76% in the Press Release, 51% Out of the Box: The Real Math of AI Support Agents in 2026

On June 15, 2026, Salesforce announced it was acquiring Fin (formerly Intercom) for roughly $3.6 billion, and the press release led with its headline argument: Fin's AI agent closes an average of 76% of support tickets without a human. That same week brought the Agentforce Summer '26 release with multi-agent orchestration, and Gartner is already forecasting that by 2029 agentic AI will autonomously resolve 80% of common tickets.

We build AI support and sales agents for clients' specific processes, so we look at these numbers without the hype. And we want to draw attention to a different figure, the one that isn't printed in large type: 51%. That's exactly how many tickets Fin resolved out of the box (for the Fin 2 generation), per Intercom's own data. The gap between the slide's "76%" and the "51%" at launch is precisely where most AI support projects either take off or quietly die.

The loud numbers are a press-release ceiling, not your result

Let's see where the "magic" percentages come from:

76% — Salesforce/Intercom. A figure from the Fin acquisition press release: an average across the best deployments, a marketing metric, not a guarantee for a new account.
80% is a forecast, not a measurement. Gartner (March 2025) predicts 80% autonomous resolution by 2029 and a 30% cut in operating costs. That's a direction, not today's reality. Current measured figures sit well below it.
60% — Gorgias. "60% of tickets resolved instantly" is a marketing promise from a product page, not an independently audited figure.

All of these numbers are a ceiling under ideal conditions. Useful as a directional benchmark. Dangerous if you write them into a business plan as fact.

Now the numbers that hold up under audit

Strip away the marketing and the picture is more sober — and still quite positive:

Intercom, its own statistics (April 2026). Across the base (7,000+ customers, 40M+ conversations), Fin resolves an average of 67% of tickets — up from 51% out of the box for the previous Fin 2 generation. In other words, 67% is the result of tuning and data, while a new account's starting point is closer to half.
A real case (Vagaro, Zendesk data). 44% of inbound tickets resolved autonomously, resolution time down 87%, CSAT up to 92%. Below "80%" — but it's a measured result from a single company, not an aspiration.
A study of 5,179 agents (NBER/QJE; Brynjolfsson, Li, Raymond). Access to a GPT assistant raised productivity (resolutions per hour) by 14% on average — and by 34% for newcomers, with almost no effect for experienced staff. In other words, AI's greatest value is augmenting people, not replacing them.
IBM (survey of 2,000 CEOs). Only a quarter (25%) of AI initiatives delivered the promised ROI, and just 16% scaled company-wide. The gap between pilot and profit is the main problem, not the model.

The Klarna lesson — the most expensive in the industry

The loudest AI support case is still cited by everyone. In February 2024, Klarna said its OpenAI-powered AI assistant handled 2.3 million conversations in its first month — two-thirds of all support chat — cut average resolution time from 11 minutes to under 2, did the work of 700 agents, and delivered roughly $40 million in profit improvement. The perfect slide.

The sequel is rarely cited. By May 2025, Klarna had reversed its "AI-first" strategy and begun bringing human agents back. CEO Sebastian Siemiatkowski admitted that an approach driven solely by cost savings produced "lower quality" service. The takeaway isn't "AI doesn't work" but "AI without an escape hatch to a human costs more than it saves."

Why the gap appears — and what the vendors themselves did

The root of the confusion is two different words merged into one:

Deflection — the ticket never reached a human (the bot replied something, the user left). This does not mean the problem was solved.
Resolution — the issue is actually closed, and that can be confirmed.

Many of the loud percentages historically counted deflection. And the best acknowledgment of this came not from critics but from the platforms themselves — by switching to charging for confirmed results:

Zendesk introduced its "Autonomous Service Workforce" at Relate 2026 (May) and expanded its outcome-based pricing model: you pay only for genuinely resolved tickets, and each resolution is verified twice — first by the agent itself, then by a separate AI evaluator model.
Salesforce in Summer '26 added multi-agent orchestration to Agentforce along with a simplified Help Agent priced per resolution (pay-per-resolution).

The signal is simple: if a vendor is willing to charge only for confirmed resolution, then the earlier "80%" wasn't quite about resolution.

What actually works in 2026

The winner isn't "100% autonomy out of the box," but a soberly designed system:

Design for the resolvable slice. The 50–67% that the agent genuinely and verifiably closes is an asset. The rest are candidates for escalation, not for forced automation.
Measure confirmed resolution + CSAT, not deflection. Otherwise you optimize the wrong thing and repeat Klarna's mistake.
Make escalation to a human the norm, not an emergency. A seamless handoff to an agent with full context is a feature, not a defeat.
Augment people, don't just replace them. +34% productivity for newcomers (NBER) is also a result — and often cheaper and safer than chasing full autonomy.
Integrate the data. An agent that can see orders, policies, and the CRM closes far more. Narvar NAVI automates post-purchase (returns, exchanges, refunds) within the retailer's policies and margins; Sierra, with the utility platform Kraken, embedded agents directly into billing (the platform covers 70M+ accounts). An agent's ceiling is first and foremost its access to data, not the model.

And one more thing: the value of agents isn't only in support savings, but also in revenue growth. At Salesforce, in Q3 of fiscal year 2026, half (50%) of Agentforce bookings came from expansions with existing customers — meaning the agent deployments themselves drive upsell. And per the State of Sales 2026 report, 54% of salespeople already use AI agents and expect 34% less time on research and 36% less on drafting emails. Support and sales are converging into a single agentic logic.

How we look at this (an integrator, not a box)

We don't sell "80% autonomy out of the box" — in 2026 that's an anti-product. We build an agent for your slice of tickets and on your data (CRM, orders, policies), turn on seamless escalation to a human, and measure confirmed resolution rather than deflection. It's duller than a press release — but the numbers survive an audit, and the brand doesn't learn the hard way through its own "Klarna case."

Rakes worth not stepping on

Writing "80%" into the plan as fact. It's a ceiling or a forecast; your number is calculated on your data.
Confusing deflection with resolution. Deflected ≠ helped.
Removing the human where humans are needed. Klarna's lesson cost it reputation; escalation is mandatory.
A pilot without data integration. An agent without access to orders and policies will hit a low ceiling — and you'll be convinced that "AI doesn't work."
Measuring the wrong thing. The count of deflected tickets is a vanity metric; count confirmed resolutions and CSAT.

Where to start

Not "AI support out of the box," but four sober steps:

Calculate your real ceiling. Take 90 days of ticket history and mark what percentage is genuinely automatable on your specific data. That's your number, not the vendor's.
Launch the agent on that slice — with integration into the CRM/orders and seamless escalation to a human.
Measure confirmed resolution and CSAT, not deflection.
Expand the slice only when the numbers hold; don't force up the autonomy.

Why this matters right now

2026 is the year the industry quietly shifted from asking "how many tickets did we deflect" to "how many did we actually resolve — and are we willing to pay for it." Vendors are already voting with their wallets, moving to pay-for-confirmed-result. The winner is the one who builds support around a number that survives an audit, not around a slide. In support, AI is a lever for human quality and speed, not a full replacement for it.

Will AI fully replace support?

No — and the 2025–2026 experience confirmed it. After a loud "AI-first" push, Klarna reversed course and brought human agents back, admitting the "lower quality" of a purely cost-driven approach. What works is a model where AI closes the resolvable slice of tickets while complex or sensitive cases flow seamlessly to a human. The study of 5,179 agents showed +34% productivity for newcomers precisely from augmentation, not replacement.

76% or 51% — which to believe?

Both figures are from Intercom, but they describe different things. "76%" is a press-release number (an average across the best deployments), "51%" is out-of-the-box resolution for Fin 2 before tuning, and 67% is the average across the whole base after tuning. None of them is your number: the real ceiling can only be calculated on your own tickets and data.

What's the difference between deflection and resolution?

Deflection means the ticket never reached a human (the bot replied, the user left), which doesn't mean the problem was solved. Resolution means the issue is actually closed and that can be confirmed. Many of the loud percentages historically counted deflection; it's no coincidence that in 2026 Zendesk and Salesforce switched to charging for confirmed resolution with double verification.

Where should a small or mid-sized business start?

With an audit of 90 days of tickets to calculate your own automation ceiling, then launching the agent on that slice with integration into the CRM/orders and escalation to a human, and measuring confirmed resolution and CSAT. Autonomy should be expanded only when the metrics hold. That's exactly the approach we at Aceverse build agents around: for your process, on your data, with real numbers.