Hybrid customer service: how to combine AI and people in customer service

By Bernardo Gontijo, founder of NeftPublished on Updated on 7 min read

Summary

Hybrid customer service is an operation in which AI and people share the same service: AI resolves simple requests and prepares the case, and the team takes over what requires judgment. A study of 5,172 agents measured an average 15% productivity gain with AI assistance, and Brazil's 2026 SAC Decree requires a real-time human option when AI is used.

What hybrid customer service is

In this article, hybrid customer service is the model in which artificial intelligence and human agents work in the same operation, with the same protocol number and the same history. AI handles, classifies or supports the agent; the person takes over when the case requires judgment, negotiation or empathy, without the customer having to repeat the request.

The term also has two other meanings: service that combines in-person and remote channels, and hybrid work arrangements for agents. All three can coexist, but here the focus is the division of work between AI and people.

Why not automate everything

A high rate of conversations resolved by AI can hide little work saved. The easiest requests tend to be the most frequent and shortest; the ones that consume the team's time, such as disputes, exceptions and cases with a long history, keep reaching people.

Marty, co-founder and CEO of Pylon, a company that sells support software, described this effect at SaaStr AI Day. He told the story of a company with about a thousand people in support that adopted an automatic resolution agent:

“It deflects around 50% of their tickets, but their headcount didn't change.”

Marty, Pylon, SaaStr AI Day, from 4:54. Original quote: “it deflects around 50% of their tickets, but their headcount didn't change”. Vendor account, not an independent study.
Video: SaaStr AI Day: The New Agentic Support Playbook (SaaStr AI, YouTube, ). Published on July 22, 2026; playback starts at the passage on ticket deflection, at 4:54. Talk in English.

According to him, the reason is that automation resolved the easiest questions, which took the least work. Pylon sells tools to support agents, so the account also argues for its product. The outcome figures cited in the talk come from the company's customers and were not audited, so they are not used here.

What research shows about AI supporting agents

The most cited study on the subject is Generative AI at Work, by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, published in the Quarterly Journal of Economics in 2025. The authors followed 5,172 customer support agents as an AI assistant was rolled out in stages.

Access to the assistant increased issues resolved per hour by 15% on average. Less experienced and lower-skilled agents improved both speed and quality, while the most experienced saw small speed gains and slight quality declines. The authors also observed more polite customers and fewer requests to speak to a manager.

The study evaluates AI that suggests responses to people, at a single software company. It does not measure an AI that serves customers on its own. Even so, it reinforces a point of the hybrid model: supporting the team can pay off as much as replacing conversations.

How to split the work between AI and the team

The split starts from the type of request and the consequence of an error, not from the technology available. A practical way is to separate requests into three tiers:

TierWho leadsExamples
Informational and repetitiveAI, with a human option at any timeOpening hours, required documents, steps of a request, protocol status
With an action in a systemAI with limited permission or human approvalScheduling, record updates, duplicate bill after identification
With judgment or riskTeam, with AI preparing the summaryBilling disputes, negotiation, serious complaints, urgency
Example split; each operation defines the tiers according to its own policy and regulation.

The boundary between tiers changes over time. Review transfers and reopened cases every month to move requests between tiers based on what actually happened.

The handoff is the critical point

The hybrid model fails when the handoff from AI to a person loses information. The customer repeats the problem, the agent starts from scratch and the advantage of automation turns into rework. A good handoff carries the protocol number, a summary of what was said and what has already been checked.

In Brazil, this is also a rule for suppliers covered by Decree No. 13,135/2026. The decree allows AI in triage and resolution as long as there is an option for real-time service by a person (art. 3, § 3), prohibits asking consumers to repeat their request (art. 6) and requires continuity until conclusion (art. 7). See the summary in new SAC Decree.

Test this handoff as part of the process, not just the final outcome. In the October 2026 preprint CXBench, 80.1% of AI agent runs with a correct outcome breached some process requirement in simulated regulated tasks.

Metrics to track

  • Resolution by tier: how much the AI resolved alone and how much the team resolved after the handoff.
  • Handoffs with context: share of cases in which the agent did not need to ask for the information again.
  • Reopening: requests that come back within a few days, separated by who handled the first contact.
  • Time to a person: wait between asking to talk to someone and the start of human service.
  • Satisfaction by path: ratings from customers served only by AI, only by the team or by both.

Compare these numbers with the operation before AI. A high automation rate with high reopening is a sign that the boundary between tiers is wrong.

How Neft organizes service with AI and people

At Neft, AI customer service works within the same queue, protocol number, recording and reports as the team. The virtual agent talks by voice, follows the limits set by the company and, when the customer asks for a person or the case falls outside the scope, passes the service to the team with a summary of the conversation.

The company chooses when the AI handles service, for example outside business hours or when no agent is free, and which actions it can perform. Cases that call for a face-to-face conversation continue in video service with the team, and reports show metrics by period and by agent.

Frequently asked questions

What is hybrid customer service?
It is the model in which AI and human agents work in the same operation, with the same protocol number and history. AI resolves simple requests or supports the agent, and the person takes over cases that require judgment, without the customer having to repeat the request.
Is hybrid customer service required by law?
Brazil's Decree No. 13,135/2026 does not require AI, but, for covered suppliers, it only allows automation if there is an option for real-time service by a person, from 8 a.m. to 8 p.m., every day. In practice, anyone using AI in the SAC must operate in a hybrid way.
Will AI replace agents?
Available studies point another way. In the study of 5,172 agents published in the Quarterly Journal of Economics, AI that supported people increased issues resolved per hour by 15%, with a larger gain among novices. Complex cases still require the team.
Which requests should AI handle?
Start with informational and repetitive ones, with clear rules and low risk, such as opening hours, documents and protocol status. Actions in systems call for limited permission or human approval. Disputes, negotiation and urgent cases should stay with the team.
How do you avoid the customer repeating everything at the handoff?
Make the handoff carry the protocol number, the conversation summary and what has already been checked. The agent should start where the AI stopped. Measure the share of handoffs in which nothing had to be asked again.

Sources

Videos, studies and documents used in this article. Preprints are versions published by their authors before peer review. Article researched and written with AI assistance. Study summary checked on NBER, the preprint on arXiv, the decree on the Planalto website and the video's date and channel on YouTube on November 18, 2026. The library transcript was used to locate the quoted passage. The tier split is an editorial recommendation, not a measured result.

  1. Research paper

    Generative AI at Work (opens in a new tab)

    Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond (NBER Working Paper 31161; Quarterly Journal of Economics, vol. 140, no. 2, pp. 889–942) · published on

    5,172 support agents; average increase of 15% in issues resolved per hour, with larger gains among less experienced agents. Figures from the published journal version. Date of the journal volume; working paper from April 2023.

  2. Video

    SaaStr AI Day: The New Agentic Support Playbook (opens in a new tab)

    Marty, co-founder and CEO of Pylon; SaaStr AI channel · published on

    Date and channel checked on YouTube. Account from 4:54 on deflecting about 50% of tickets without a change in headcount. Support software vendor; unaudited account.

  3. Research paper (preprint)

    CXBench: A Benchmark for Process-Compliant LLM Agents in Regulated Customer-Experience Workflows (opens in a new tab)

    Boxi Yu and nine other authors · published on

    Version 1. 80.1% of runs with a correct outcome breached process requirements in simulated regulated tasks.

  4. Legislation

    Brazilian Decree No. 13,135/2026 (opens in a new tab)

    Office of the President of Brazil · published on

    AI allowed with a real-time human option (art. 3, § 3), ban on asking to repeat the request (art. 6) and continuity (art. 7).

Want to evaluate AI in your customer service?

Request a demo and assess the workflow with fictional data from your operation.