Skip to content

Gartner Customer Service & Support · Denver · 4–5 November 2026

Who checks the AI? What we learned at Gartner CSS 2026.

Two days in Denver, a busy booth and the most popular roundtable at the event, with 26 support leaders from 19 enterprises. Here is what it means for support teams.

Line of the week

The score isn't the insight. The real question is why customers contact you at all.
From the Auto QA session, and it matched what we heard at our booth.

From the Gartner sessions

Four things support leaders took home from Denver

  1. 01

    The market is split in two

    Most teams are still piloting AI, careful not to break experiences that work. On the same day at our booth we met teams reviewing about four calls per agent a month in spreadsheets, and teams a year into running agentic AI.

    Ask your team

    Which side of the split are we on, and what's our next step?

  2. 02

    Scores aren't insights

    50 to 70 question scorecards just automate noise. Measure a few behaviours tied to outcomes, then ask why contacts happen. 20,000 well-handled "where is my order?" contacts isn't a success story. It's a broken journey.

    Ask your team

    Which contacts do we handle well that shouldn't exist at all?

  3. 03

    Nobody owns the AI yet

    AI often lands as one more priority, with no named owner, authority or time. One company's top AI initiative had 0.3 of a full-time role, spread across three people. Capacity comes from subtraction: remove, reduce, reassign, then reinvest.

    Ask your team

    Who owns our top AI priority, with what authority and how much time?

  4. 04

    Your next customer may be an AI

    Customers' personal AI agents are starting to call contact centres, and they can know the answers to security questions. The model: predictable flows for routine work, AI agents for complex work, and safeguards before any sensitive action.

    Ask your team

    Could our journeys and authentication handle a customer's AI agent today?

From our roundtable

Who checks the AI? What enterprise support leaders told us.

26 VPs and directors of service, operations, AI strategy and QA from 19 enterprises, across telecoms, banking, insurance, utilities, retail, healthcare and media. Shared without names.

Most are piloting. A few run AI at scale.

  • Piloting · most

    Don't break what works

    Informational use cases first. Some pilots run up to a year, even with strong deflection and CSAT.

  • Scaling · a few

    Rolled out in stages

    Customer-facing AI expanded step by step. AI agents diagnose tickets; people take complex cases.

  • Deciding · rare

    AI makes the call

    AI agents decide simple claims, with people reviewing exceptions.

Most companies have business rules, not governance.

Permissions and escalations are set case by case. Few teams regularly review what their AI actually does. Practices we did hear:

  • Post-incident review. Check the transcript, expected answer and logic before fixing.
  • Supervise bots like people. If a bot has responsibilities, someone owns its performance.
  • Smart routing. A third contact about the same issue should skip the bot.

The most mature model we heard

Monthly human-vs-AI calibration

Each month, people and the AI score the same ~100 interactions. Operations and the former QA team review the gaps: fix the model, or the knowledge base?

The QA team didn't disappear. It became a process team.

Scoring everything changes the conversation with agents.

  • Set expectations first

    Agents used to a weekly sample suddenly had every contact reviewed, and some felt it was not fair.

  • Score what agents control

    Teams that made it work score a few behaviours agents influence, and supervisors coach.

  • Don't grade yourself

    Separate scoring removes self-marking bias. Keep the human context, and retrain.

Still open: where the room disagreed

Trust one bot before connecting many
Some won't link bots until QA proves a single one.
One rubric for people and AI?
Shared criteria show health; separate ones show who caused it.
One metric can mislead
Governance can't hang on a single number like repeat contacts.
ROI pressure pushes AI out early
Seen only as a cost cut, use cases go live before value is proven.

Downloads

Get both one-pagers

The four takeaways and the roundtable learnings, each on one page, ready to share with your team.

  • Recap: four takeaways from the Gartner sessions, each with an "Ask your team" question.
  • Roundtable learnings: where the room is, the governance gap, the calibration model, the people side and open questions.

Send me both one-pagers

Work email and company. The download opens straight after.

Curious how other teams are answering these?

Happy to compare notes.

Trusted by global support teams

  • Foot Locker
  • Canva
  • GetYourGuide
  • Instacart
  • BlaBlaCar
  • UiPath