We put 10 AI support agents to work at the same company. Here is what we learned.

The average support representative stays in a seat for less than a year, so companies retrain endlessly or hand the queue to business process outsourcers, then live with unwanted ramp time and uneven quality. Seasonality makes the problem worse: retailers can grow support staff by as much as five times for the holidays, hiring in September and letting people go in January.

This is the work AI agents absorb best: repetitive, policy-bound, and spiky. Teams getting value from agents can redeploy people to complex and proactive work instead of rehiring and retraining every year.

Buyers also want that capacity gain to show up in operating results. The Head of Customer Experience at Flipkartdescribed starting at 60–65% efficiency and building toward roughly 85%. That is one buyer's experience, not a universal ceiling or a direct comparison with G2's task accuracy scores. Interviewees assess resolution rates alongside satisfaction on resolved conversations and the effect on support staffing.

“The bottom line every institution looks at: a real reduction of FTEs without reducing satisfaction. Until that materializes, it's just numbers on paper.”

Every agent received identical tasks, tools, and policies inside G2's simulated company. The best Overall Score was more than double the lowest, while task accuracy ranged from 64% to 88%. Measuring actual performance helps buyers look beyond vendor claims.

We expect agents to improve quickly and take on more complex work involving payments, sensitive private data, and multiple integrations across a complete CX workflow.

Nearly every agent stayed on topic: relevance ranged from 82% for the lowest-performing agent to 100% for Retell. Larger gaps emerged in policy compliance, which ranged from 56% to 92% for Tidio.

Recurring failures included answering before checking the customer record, escalating tickets the agent could have resolved, and taking the wrong action while reporting success.

“The gap we see is not the models, but in the harness. Products that let the agent work straight from the policy and tools scored at least 20% higher than the ones that route everything through their own workflow builder.”

The cost of a wrong action shapes how buyers roll out agents. Interviewees described human checks for refunds, payments, and account changes, including at companies using Sierra and Decagon. A senior customer care leader at an ITSM software vendor described progressing from drafts to routine answers, then low-risk tasks, before real transactions.

“Everything informational is autonomous. Everything transactional needs a human validating before action, AI is mostly used for recommendation.”

For buyers, policy compliance is therefore a deployment decision as well as a benchmark score: define which actions the agent can take alone, which require approval, and when it must hand off to a person.

Salesforce and Jira came up as must-have integrations in separate buyer conversations. One buyer reported dropping a vendor over a three-hour Salesforce sync and walking away from another when the integration was quoted at $100,000. Buyers need to test the connection to their own systems, including sync speed and implementation cost, before committing.

“Without integration to our membership platform, the AI agent can't do anything except surface FAQs. It has to connect to the system that drives our business.”

Since G2 began evaluating CX agents, the market has continued to consolidate, including Fin being acquired by Salesforce for $3.6 billion and Forethought by Zendesk. Sierra, which is under evaluation, and Decagon, which declined evaluation, are differentiating through a forward-deployed engineering strategy rather than a self-serve model. Both are also developing frontier skills intended to handle longer-running tasks.

Interviewees differed on paying per resolution versus per conversation. Some preferred paying only for solved issues; others valued predictable conversation costs. Usage-based credits and long commitments drew objections. Buyers also described how changes in vendor ownership affected their plans and contract expectations.

“Per resolution puts me in a position where the better I configure the agent, the more I pay.”

Time to value matters alongside price. Buyers weighed a six-month setup resolving 50% of tickets against a one-month setup resolving 30%, and wanted clarity on whether implementation was self-serve or required the vendor's engineers. Ashley Wagner, VP of Customer Experience at Blackthorn, described how both vendor support and cost per resolution affected the economics of deflection.

Selection is only the start of testing. One buyer described running two agents side by side on 1% of traffic for three to six months before choosing. Buyers plan to keep monitoring after launch because agent quality can change over time. A pilot should establish both the service outcomes and the ongoing effort needed to sustain them.

G2's August 2026 AI agent buyer study of more than 1,000 respondents found that buyers now count AI search (17%) and evaluation sites (15%) among their key research sources. Seventy-three percent trust open research or independent test labs most, compared with 11% for vendor-published tests.

Buyers' most requested benchmarks include task completion (33%), safety and robustness (31%), and tool-calling reliability (29%).

The interviews reinforce that demand for evidence. Buyers want a third party that has actually run the agents, plus customer references they can speak with independently of the vendor. As PureGym's Head of Customer Experience put it: “I'd trust the numbers, but I'd ask to speak to one of their customers, ideally without the vendor present.”

“This is honestly a game changer... a reputed third party that has actually gone and implemented it, seen results, and knows what to do.”

Use independent evaluations to build a shortlist, then test the finalists against your own policies, integrations, approval rules, and cost model. The strongest benchmark result still needs to translate into a deployment that works for your customers and team.