: an 83-92% cost reduction per resolved ticket, payback in months, and ten named case studies with real dollar figures behind them.
None of that means a vendor’s headline deflection number is the number you’ll actually get. Before you configure anything:
- Score your own readiness first
- Sort ticket volume into safe-to-automate, safe-to-assist, and never-automate
- Track Net Minutes Removed, not deflection rate, as your real success metric
- Fix the knowledge base before blaming the model
- Build the handoff so a customer never repeats themselves
- Re-measure at 90 days; the numbers move as fast as Klarna’s did
Ready to Find Your Team’s Real Number, Not a Vendor’s Demo Number?
Most companies evaluating AI agents for customer support get a deflection percentage from a sales call and back into a budget from there. That’s backward.
GVM Technologies AI scopes the ticket-tiering, knowledge base audit, and ROI baseline before a single line of code ships:
- AI agent deployments for support and operations
- Workflow automation and CRM integration
- Industry-specific AI applications
See that discipline in practice in our customer outreach case study, where the baseline was set before the agent ever went live.
Talk to our team about what your ticket mix would actually support, before you buy against someone else’s case study.
