On this page · 12 sections
- The economics: why the pattern beats the model
- Pattern 1: Autonomous resolution
- Pattern 2: Agent assist, the copilot
- Pattern 3: Intelligent triage and routing
- Pattern 4: Post-contact automation
- Pattern 5: Proactive and self-service deflection
- Where cost-only automation breaks: the Klarna lesson
- What it means for India
- How to choose and sequence the patterns
- FAQ
- How eCorpIT can help
- References
Summary. AI is changing the economics of customer support, but the saving comes from the pattern you deploy, not the model you buy. Gartner expects conversational AI to cut contact-centre agent labour costs by $80 billion in 2026, and it pegs an agent-assisted contact at about $13.50 against $0.50 to $2.00 for an AI-handled ticket. The market is moving with the money: the global conversational AI market is forecast to grow from $14.79 billion in 2025 to $17.97 billion in 2026, and Salesforce agreed in June 2026 to buy the AI support platform Fin, formerly Intercom, for $3.6 billion. Klarna's OpenAI-built assistant showed both the upside and the limit, handling two-thirds of chats and the work of 700 agents within a month, then forcing a public rethink in 2025 when quality slipped. This guide sets out five conversational-agent patterns that cut support costs, what each automates, the per-resolution economics behind them, and the Klarna lesson on where cost-only automation breaks. The pattern that wins is rarely full automation. It is the right mix of machine speed and human judgement, measured on quality as well as cost.
Customer experience leaders are under two pressures at once: budgets that expect AI-sized savings, and customers who punish bad automation. The five patterns below are ordered from the most autonomous to the most human-supported, with the economics and the trade-offs for each, so a support or operations head can pick the ones that fit their volume, their risk, and their team.
The economics: why the pattern beats the model
Before the patterns, the money. A human-handled contact and an AI-handled one differ by an order of magnitude. Gartner puts the average agent-assisted interaction near $13.50 and an AI-resolved ticket between $0.50 and $2.00 once infrastructure, licensing, and configuration are counted. That gap is why every major platform now charges per outcome rather than per seat alone.
The pricing tells you what each vendor is willing to stand behind. Fin, formerly Intercom, charges $0.99 per resolution. Zendesk's AI agents run about $1.00 to $1.50 per automated resolution. Salesforce Agentforce charges $2.00 per conversation, billed whether or not the issue is resolved and even when it escalates to a human. At 100,000 monthly resolutions, the spread between a $0.99 and a $2.00 model is over $1.2 million a year, which is why the billing unit, resolution versus conversation, matters as much as the rate.
| Platform | Pricing model | Indicative cost |
|---|---|---|
| Fin (formerly Intercom) | Per resolution | $0.99 per resolved issue |
| Zendesk AI agents | Per automated resolution | $1.00 to $1.50 |
| Salesforce Agentforce | Per conversation | $2.00, resolved or not |
| AI-handled ticket (blended) | Cost per contact | $0.50 to $2.00 (Gartner) |
| Human agent-assisted | Cost per contact | About $13.50 (Gartner) |
The lesson for a CX leader is to buy on resolved outcomes, not interactions, and to model the cost at your real volume before signing. A pattern that deflects cleanly at $0.99 a resolution is a saving. A pattern that charges per conversation and still escalates half its cases can cost more than the agents it was meant to replace.
Pattern 1: Autonomous resolution
The most direct saving is letting AI resolve common, repetitive queries end to end, with no human in the loop. Order status, password resets, refund rules, store hours, and policy questions are the textbook cases, because the answer is knowable and the action is bounded.
Klarna's assistant is the reference point. Within a month of launch it handled two-thirds of the company's customer service chats, around 2.3 million conversations, doing the work of 700 full-time agents, and it resolved issues in under two minutes against eleven for a human. Klarna estimated a $40 million profit improvement in its first year.
The economics are strongest here because a deflected ticket never reaches a person. The risk is also highest, because a confident wrong answer on a refund or a fraud claim does real damage. The discipline that makes this pattern safe is scope: automate the intents you can verify against a source of truth, and route everything else to a human. This is also where AI agent governance matters most, because an autonomous support agent is an autonomous agent with access to customer data.
Pattern 2: Agent assist, the copilot
The lowest-risk pattern keeps the human in charge and makes them faster. An agent-assist copilot listens to the live conversation, surfaces the right knowledge-base article, drafts a reply, and summarises account history, so the agent spends less time searching and more time deciding.
The gains are measurable. McKinsey reports that a European media and telecommunications company cut average handle time for finding relevant knowledge by 65% with a gen-AI copilot. Because the human still owns the outcome, quality holds up while capacity rises, which makes this the safest first step for teams handling sensitive or complex contacts. It also lifts new-hire ramp speed, because the copilot encodes the answers a veteran would already know.
Agent assist rarely deflects a ticket outright, so its saving shows up as higher throughput per agent and shorter calls rather than headcount cut. For most enterprises that is the more durable win, because it improves service while reducing cost, instead of trading one for the other.
Pattern 3: Intelligent triage and routing
A large share of support cost is spent before anyone solves anything: identifying the customer, working out what they want, and getting them to the right queue. Conversational AI handles that front end. It detects intent from the first message, collects the account details and the reason for contact, checks entitlement, and routes the case to the agent or workflow best suited to it.
Gartner notes that even partial containment, automating the identification of a customer and their reason for calling, can cut up to a third of the time a human agent spends on an interaction. The case may still reach a person, but it arrives qualified, prioritised, and pre-filled, so the expensive human minutes go to resolution rather than discovery. Done well, triage also reduces transfers and repeat contacts, both of which are pure cost.
This pattern is low risk because it does not decide the outcome, only the path. It pairs naturally with the first two, feeding simple intents to autonomous resolution and complex ones to a copilot-equipped agent.
Pattern 4: Post-contact automation
The work does not stop when the call ends. Agents spend minutes after each contact writing summaries, coding the disposition, and updating the record, and quality teams review only a small sample of contacts by hand. Conversational AI automates that after-call work. It generates the summary, tags the case, drafts the follow-up, and can score every interaction for quality rather than the usual 1 to 2% a human team can review.
The saving is twofold. After-call work falls, which returns agent time directly, and automated quality scoring catches problems across 100% of contacts instead of a sample, which improves the whole operation rather than one agent at a time. Because the customer has already left the conversation, the risk is contained: a wrong summary is an internal error to correct, not a bad answer sent to a customer.
This is one of the easiest patterns to justify, because it touches no customer directly and the time saved is easy to count: minutes of after-call work multiplied by contact volume.
Pattern 5: Proactive and self-service deflection
The cheapest ticket is the one never opened. The final pattern pushes answers to where the customer already is, an in-product help widget, a search box, a status page, or a proactive message, so the issue is resolved before it becomes a contact. A clear answer surfaced at the moment of need deflects the contact at its source, ahead of any per-resolution charge.
Proactive support extends this. When a system knows an order is delayed or a payment failed, it can reach out with the fix before the customer calls, which both lowers contact volume and improves how the customer feels about the brand. Self-service has the lowest marginal cost of any pattern, because a well-written help article answers thousands of customers for the price of writing it once, now kept current by AI.
The limit is honesty. Self-service that hides the path to a human, or a proactive message that misfires, erodes trust quickly. The pattern works when it makes reaching help easier, not when it is used to wall customers off from people.
| Pattern | What it automates | Where the saving comes from |
|---|---|---|
| 1. Autonomous resolution | End-to-end answers to common intents | Tickets deflected before they reach a person |
| 2. Agent assist | Knowledge lookup, drafting, summaries | Shorter handle time, faster agent ramp |
| 3. Triage and routing | Intent, identity, prioritisation | Less time on discovery and transfers |
| 4. Post-contact automation | Summaries, tagging, quality scoring | Lower after-call work, full QA coverage |
| 5. Proactive and self-service | Answers before a ticket is opened | Contacts avoided at the source |
Where cost-only automation breaks: the Klarna lesson
The cautionary tale comes from the same company that proved the upside. After going AI-first, Klarna reversed course in 2025 and began hiring human agents again. In a May 2025 Bloomberg interview, chief executive Sebastian Siemiatkowski said that when cost becomes "a too predominant evaluation factor," what you end up with is "lower quality," and that "investing in the quality of human support is the way of the future for us." Klarna kept its AI but rebuilt the human option around it.
Gartner has made the same point from the other direction, challenging the assumption that AI support is automatically cheaper as the cost of running generative models in customer service rises. The takeaway for CX leaders is not to avoid automation. It is to optimise for resolution quality and customer outcome, with cost as a constraint rather than the only goal. The patterns that endure keep a clear, easy path to a capable human for the cases that need one, especially disputes, fraud, hardship, and anything with a vulnerable customer.
| Reported outcome | Result | Source |
|---|---|---|
| Chats handled by AI (Klarna, month one) | Two-thirds of volume | Klarna |
| Work equivalent (Klarna) | About 700 agents | OpenAI, Klarna |
| Handle-time for knowledge lookup | Down 65% | McKinsey |
| Resolution time (Klarna) | Under 2 min vs 11 | Klarna |
| Contact-centre labour cost, 2026 | Down $80 billion | Gartner |
What it means for India
India sits at the centre of this shift, because it manages roughly 40% of the global outsourced customer experience market. The BPO sector employs around 1.65 million people in voice, chat, and back-office support, and hiring has slowed sharply as the routine work these teams handled becomes the work AI deflects first. That is the risk side.
The opportunity side is a move up the value chain. Leading Indian providers are shifting into AI-adjacent work: building and tuning the conversational agents, annotating and governing the data behind them, running quality and analytics on top of them, and handling the complex, judgement-heavy contacts AI routes to people. A NASSCOM and Deloitte estimate puts the AI talent India needs at about 1.25 million by 2027, up from roughly 600,000 to 650,000, which is the scale of the reskilling involved.
For Indian CX operations, two rules follow. First, customer conversations carry personal data, so any deployment falls under the Digital Personal Data Protection Rules notified on 13 November 2025, with their consent, security, and 72-hour breach-notification duties. Second, the providers that win will sell outcomes and quality, not seat-hours, because the seat-hour is exactly what automation removes.
How to choose and sequence the patterns
Most teams should not start with full autonomous resolution. The lower-risk sequence is to deploy agent assist and post-contact automation first, because they cut cost without putting AI between the customer and the answer, then add triage, then automate resolution only for the specific intents you can verify against a source of truth. Proactive and self-service deflection runs alongside throughout.
Measure each pattern on two axes, never one. Track cost per contact and deflection rate, but track customer satisfaction, resolution quality, and escalation rate next to them, so a saving that quietly damages service shows up before customers leave. The same operating discipline that governs any enterprise generative AI strategy applies here: clear ownership, measured outcomes, and a human accountable for the result.
FAQ
How eCorpIT can help
eCorpIT is a CMMI Level 5 technology organisation in Gurugram whose senior engineering teams design and integrate conversational-agent systems into existing helpdesks and contact centres. We help CX and operations leaders pick the patterns that fit their volume and risk, wire them to a verified source of truth, keep a clean human escalation path, and measure quality alongside cost. You can read more about eCorpIT and its director Manu Shukla. To scope an AI customer-experience project, contact our team.
References
_Last updated: 21 June 2026._
