Customer support trends from 5.85 million real conversations

Every figure here rests on one definition, so it is worth setting out before any of the numbers.

A conversation is resolved when the AI agent handled it end to end without human intervention and the customer got what they came for. A conversation is unresolved when the agent did not complete it on its own.

Unresolved covers four outcomes, and they carry different weight:

• Routed to a colleague. Often the correct result, and how a well-configured support operation is designed to behave.
• Outside the agent's configured scope. A boundary the team set on purpose.
• The question was ambiguous as the customer asked it.
• The answer did not exist anywhere the agent could reach.

The last one is the gap worth closing. We separate the four throughout, because a single unresolved figure read as failure would misdescribe most of what sits inside it.

How we score a conversation

• An AI reviewer reads each conversation 15 minutes after the last message.
• It judges, from the exchange and the customer's tone, whether the person got what they needed.
• If the customer comes back, the conversation is scored again after the next quiet period, so the judgment reflects where it ended up.

Published resolution figures rarely describe the same unit. A single scored chat is narrower than a full self-service journey across several channels, so two figures can differ on definition alone. Ours is stated here so you can compare like with like.

All data used in this research is aggregated and anonymized. Businesses that explicitly opted out of their data being used for aggregation or research are excluded from the dataset.

On the questions shown in this report. All data is reported in aggregate. The questions throughout are sample questions, written to represent the question types identified in the dataset. They reproduce no customer’s wording, and no business, user or individual conversation is identifiable from them.

Every figure comes from the Wonderchat conversation dataset covering March 2023 to August 2026. All of it is free to cite with attribution.

Gartner surveyed 321 customer service and support leaders between September and October 2025 and found that 91% report pressure to implement AI in 2026.¹

That pressure lands on the person who has to sign the contract. They are asked to forecast what an AI agent will do to their customer service costs and their service quality, and most available guidance compares vendors against each other without explaining what produces a result.

Our data points at the operation underneath. Across 2,794,330 scored conversations, AI resolved 75.3% of customer conversations end to end in 2026, up from 56.7% in 2024. The share the agent did not complete alone fell from 43.3% to 24.7% over the same two years.

Modern customer service operations are absorbing more volume without proportional headcount growth, which frees experienced agents for the conversations that carry judgment, revenue or risk. The direction is established. What remains open is how much of the remaining gap each team closes.

What the AI agent handles end to end

A consistent band of requests resolves without a colleague stepping in:

• Password resets and account access
• Delivery timelines and order status where the agent is connected to the order system
• Opening hours, locations and contact routing
• Policy questions covering returns, cancellations and eligibility

These routine tasks share one property. The answer already exists in published material and the question maps cleanly onto it, which is where AI for support teams performs most reliably.

AI agents also work continuously, which matters for support volume arriving outside staffed hours, and they carry multilingual support without additional headcount. Our dataset spans 13 languages with meaningful volume, so multilingual capability appears here as a measured pattern.

Where the agent supports the team rather than replacing it

Alongside end-to-end resolution, agent-assist capabilities change how experienced people spend their day. The capabilities below are standard across the category, and our data speaks to their effect on resolution rather than to their internal mechanics:

• Automated triage classifies and prioritises incoming support requests by customer sentiment and topic, shortening the path to the right queue for support agents.
• An AI copilot retrieves relevant information and drafts replies mid-conversation, so agent assistance lifts agent productivity while the colleague stays in the exchange.
• Complete customer context, drawing on customer history and past interactions, makes personalized support workable at a scale lean support teams cannot staff for.
• Natural language processing reads customer intent across support interactions, so high volume periods are absorbed without the queue lengthening.

Four patterns run through the dataset, and each carries a consequence for how support operations are staffed and documented.

Taken together they describe the customer service industry trends that matter operationally, and they are the customer service trends 2026 a support leader can act on rather than watch.

The through line across all four is operational. What separates strong performance from weak sits in documentation coverage, routing rules and the systems the agent can reach.

Resolution is rising, and unevenly

The average climbed 18.6 points between 2024 and 2026. Over the same period the distance between the strongest and weakest businesses held near 42 points. Both are true at once, which is why an industry average tells a support leader very little about their own deployment. Benchmarking against the average compares you to a number no individual business is obliged to hit.

Support volume concentrates where resolution is hardest

Every conversation in the dataset falls into one of four stages, and the shares below describe all conversations, not only unresolved ones. Support conversations, meaning existing customers with a problem to fix, make up 53% of the total and resolve at 60.8%. Pre-sales conversations make up 23% and resolve at 73.6%. The largest category is also the hardest, so a blended resolution figure conceals the harder problem underneath it. Two teams reporting the same overall rate can be performing very differently once the mix is accounted for.

Documentation coverage is the constraint

In 58.5% of unresolved conversations the answer had never been documented anywhere the agent could reach. Content that existed but had gone out of date accounted for 1.2%. Among emerging customer service trends this carries the clearest operational consequence, since the fix belongs to the team that owns the content and needs no procurement cycle.

The mix is multilingual by default

Thirteen languages carry meaningful volume. English accounts for 38% of conversations and German for 20%, with French at 9% and Dutch at 8% behind them. A support operation planning English-first coverage is planning for a minority of its own volume, and documentation gaps compound in every language the knowledge base was not written in.

Figures for AI resolution circulate widely, and they describe different things. Some are forecasts of where a category might land. Others are measurements of what happened. This section sets the two side by side so a business case can use each for what it is.

A forecast describes where a category might land by a future date. A measurement describes what happened across 1,000+ businesses over three and a half years, and it can be checked against your own numbers. Both belong in a business case, in different columns.

The third row is the one worth settling early. Treating failure as a model problem leads to a vendor change. Treating it as a documentation problem leads to writing the answers that are missing, which costs less and can start this week.

Among 687 businesses with enough conversation volume to compare fairly, resolution ranged from under 46% at the bottom decile to 88% at the top, with a median of 69%.

Every business in that comparison runs Wonderchat, on the same models and the same release. In their case the technology was held constant and the results still differ by 42 points, which places the variable inside each operation. Outcomes across customer service AI agents track what those agents have been given to work with.

Where your own rate sits

If you do not hold this number, it is the share of conversations your agent completed without a colleague stepping in. The methodology section above sets out how we score it, so you can measure the same way.

We sampled 5,000 conversations the agent did not complete alone and classified the cause of each. The five causes below account for all of them.

• 58.5% failed because the answer had never been documented anywhere the agent could reach. A separate internal check, run by support teams tracking their own gaps, placed the same figure near 63%.
• 19.3% could not be attributed to a cause from the conversation alone. That residual is worth naming rather than folding into one of the others.
• 12.6% fell outside what the agent had been configured to cover, which is a boundary the team set on purpose.
• 8.3% involved a question that was ambiguous as asked, and documentation does not resolve a question the reader cannot parse.
• 1.2% traced to content that existed but had gone out of date.

Read together, around six in ten unresolved conversations are a documentation problem the team owns and can fix this quarter. The remainder divides between deliberate scoping, ambiguity and cases outside our classification.

That balance changes what a team does next. Assuming stale content is the problem leads to a review cycle across relevant knowledge base articles, when the customer requests arriving in the support queue without answers are usually questions nobody has written about at all. Incoming support requests carry customer intent the knowledge base has never addressed, and ticket volume accumulates around those gaps until somebody writes the answer.

The industry is arriving at the same conclusion from another direction. In the same February 2026 research, Gartner reported that 58% of service leaders plan to move agents toward knowledge management work, on the basis that AI systems need accurate and current content to function.¹

Every industry in the dataset leaves a share of questions unresolved. What differs is which questions, and whether writing an answer would close them.

Across the 15 industries with enough volume to measure, resolution runs from 74.4% at the top to 58.1% at the bottom, against an overall average of 65.5%.

E-commerce and retail carries more volume than any other industry in the dataset, while sitting near the bottom of the range. That combination makes it the single largest pool of unresolved conversations, and the reason is structural rather than a question of documentation quality.

What e-commerce customers ask that goes unanswered

E-commerce and retail resolves 61.2% of its conversations. The unanswered questions cluster around accounts and transactions.

• 16.5% mixed account and payment issues
• 15.9% selling, payouts and payment processing
• 12.8% product advice and order tracking
• 11.3% marketplace account and transaction procedures
• 9.6% cancellations, refunds and listing issues
• 9.1% account and profile settings
• 8.0% login and account recovery

Each of these is account-specific and transaction-specific, so the answer lives in the order system rather than in a help centre. Additional articles will not close them.

What higher education students ask that goes unanswered

Higher education resolves 72.7%. The unanswered questions here are largely answerable from content, which is part of why the sector sits near the top of the range.

• 13.3% admissions and enrolment
• 12.3% programs, registration and university information
• 11.6% mixed multi-intent requests
• 10.5% schedules and campus logistics
• 10.0% program and course information
• 9.1% systems and portal access
• 9.1% HR and employee records

Each of those has a documentable answer, so writing them into the knowledge base moves the resolution rate directly.

What BFSI customers ask that goes unanswered

BFSI resolves 61.0%, and the pattern follows e-commerce more closely than education.

• 17.5% case status and support access
• 12.9% recurring payments and scheduled transactions
• 12.4% payments and transfers
• 12.2% product and account support
• 10.7% login and access errors
• 8.6% applications and renewals
• 7.7% account, app and promotion questions

The first two carry the point of this chapter. A business waiting on a payout, or a customer asking why a transaction is on hold, needs the record behind the answer. Either the agent reaches it or the conversation reaches a person, and the cost of getting that wrong is not measured in resolution points.

Login failures sit differently again. A customer locked out of their account has a question with a documentable answer behind it, and a repeat attempt that still fails is a signal to route them to a colleague. A team that has not written that rule down is leaving the decision to the system.

What industrial supply customers ask that goes unanswered

This cut has not presented enough volume to publish as a rate. Businesses field internal directory, routing and document requests, answerable from an internal knowledge base nobody has built, alongside catalogue questions. Both are documentable, which is the useful finding here.

The gaps that showed through recur often enough to name:

• product availability
• specifications and compatibility
• regional ordering contacts
• internal directory and routing
• document and contract administration requests

The AI agent resolved 73.6% of conversations with people still deciding whether to buy, and 60.8% of conversations with existing customers holding a problem.

The split follows the nature of the questions. Pre-sales customer inquiries cover pricing, features and comparisons, and those answers usually exist in published material already. Support questions carry customer history, account context and consequence, which places them beyond what documentation alone can settle.

Part of that gap belongs where it is. A question requiring judgment, a policy exception or access to a customer's account is better handled by a colleague, and the difference between 73.6% and 60.8% partly reflects the system routing correctly rather than falling short.

What this changes about benchmarking

For a support leader tracking customer service employee metrics, the conversation mix matters as much as the rate. A team resolving 60% on a support-heavy mix may be outperforming a team resolving 70% on a pre-sales-heavy one. Comparing resolution without comparing mix produces a misleading read, so the mix is the first thing to check before benchmarking against anyone, including the figures here.

The same applies to targets. A single resolution goal across an operation with a shifting mix rewards the quarters where pre-sales volume happens to rise, rather than the quarters where the team improved.

Cost per contact is the number most support teams reach for and few can state precisely. This section gives you two published benchmarks, our own list pricing, and a calculator that returns a figure you can defend.

Two published benchmarks frame the range. Gartner places the median cost per contact at $1.84 for self-service against $13.50 for agent-assisted interactions.⁶ Our own list pricing sits below the self-service benchmark, at $0.12 per resolution on the Basic plan and $0.08 on Scale.⁷

That distance between channels is where customer support automation produces a saving. Once the content exists, the marginal cost of an automated resolution sits close to zero, while every contact reaching a colleague carries their fully loaded time.

Pressure runs the other way at the same time. Support costs climb even when volume holds flat, because each ticket carries more expectation and more risk of a poor outcome than it did two years ago.

What moves the number without adding headcount

• First contact resolution. Each repeat contact carries close to the full cost of the original, so a deflection that fails to resolve raises total cost.
• Automated triage. Routing a request correctly on arrival removes the handling time attached to every ticket that lands in the wrong queue.
• Automated ticket documentation. Writing up the conversation afterwards returns that time to assisting customers.
• Predictive analytics on customer sentiment. Identifying an escalating issue early removes the second and third conversation it would otherwise generate.
• Consistent responses across channels. The same question answered three different ways is what generates the repeat contact.

What closing the gaps is worth

Across 84 businesses that added missing documentation, resolution moved from 75.4% to 80.8% over three months. Of those, 57% improved, 23% declined and the remainder stayed broadly flat, so the average describes a tendency rather than a promise.

How the calculation works

• The saving is colleague time, not software cost. If you already run an AI agent, closing documentation gaps does not change what the platform costs you. What changes is how many conversations reach a colleague, so the saving is the colleague-handled contacts you no longer pay for.
• Only escalated conversations count. A customer who gives up instead of contacting you costs no colleague time, so those are excluded here. They carry a cost to the relationship, which sits outside this calculation.
• Headroom is applied. The 5.4 point movement came from businesses sitting near 75.4% before the change. The calculator uses up to 5.4 percentage points, capped at the available headroom below the 88% top-decile resolution rate observed in this study, and explains when that cap applies.
• Full resolution is not the target. A share of conversations should reach a colleague by design, which places a realistic ceiling around the 88% the top decile reaches.

Support becomes a smaller cost center across customer service operations where teams automate support for the questions documentation can answer. Workflow automation and customer service automation reduce ticket volume by removing repeat questions from the queue, which lowers operational costs across service operations and returns experienced people to the work that needs them. Teams that reduce customer support costs this way do it by lifting resolution, and the same lever is what lets a business reduce customer service costs overall.

Gartner forecasts that agentic AI will resolve 80% of common customer service issues by 2029 while cutting operational costs by 30%, with customer satisfaction scores improving 15% to 20% where deployments work.² That is a forecast, so it belongs in a business case as direction of travel.

Two properties of documentation separate the top of the range from the bottom, and both are measurable.

Businesses with knowledge bases above 150,000 characters resolved about 74% of conversations, against about 61% for those under 10,000. That comparison isolates size, giving a 13 point difference attributable to depth.

Recency moves alongside it. Sorting businesses into four groups by resolution shows both properties rising together.

That table compares the worst-run accounts with the best-run ones, so it carries documentation depth together with everything else a strong team does well. The isolated size effect, holding operational quality aside, is the 61% against 74% comparison above. Only the second supports a claim about size on its own.

Depth alone is not enough. A large archive untouched for five months behaves like a small one, which makes an AI-ready knowledge base a matter of two properties held together.

What belongs inside

• The questions arriving repeatedly, written in the words customers use. A customer searching for “where is my order” will not match an article titled “Fulfilment status enquiries”, and the agent inherits that mismatch.
• Frequency order. Customer data about which questions recur gives customer service teams the priority, and support efficiency improves fastest when the most common unanswered question is documented first.
• A review cadence. The strongest groups touched their content within about 48 days.

Writing an answer closes some unresolved volume. The rest needs a connection to the system holding the record.

The industry data above draws the line. Questions about programmes, policies, eligibility and process are answerable from documentation, which is part of why higher education sits near the top of the range. Questions about a specific order, refund or account balance need live data, which is why e-commerce and BFSI sit lower while carrying more volume than any other sector in the study.

What each category needs

• Order tracking requires a live link through seamless integration with the commerce system, so the agent reads the actual shipment rather than describing the policy.
• Account questions require a link to the customer record, including balance, status and history.
• Refunds and cancellations require both, plus permission for the agent to act rather than only report.

This is where the choice of AI customer service platform becomes material. An agent connected to Shopify, Salesforce, Zendesk, HubSpot or Microsoft Dynamics can answer from the record itself, which converts a category of conversation documentation was never going to close. Teams comparing customer service tools should weigh those connections above surface features, since the key features that decide outcomes are whether the platform supports chat alongside your other channels and whether it reaches the record the customer is asking about. Wonderchat supports those connections directly, and no code tools have brought them within reach of a support team without engineering time.

The same connection carries a second return. An agent reading live account and order data can recognise a buying signal as readily as a support issue, qualify the visitor and capture the lead in the same conversation, which turns the support surface into a revenue surface rather than only a cost centre.

Basic self service stops at this boundary. A help centre answers the documented question and ends there, while teams that connect AI agents to the systems holding customer context extend resolution into account-specific territory, which is where the remaining volume sits once documentation is in good shape.

Since 2 August 2026, Article 50 of the EU AI Act requires an AI system interacting with a person to disclose that it is AI.⁸ For support operations serving EU customers, that turns disclosure and escalation into build requirements.

This applies to more of the dataset than a regional note would suggest. German accounts for 20% of all conversations, with French at 9% and Dutch at 8%, so a substantial share of the businesses represented here operate at least partly inside the regulation's scope.

What to put in place

• Disclose at the first message. State clearly that the customer is talking to an AI agent, at the opening of the conversation rather than in terms nobody opens.
• Make the human path visible. Give the customer a stated way to reach a colleague at any point, and name the conversation types that route automatically.
• Keep the transcript. Retain a record of what the agent told the customer, so a disputed answer can be checked.
• Document the routing rules. Write down which conversations escalate and why, which is also what Chapter 12 asks for operationally.
• Check the retention position. Conversations carrying customer history, account context or payment detail remain personal data under GDPR, so confirm what your platform stores and for how long.

Most of this is configuration rather than engineering. If you already run an AI agent, the practical step is to confirm your platform supports disclosure at the opening message, a visible handover to a colleague, transcript retention and a documented data processing position.

Any platform serving EU customers should carry these as standard rather than leaving them to the buyer to assemble. Wonderchat ships human handover on every plan, with SOC 2 and GDPR alignment on the platform, so the disclosure and escalation requirements are configuration steps rather than a build. Whichever platform you run, confirm it before your next EU-facing deployment.

The businesses at the top of the range documented their common questions before automating. The sequence matters more than the tooling, since an AI agent can only answer from what exists.

Step one. Rank your questions by frequency

Export the last quarter of customer inquiries from your helpdesk or chat logs. Group them by topic and sort by volume rather than by how interesting they are to answer. The top twenty topics usually cover the majority of contact.

Step two. Check each answer exists

Work down that list and search your own knowledge base for each question using the words a customer would type. Where the search returns nothing usable, mark the row. That list is your writing backlog, already in priority order.

Step three. Check the dates

Sort your existing articles by last-modified. Anything written against a product that has changed since is a liability rather than an asset, and the strongest groups in our data had touched their content within about 48 days.

Step four. Write the escalation rules down

Name the conversation types that should reach a colleague every time, such as refund disputes, policy exceptions, account changes and anything carrying a complaint. Configure those as routing rules before launch, so the decision sits with your team rather than emerging from how the agent behaves.

Step five. Measure the same way twice

Record your resolution rate before deployment using the definition in the methodology above, then measure it the same way afterwards. Without a matching baseline, any change is unattributable.

Those five steps describe readiness. Capabilities across AI powered customer service platforms converge quickly, and the differences between customer support AI agents matter less than whether your documentation can support any of them. The AI tools available to a customer support platform are broadly comparable at this point. Support AI agents deployed onto thorough content perform well, and the same AI support deployed onto gaps produces the range this study opened with.

Done in that order, the gain reaches past cost. AI agents for customer service that resolve reliably protect customer relationships and customer experience during growth, and customer satisfaction holds as volume rises. An AI customer service agent handling documented questions returns a customer support team to the work that builds loyalty.

Findings here are free to cite and republish with attribution. Please credit Wonderchat, The Deflection Gap, State of AI Customer Support 2026 and link to this page. Every chart carries its source line, so attribution travels when a chart is shared or screenshotted.

If you would like further breakdowns by industry, customer stage or language, they are in the downloadable report.

References

1. Gartner, survey of 321 customer service and support leaders conducted September to October 2025, published 18 February 2026. Source for the 91% pressure figure and the 58% knowledge management finding.
2. Gartner, forecast that agentic AI will autonomously resolve 80% of common customer service issues by 2029 with a 30% reduction in operational costs, published 5 March 2025.
3. Salesforce, State of Service, expectation that AI resolves 50% of service cases by 2027.
4. Gartner, finding that 14% of customer service issues are fully resolved in self-service.
5. Klarna, publicly reported AI assistant workload and subsequent rehiring of human agents.
6. Gartner, benchmarks to assess customer service costs, $1.84 per self-service contact against $13.50 agent-assisted.
7. Wonderchat published list pricing, annual billing. Basic plan $124 per month for approximately 1,000 resolutions, Scale plan $417 per month for approximately 5,000 resolutions.
8. European Union, Artificial Intelligence Act, Article 50 transparency obligations, applicable from 2 August 2026.
9. CSA Research, Can't Read, Won't Buy, survey of 8,709 consumers across 29 countries conducted with Kantar. Source for the 76% native-language preference and the 40% who will not buy from a site in another language.

The findings on this page are complete and free to read. The report adds the working material, built so a support team can act on the data rather than only read it.