The State of AI Customer Support in 2026: The Receipts

If you buy customer support software in 2026, the AI in it will probably arrive as its own line on the invoice, priced by the answer. Published rates run from about $0.05 per conversation to $1.25 per resolved ticket. Three years ago, almost none of this existed: you bought seats, and whatever intelligence the tool had came with them.

There is a second number every buyer should know before signing. Vendors advertise that their AI resolves 65 to 86 percent of conversations on its own. The numbers in their own published case studies cluster at 40 to 70 percent. When the AI is billed per resolution, that gap is not a rounding error. It is the difference between the bill you modelled and the bill you get.

This report documents both shifts with evidence you can check. Instead of surveying people about their feelings, we tracked what vendors actually did: every dated claim below rests on a vendor announcement or an archived, dated copy of the vendor's own pricing page, all listed in the methodology.

One disclosure before we start. Drag, our product, is one of the 33 vendors tracked here, and we sell the opposite model to most of this market, with AI included in the seat. Keep that in mind as you read. The evidence does not care who compiled it.

What this report covers

• The three tiers of AI support, and which one you are actually buying
• How AI took over the pricing pages, year by year
• Pricing comparison: what AI support actually costs
• The resolution-rate gap
• What leaders say vs what the market did
• The demand signal: AI assistants now send the traffic
• The consolidation year
• The MCP frontier
• Cite this report
• Methodology and receipts

The three tiers of AI support, and which one you are actually buying

"AI customer support" is sold as one thing. In practice it is three different products, and knowing which tier you are buying explains most of the pricing and most of the disappointment.

Tier 1: the copilot. AI that helps your agents: drafting replies, summarizing threads, tagging and routing. The human still sends every message. This is where most real usage lives today: in Hiver's survey, AI-suggested replies are the top use case at 39 percent. Copilots are usually priced per seat, either as an add-on (Freshdesk's Copilot at $29 a seat, Zendesk's at around $50) or included in the plan (Drag from $18, Gmelius, Superhuman's Business tier).

Tier 2: the autonomous agent. AI that faces the customer and resolves conversations on its own, with no human in the loop. This is the tier the meters bill for, the tier the resolution-rate claims are about, and the tier most of this report's pricing table covers. It is also the least mature: only 9 percent of leaders in Hiver's survey say AI assists with most of their ticket volume, and the production resolution rates cluster well below the advertised ones.

Tier 3: the operator. The newest tier flips the direction: instead of the support tool containing AI, an outside assistant like Claude or ChatGPT operates the support tool, through an MCP server. It is early, most vendors do not support it, and it is covered in its own section below.

The tiers get conflated in marketing because the same vendor often sells all three under one "AI" label. The practical rule: copilots are priced like software, agents are priced like electricity, and operators are mostly not priced at all yet. When a pricing page shows one AI headline, ask which tier it belongs to before you compare numbers.

How AI took over the pricing pages, year by year

Surveys measure what people say. Pricing pages record what companies actually commit to. So instead of asking vendors about AI, we tracked when AI first appeared on the public pricing page of each of the 33 support tools in our pricing receipts program, which has kept dated archive copies of every vendor's pricing page since 2021.

2023: two vendors. Tidio launched Lyro, a paid AI add-on. Intercom launched Fin, the AI agent that would come to define the category, and its $0.99 per-resolution price was on the pricing page by 2024.

2024: six. Superhuman moved its newest AI behind the Business tier in August. Kayako relaunched at $79 a seat plus $1 for every AI-resolved ticket. The Groove team launched Helply, priced per ticket. HubSpot announced Breeze on September 18 at INBOUND.

2025: ten. Help Scout's AI Answers meter arrived at $0.75 per resolution. Zendesk began metering AI resolutions above plan allocations, on top of $50-a-seat add-ons. Hiver introduced a $20-a-seat AI add-on. Front announced Autopilot on September 9.

2026: fifteen. Crisp put AI credit balances on every paid plan card. Gorgias rebranded around its AI Agent. Gmelius introduced its Meli tier, Sortd rebuilt its plans around AI, and Freshdesk's Copilot-plus-session-packs stack was confirmed on its live pricing page.

Two findings sit inside that curve, and they matter more than the count itself.

The first is the pricing shape. Of the 15 vendors that price AI, 11 charge by usage: per resolution, per conversation, per task, or through credits that expire monthly. Only 4 build it into the seat or tier. The category has made its bet, and the bet is that AI should be billed like electricity.

The second is the absence. Roughly half of the tracked market, mostly the free, open-source, and legacy tools, still shows no AI price line at all in 2026. This shift is real and it is fast, but it is not finished.

Pricing comparison: what AI support actually costs in 2026

The published usage rates span a 25x range. The sticker matters less than the unit each vendor bills on, so the table lists both:

Rates come from each vendor's own pricing page or published rate card, held in our July 2026 dataset and spot re-checked in September 2026. Confirm current rates before you buy.

There is a counter-bet on the other side of the market. Drag includes its six AI assists in the seat from $18 with nothing metered. Hiver folded its AI add-on into its plans from the Growth tier in 2026. Gmelius and Sortd price AI into their tiers. Whether the meter or the seat is better for you depends entirely on your volume, which is why we built a calculator instead of an argument.

The meters also sit on top of seat prices that have been climbing well ahead of inflation. Across the tracked vendors, entry prices moved as much as 257 percent between 2021 and 2026 (Hiver, from $7 to $25) against a US inflation reference of 23 percent:

To turn any of this into your own number, model your team size and expected AI volume in the shared inbox cost calculator. It prices the meters and the seats side by side using this same dataset, and it is the honest benchmark for this report: where your team's bill sits against the shapes above.

The resolution-rate gap: advertised vs production

The number that sells AI support is the resolution rate. It is also the least reliable number in the category.

Vendors advertise 65 to 86 percent autonomous resolution. The numbers in their own published case studies cluster at 40 to 70 percent, and Fin's own case studies sit at 42 to 50 percent. Our resolution-rates audit lines up every vendor's advertised claim against its published evidence, with links.

The gap has a mechanical cause: definitions. A billable resolution is whatever the contract says it is. At several vendors that includes an "assumed resolution," meaning a customer who left the conversation without asking for more help. In practice, that includes customers who gave up.

Here is what the gap costs. An AI agent that resolves 1,000 conversations a month at $0.99 each costs $11,880 a year on the meter alone. If its true resolution rate is 45 percent rather than the advertised 80, the cost per genuinely resolved conversation roughly doubles. Run your numbers at a realistic rate, not the advertised one, before you sign anything multi-year.

What leaders say vs what the market did

The best sentiment data this year comes from Hiver's August 2026 survey of more than 700 US support leaders, published as the gated report "State of AI Customer Support in 2026." Credit where due: it is a real survey with a documented respondent base, and it is worth reading. Its headline findings, quoted with attribution:

• 48 percent of leaders describe their teams as only moderately confident with AI
• The top uses are AI-suggested replies (39 percent), first-line chatbots (35 percent), and AI quality checks (30 percent)
• Only 9 percent say AI assists with most of their ticket volume
• 87 percent are cautious about letting AI represent their brand
• 25 percent have not measured AI ROI at all, the largest single answer in their ROI data

Now put that next to the pricing-page record above. Leaders are cautious. ROI is mostly unmeasured. Only a sliver of teams let AI carry most of their volume. And the vendors monetized anyway, at speed, mostly by the meter.

That tension is the honest state of AI customer support in 2026. The people buying the software are not yet sure it works, the companies selling it have already priced it like a utility, and the gap between what is advertised and what is delivered lands on the buyer.

The demand signal: AI assistants now send the traffic

One more behavioral series, this time from our own first-party analytics, published as shares and multiples rather than raw counts.

Traffic referred to this site by AI assistants (ChatGPT, Gemini, Claude, Perplexity and peers) roughly quadrupled between June and September 2026, doubling about every six weeks, and Gemini drew level with ChatGPT as a referrer over the summer. Assistants are still a small share of total traffic, in the low single digits of percent, but they are the fastest-growing channel we measure, and the questions they send are longer, more specific, and more purchase-shaped than classic search queries.

The practical reading: support buyers increasingly arrive having already asked an assistant what to buy. That is exactly why this report is an open page with anchored, citable figures rather than a gated PDF. Assistants cannot cite what they cannot read.

The consolidation year

The market's structure moved as fast as its pricing.

Intercom renamed the entire company after Fin in May 2026, and shortly afterwards Salesforce agreed to acquire it for approximately $3.6 billion. Zendesk acquired Forethought. Grammarly had already absorbed Superhuman in 2025 and renamed the combined company Superhuman that October. At the small end, the Groove team's pivot to per-ticket Helply shows the same pressure from below.

For buyers, the practical reading is simple. The vendor you sign with may not be the vendor you renew with. Pricing leverage is shifting toward platforms. And the per-outcome meter has so far survived every acquisition it has been through. Read your renewal terms with all three in mind.

The MCP frontier

The newest line on 2026 pricing pages is not a feature but an interface. MCP servers let AI assistants operate a support tool directly: reading the queue, drafting replies, assigning owners, reporting on what is still open.

Most support vendors do not ship one yet. Drag ships its own MCP server, with 47 tools, on every plan. Superhuman's sits on its Business tier. Zendesk, as of mid-2026, is building MCP client capabilities, which is the opposite direction. The MCP-native support stack guide covers how the architecture fits together, and the practical question for 2027 budgets is whether the tools on your shortlist can be operated by the assistants your team already uses.

If your queue is really a shared email address and you would rather have the AI included in the seat than running on a meter, that is the exact case Drag was built for, and the 7-day trial needs no card. If you need an autonomous, customer-facing agent at enterprise scale, the metered vendors above are the honest place to start, and the calculator will tell you what the meter costs at your volume.

Cite this report

This page is open on purpose. Every section heading is a stable anchor, the figures live in the text rather than inside a locked PDF, and you are welcome to cite any of it with attribution.

Suggested citation: DragApp, "The State of AI Customer Support in 2026: The Receipts," September 2026, dragapp.com/blog/state-of-ai-support/. If you need the vendor-by-vendor dating behind the adoption chart, the methodology below lists every source.

Methodology and receipts

The panel. The 33 support tools in our pricing-receipts program, tracked continuously since 2021 with dated Internet Archive copies of each vendor's own pricing page. The panel leans toward shared inbox and help desk tools, and it includes our own product, Drag.

What counts. "AI on the pricing surface" means a generative-AI feature with its own price line, add-on, credit balance, or explicit tier inclusion on the vendor's public pricing page.

How dates are read. "In [year]" marks a dated vendor announcement, such as Front's Autopilot announcement of September 9, 2025, or HubSpot's Breeze announcement of September 18, 2024. "By [year]" marks the first dated archive copy in which the AI price line appears, so the true date may be earlier.

The dates behind the chart. Tidio (in 2023) · Intercom Fin (product announced 2023, meter on the pricing page by 2024) · Superhuman (by August 2024) · Kayako (by 2024) · Helply (in 2024) · HubSpot (in September 2024) · Help Scout (by 2025) · Zendesk (by 2025) · Hiver (in 2025) · Front (in September 2025) · Crisp (by August 2026) · Gorgias (by 2026) · Gmelius (by 2026) · Sortd (in 2026) · Freshdesk (Copilot price line by 2026). Two tracked vendors, Drag and Missive, carry AI without a dated first appearance in our captures and are excluded from the bars. Including them would raise the 2026 count, not lower it.

Rates. The cost table reads from each vendor's own pricing page or published rate card as held in our dataset (as of July 2026), spot re-verified in September 2026.

Survey figures. All survey numbers are Hiver's, from the full PDF of its August 2026 report, cited with attribution. Where Hiver's landing page and its PDF disagree (brand caution: 61 versus 87 percent), we cite the PDF.

Our traffic data. Published as shares and multiples only, never raw counts.

Corrections. If a vendor believes any receipt here is wrong, email us. Corrections are recorded, not silently overwritten, consistent with our pricing-history program.