AI Pulse

Every AI Lab Powering Your Support Tools Just Got Graded. The Best Score Was a C+.

By Felix Maru · July 16, 2026 · 7 min read

The Future of Life Institute published its AI Safety Index on July 7. By July 15, it was leading the conversation on The AI Daily Brief and across every AI news outlet covering the week. The central finding: the best score any AI company received was a C+, earned by Anthropic. OpenAI and Google DeepMind both landed at C. Meta pulled a D+. xAI, DeepSeek, and Mistral essentially failed. Not a single lab earned an A or a B.

If you work in customer support and your team is deploying AI agents on your help desk, those grades belong on your radar. Not as a panic signal, but as a vendor due diligence prompt you probably have not run yet.

What a C+ Actually Means for Your Support Stack

The FLI index grades six domains: technical safety, governance, transparency, risk assessment, security, and something called existential safety. Anthropic led because it has a more documented safety framework and publishes more information about how its models behave and where they can go wrong. That documentation matters for support teams specifically, not because of abstract risk concerns, but because it tells you how seriously your AI vendor takes model failures at the product level.

The domain most relevant to support teams is the combination of transparency and risk assessment. Does your vendor document known failure modes? Do they give you override controls? When the AI drafts a response to a frustrated customer and gets it confidently wrong, what happens next? Who is accountable, and what does the audit trail look like?

A C+ at the top of the industry means we are still early. The labs powering most commercial AI support tools are passing, but not comfortably. That is not a reason to stop deploying AI in support. It is a reason to deploy with clear human checkpoints, and to treat vendor transparency as a non-negotiable. If you have not asked your AI tool provider directly about their failure mode documentation and override controls, this week is the right time.

The Consumer Trust Gap Nobody Is Measuring

Here is a number that should be on every support leader's dashboard: research from CX Today and Hiver's 2026 State of AI Customer Support report shows that fewer than half of consumers say they trust AI to handle their service needs. Service professionals, including directors and support managers, consistently believe the number is much higher. The gap between what customers actually feel and what leaders assume they feel sits at roughly 20 percentage points.

That gap is producing real operational consequences. The same body of 2026 research finds that a large share of enterprises that deployed an AI customer support agent had to roll it back or shut it down after deployment, often more than once. The pattern is predictable: the pilot looks great because the scope is narrow. Then volume scales, edge cases appear, a customer gets stuck in a loop asking for a human, trust erodes, and the rollback happens.

The fix is not to slow down on AI. The fix is to design for the trust gap from the start. Specifically:

An AI that knows when to stop and hand off cleanly to a human is not a product failure. It is good CX design. The agent on the receiving end of that escalation is not a fallback. They are the product, doing the work the AI correctly identified as out of its depth.

Mews, the Layoff Headline, and the Part That Got Cut

On July 7, Mews, a hotel software company serving roughly 15,000 properties across 85 countries, announced it was cutting approximately 15% of its staff, around 170 roles out of a team of roughly 1,350. CEO Richard Valtr was direct: the roles being eliminated were, in his framing, built for an era that is ceasing to exist. The company had raised a $300 million Series D partly to build AI agents, and the restructuring reflected that shift.

That headline spread everywhere. But here is the detail that most coverage buried: a company spokesperson confirmed to both Skift and PhocusWire that customer-facing roles were largely unaffected. The cuts came from internal design, product, and engineering coordination layers that AI now handles, removing handover overhead that built up during fast growth.

This distinction matters if you manage a support team and your agents are asking hard questions about their future. The honest answer: AI is eliminating repetitive, low-judgment work. The people most exposed are those whose entire day is tier-0 lookups, copy-paste resolutions, and tasks a well-configured workflow could handle. The people with the clearest path forward are those doing the work AI cannot: reading a customer's frustration between the lines, making judgment calls on edge cases, and building the kind of trust that creates long-term retention.

If you have not had this conversation with your team directly, have it now. Not because layoffs are on the agenda, but because your agents deserve a clear picture of where the upgrade path leads, and you are in the position to draw it.

What Zendesk Actually Shipped in July That Support Ops Should Notice

While the industry was processing safety grades and trust research, Zendesk published its July 2026 product update, and two features are worth calling out specifically for support ops leads.

External knowledge sources are now live. You can connect Google Drive, Box, or Amazon S3 directly to Zendesk. Content from those sources becomes searchable by the help center, generative search, and AI agents. This matters because one of the most consistent failure modes in AI-supported help desks is the model giving a confidently wrong answer because the knowledge base it queries is stale or incomplete. Most support teams keep their real, up-to-date documentation in Google Drive, not in the help center. This connection closes that gap without requiring a manual sync process.

Workforce Management (WFM) is now included in Suite Professional plans at no extra cost. It was previously a paid add-on. WFM gives managers real-time visibility into agent activity and average handle time, plus automated alerts when an agent is working outside their scheduled task. For support teams now managing a mix of human agents and AI agents, this kind of observability is not optional. You cannot tune what you cannot see. WFM is how you find out whether AI is actually deflecting the right ticket types or just handling the easy ones and leaving agents with a harder queue than before.

Neither of these is a headline announcement. Both are the kind of foundational work that makes an AI-human support operation actually function at scale over time.

Three Actions for This Week

  1. Run a vendor safety audit. Ask your AI tool provider directly about their failure mode documentation and override controls. Missing documentation is your answer.
  2. Measure your trust gap. Add one question to your next post-interaction survey: "Did you feel confident the AI could handle your issue?" Compare it to what your team assumes. The gap is where your next deployment decision lives.
  3. Have the team conversation. The Mews story is circulating. Your agents will see it. Frame "what AI does to our work" as an upgrade path, not a replacement timeline, before they fill in that blank themselves.

The Safety Index is not the last time this industry gets graded. Support teams building human-in-the-loop operations now are the ones positioned well for whichever direction the next report goes.

Sources

If you're working through how much autonomy to give your AI agents and where to keep humans firmly in the loop, drop me a line. Happy to think through the specifics with you.

Share 𝕏 in ✉

Comments