Gartner Says AI Support Is About to Get Expensive, So What Should CX Leaders Build Instead?

Most probably, you have already heard the news, but in case you might have taken a sabbatical, Gartner predicts that the cost of generative AI will exceed $3 by 2030. 

Harsh news to take in for the industry that put all of its bets on customer support AI as a cost-cutting measure. But does it actually mean that AI customer service will be more expensive than the regular human one? And is there a way to still benefit from AI in support without ballooning the company budget?

What did Gartner find about the cost of generative AI?

What Garther has found refers only to the cost per resolution, meaning the cost would cover the full AI-driven ticket processing from start to finish. Given that current prices for offshore agents range from $4.50 to $6.20+ per ticket, AI might not be a favorable choice for customer support in the near future.

This wasn’t the only finding of the article, though. Also interesting were the reasons behind the increase in AI prices. Here are just a few factors that seem to influence the climb:

1. As most LLM providers are currently subsidizing their services by up to 90% as a market-share strategy, we can say that the upcoming cost increase is just a regular vendor pricing normalization.
   

2. Running LLMs at enterprise scale takes serious computing power, and the need for it keeps growing – especially for systems that handle complex reasoning, long inputs, or a mix of text, images, and audio. As such, the new price simply accounts for these AI infrastructure requirements.
   

3. To effectively run AI operations, businesses also require orchestration layers, governance controls, RAG pipelines, compliance tooling, and monitoring systems. And each of these functions inflates the cost of AI–facilitated resolutions.
   

4. The newest models use three to ten times as many tokens per interaction as older ones. And as the tasks you ask them to handle get more complex, they burn through even more tokens, pushing the total resolution bill higher.

Overall, I would say that the key driver behind the “sudden” increase in the cost of generative AI is a deeper understanding of the technology itself and of what it takes to run it. 

The lying metric: Cost-per-hour vs. cost-per-resolution

Since we are talking about the cost of customer support, I want to dig a bit deeper into why businesses felt that AI was helping them to cut costs. For years, the metric that drove support decisions was cost per hour, and on paper, AI looked like it had demolished it. But the cost per hour was telling a flattering story long before AI entered the scene.

Procurement teams default to hourly rates because they're easy to compare. And outsourcing rates usually fall neatly into geographic bands:

• roughly $9–$17/hr for Asia-based providers

• $12–$22/hr in Eastern Europe

• $25–$50/hr in the US and Canada

This has made cost-cutting look like a simple matter of picking the cheapest destination. But paying for an hour of agent time doesn’t tell you what you're getting. And what businesses expect to get are closed tickets (and resolved customer issues). An hour can be spent thoroughly resolving one issue, or it can result in a repeat contact, trigger an escalation, or close a ticket incorrectly – each of which inflates your support costs down the line. 

Instead, leaders should focus on the cost per resolution: total support costs divided by the issues you actually resolved. In this case, “resolved” means a ticket that got closed without repeated contacts on the same issue within a set window (usually 7–30 days) and no escalations after the last interaction.

Example: A center spending $250,000 a month on 50,000 contacts looks like $5.00 a resolution. But if weak first-contact resolution means only 40,000 of those were actually solved, the real figure is $6.25, which is already 25% higher.

Cost per resolution is probably the most important metric teams need to track to properly budget their support operations.

Where the “hidden” costs actually “hide.”

When talking about the hidden costs of human-led customer support, I usually highlight 4 key areas:

• Repeat contacts and rework

Every unresolved ticket tends to come back one to three more times. SQM Group has found that each 1% gain in first-contact resolution (FCR) knocks operating costs down by about 1%, which for a typical midsize center is around $286,000 a year per point.

• Escalations

A ticket escalated to Tier 2 costs 3–5 times more than one resolved at Tier 1, and issues that reach Tier 3 specialist support routinely land in the $80–$100+ range — compared to a Tier 1 average of $15.56 in North America, making the multiplier 5–7× for a full Tier 1→Tier 3 escalation path. Of course, you wouldn't budget for all of this when looking at the Tier 1 rate. 

• Agent churn and ramp drag

The average US contact center loses 30–45% of its people every year, and replacing one agent costs $10,000–$20,000 when you factor in hiring, training, and lost productivity. New hires also need time to hit full speed, handling tickets more slowly and generating more repeat contacts. High-churn vendors pass this cost to you continuously

• Customer churn

It was found that any given service interaction is 4 times more likely to create disloyalty than loyalty among the customers, with repeat contacts, transfers, and channel switching as the primary drivers. And as we all know, acquiring new customers always costs more than keeping the ones you have. So when a frustrated customer leaves after a poor experience, the cheap hourly rate didn't save you anything – it just moved the cost over to your acquisition budget.

The real math behind cheap support

Let’s imagine a scenario where we have agents A and B handling the same volume, but at different hourly rates and with different numbers of repeat contacts. 

Metric

Agent A

Agent B

Hourly rate

$10/hr

$18/hr

First-contact resolution

60%

85%

Tickets handled per hour

8

8

Cost per resolved ticket

~$2.08

~$2.65

Repeat contacts per 100 tickets

40

15

It’s true that the work of agent A costs less, but they also generate more repeated contacts than agent B. And every one of those is basically another ticket being solved for $2.08. So, even 30 extra contacts add about $62 per 100 original tickets, practically erasing the original cost savings provided by agent A. 

This is the same trap that made AI look like a cost-cutter. Measured by marginal cost per query, automation seems like THE solution for bloated support budgets. Yet, if we consider the actual ticket resolution with the poorly deployed AI, we would need to factor in the same escalations and repeat contacts as we would for the human team. 

That’s exactly why I believe optimizing for money (almost) never works. What businesses need to start doing is optimizing for actual better performance and outcomes, and only then will there be a chance to cut operational costs.

So what does this mean for future businesses?

I’ve talked about it before, and I won’t get tired of mentioning it again: AI in customer service is a great tool for improving agent productivity, boosting FRT, and facilitating faster case resolutions. But only as long as it has humans to keep it in check.

Take this as an example. On July 19, 2024, a botched CrowdStrike update set off one of the biggest IT outages we've seen. According to Cirium, of the 411,009 flights scheduled worldwide over the next 72 hours, around 16,896 were canceled, more than double the previous week. Delta alone grounded 1,326 flights.

Now picture being one of those passengers, opening the airline's chatbot to rebook. The chatbot can't help you because the automated rebooking it relies on runs on the same infrastructure that just collapsed. And you know what would have helped sort the situation out in minutes? A human agent, working from a different system with override rights. And this is, basically, the whole problem of going full AI in one example.

Where AI has its place, and where it doesn't 

AI won’t be of help in situations involving system failures, policy exceptions, fraud reviews, or genuinely upset customers looking to let their anger out (or maybe that’s where AI could help save poor agents?) Anyways, this doesn’t mean AI is the enemy of good support.

The experience of deploying AI for EverHelp clients has shown me that the first step to successful implementation is understanding what the AI could realistically handle for your business. Here’s a little cheat sheet with some of the functions that can be easily delegated to AI and those that are better left to humans:

AI handles well

Human judgment required

Routine FAQs, order status, tracking

Travel disruptions, system outages

Ticket triage and intent routing

Billing disputes with policy exceptions

Summarization and agent note-taking

Emotionally charged escalations

Sentiment detection and smart routing

Fraud reviews, VIP retention

24/7 first contact for simple queries

Regulatory and compliance decisions

Having come to these conclusions in the early days of implementing our AI agent, my team and I have decided to offer our clients a unique collaboration model – human+AI support outsourcing. 

How does it work in practice?

The setup can be arranged based on each business’s unique needs, but the main premise is that it runs in tiers or layers:

1. The AI-only layer (tier 0/1) → AI is used to cover routing, triage, FAQs, and summarization. They can run 24/7 at near-zero marginal cost per extra query and provide necessary assistance even during the peak season rush.
    

2. The layer of human agents + an AI copilot (tier 1/2 ) → Agents mostly work on complex, emotional, policy-sensitive, regulatory, and VIP cases, while the job of AI is to surface relevant knowledge, flag shifts in sentiment, and draft responses. No matter the function, it’s the agents who stay in control of the customer conversations.
   
   One important thing to establish here is the handoff – how and when the AI should escalate to a human. You can use AI to capture the full context of a conversation and transfer it, along with the ticket, to the appropriate human agent. This way, you can reduce repeated requests and prevent customers from having to re-explain the situation.

Lawmakers are already moving to guarantee customers the right to reach a human, and Gartner expects that shift to push assisted-service volume up 30% by 2028. Thus, companies that have kept a strong human layer are more likely to see their operations succeed than those that gutted their teams in pursuit of quick automation savings. What I keep seeing with teams already running this way is that their resolution costs are steadier, and their customers stick around longer.

So, customer support AI: yes or no?

My honest take is that AI in customer support is worth the hype, as long as there are people standing behind it. And stop optimizing support operations simply to cut costs. After all these years in the business, I’ve come to a simple conclusion: good support pays for itself, and bad support just hides the bill somewhere else.

So, implement AI in your support operations, keep a human in the loop, track what you actually resolve, and you will finally see the cost savings you were promised.