State of Customer Support 2026

Faster answers. Harder problems.

Based on Screendesk’s analysis of public interviews and practitioner statements from 108 distinct people: 41 with sources published in 2026 and 67 historical perspectives published in 2020–2025. We did not conduct these interviews. This is a research review, not a new survey or a representative industry sample.

Support automation promises faster answers. The more interesting question is what happens to the work that remains: the difficult investigation, the frustrated customer, the missing context, and the people responsible for making a lasting improvement.

The accounts below describe different businesses, products, and operating models. We compare their approaches rather than combining incompatible company metrics into an industry average. The complete source directory follows the report.

What should support optimize for?

A faster reply can still leave the customer with more work. Meghan Gallagher describes speed targets encouraging acknowledgments without answers. Priscilla Brooke recalls an early two-person Buzzsprout team skipping lunches while chasing shorter response times, before shifting attention to more thorough replies. Their July 2026 conversation distinguishes the first response from a useful outcome. Read the interview.

Speed still matters. In a July interview, SimpliSafe’s Elise Jeanmaire explains why operational measures help teams forecast and manage capacity. But she also recalls routing work to the strongest agents to reduce handling time: the gains came with burnout and fewer opportunities for newer colleagues to learn. Watch at 6:14.

The distinction is between improving the whole service and improving one number. A useful scorecard asks whether customers get an answer, whether they must repeat themselves, and whether the team can sustain the service.

Where does AI help—and where does it fail?

The clearest cases describe a specific task, not a promise to automate everything. Zywave’s Emre Tekoglu describes email classification and quality review. Q2’s Lou Senko describes assembling logs, configuration, and other technical context inside existing workflows. These accounts place AI within support work, rather than judging it only by the number of customers who avoid a person. Zywave, June 2026, Q2, September 2026.

RJ Young’s Amanda Carden-Chamberlain reports using Agent Assist to surface resources during human conversations. Her account separates that deployed assistance from her predictions about a more automated future. Watch at 28:43.

Other contributors emphasize the content behind the tool. Writing specialist Leslie O’Flahavan recommends testing drafted knowledge articles with real customers: can they follow the steps, and do they need more explanation or screenshots? Elise Jeanmaire emphasizes human revision to preserve the business’s voice and understanding. The comparison is not simply AI versus humans; it is generation versus accountable review. Leslie, April 2026, Elise’s opening excerpt.

When should a customer reach a person?

Escalation needs both access and continuity. Amanda calls for human involvement in complex or emotionally charged interactions. Bank of America’s Andrew Santoro describes Erica passing conversation information to an associate so the customer does not restart. Amanda at 27:20, Bank of America, September 2026.

Henry Schein One’s Nate Christiansen distinguishes a current ticket-based handoff from a further improvement being developed: allowing a caller to join the human queue while AI continues helping. A proposed experience is not a completed deployment. July 2026 interview.

Historical accounts help explain the friction. ROI Hunter’s Libor Bitala described a chatbot that neither resolved his request nor connected him to a person in 2023. MultiTracks.com’s Trevor Kimball described helpful staff inside a system of repeated transfers and waiting in 2024. These are personal service experiences, not measured results from their employers. Libor, Trevor.

What makes an issue hard to resolve?

A plausible answer is not enough when the environment is unknown. Lou Senko’s technical-support account emphasizes customer-specific configuration, logs, and component state. Libor Bitala’s historical account similarly describes using databases, logs, APIs, and permissions to understand the full issue. These examples support collecting context before asking someone to investigate; they do not establish a measured benefit for any particular tool. Q2, ROI Hunter.

Context is also human. Helen Dimgba’s complaint-handling interview emphasizes investigating facts without making the exchange accusatory. Stacy Justino’s 2021 account explains why technically correct instructions still need to establish that the customer’s problem is understood. Different customers need different levels of explanation. Helen, June 2026, Stacy.

How should support work across teams?

Support can make the problem visible before it becomes an engineering task. Alex Batchelor’s 2026 account describes engineers learning from frontline conversations and recurring patterns, rather than seeing only the subset escalated as bugs. Elise Jeanmaire describes building regular relationships with engineering and marketing so CX hears about changes before customers do. Alex’s interview, Elise at 19:01.

Historical accounts show related approaches. Andrea Saez describes combining customer-facing and product-facing colleagues in discovery conversations. Kat Gaines discusses including customer-facing teams in incident communication and reviewing gaps in information and approval. Their examples concern different workflows, but both give support a role in the decision rather than making it the last recipient of an update. See their original interviews in the source directory.

What do metrics reveal—and miss?

Contact counts, satisfaction, and resolution measure different things. Meghan Gallagher warns that first-contact and first-response measures do not establish actual resolution. In a 2022 account, Ethan Walfish separates product area, contact cause, and resolution route so the team can understand handoffs that issue counts alone obscure. Denise Twum’s 2020 account explains how satisfaction feedback can conflate the product with the support interaction. Their accounts call for interpretation, not abandoning measurement.

Vendor perspectives can inform the comparison, but need a separate label. AmplifAI founder Sean Minter argues that low handling time can coexist with poor satisfaction. We include his account as a vendor perspective, not an independent operator result. March 2026 interview.

We do not combine headline claims about automation rates, QA coverage, or customer satisfaction. Definitions, baselines, sample sizes, and measurement windows differ or are unavailable. This review does not rank companies by their reported numbers.

What skills and roles are changing?

Automation changes the work around the queue as well as the queue itself. Gamma’s Hilary Dudek discusses AI-related support roles. Planning Center’s Reagan Helms describes using an AI management assistant connected to internal knowledge to prepare coaching and one-to-ones, with employees aware of and approving the arrangement. These are distinct applications: customer-facing help and the management of the people providing it. See their March and May 2026 interviews in the directory.

Training remains part of the operating model. Elise’s routing example cautions against concentrating all difficult work in the most experienced agents. In 2025, 37signals’ Chase Clemons described using successful and unsuccessful cases to develop independent judgment, with colleague help available. The comparison suggests that removing routine work should be accompanied by a plan for how people learn the remaining work. 37signals interview.

What would practitioners change first?

Several accounts start before buying a tool. Emre Tekoglu recommends clarifying the problem, checking people and processes, and agreeing success criteria. Amanda asks vendors for evidence from comparable businesses and distrusts promises to solve every problem. Buzzsprout’s Cara Pacetti describes turning recurring inbox questions into reusable customer education, helping people before they know how to ask. Emre at 1:57, Amanda at 30:12, Cara, August 2026.

Our interpretation is to start with one customer problem and follow it end to end: what the customer already supplied, what the agent still needs, when another team becomes involved, and how the outcome is confirmed. Then choose the smallest change that removes an actual obstacle.

Screendesk’s perspective

For technical support, useful investigation context can include steps to reproduce, the expected and observed behavior, relevant configuration, and a recording when seeing the problem would help. Make that context available to the next person handling the issue. This is Screendesk’s recommendation; the interviews do not constitute an evaluation or endorsement of Screendesk.

How we compiled this review

We reviewed selected passages in public practitioner interviews, published Q&As, and statements, and organized attributable contributions around eight questions. One person counts once, even when several interviews are available. Panelists count separately only when their own contributions can be identified. Shared employers and shared interviews are not independent company cases.

The core directory includes support operators, technical-support practitioners, former operators, support-specific consultants, and relevant technology vendors. Segments, source dates, publishers, hosts, and evidence locations are provided. Unknown job titles or employers remain unspecified. Nine broader commentary or adjacent-role entries were retained as background and excluded from the 108-person core.

The 2026 group is defined by source publication date, not necessarily the date of the described deployment or original recording. David Kay explicitly dates his interview to fall 2024 despite its 2025 publication. Earlier sources provide historical perspectives; differences in interview topics cannot establish a trend over time.

The sample is not random or representative. Interview prompts differ, publishers may have commercial interests, and some transcripts are automated. We paraphrase selected passages rather than treating each interview as a complete account of a person’s views. An unmentioned topic is missing evidence, not disagreement. Company outcomes are self-reported unless otherwise noted; we have not independently audited their operations.

The coverage table counts people with at least one selected passage relevant to each question. These counts describe this source collection, not the prevalence of beliefs or practices in the industry.

Source directory: 108 public voices

Browse the selected contributions below, or download the full source ledger (CSV). The ledger includes roles, publication dates, source links, evidence locations, paraphrases and limitations. These people did not respond to a Screendesk survey.