The challenge
What was deployed
The results
Last Updated: November 20, 2026
Xero, the New Zealand-founded small business accounting platform, rolled a conversational AI service agent into the in-product help menu for all five million of its customers. Salesforce reports the agent now resolves 62% of support cases autonomously across more than 800,000 cases a year. The interesting part is not the model: it is the seven-year-old knowledge base and the zero-copy data layer underneath it. What the public record does not give us is a CSAT number, a cost-per-contact number, or a precise definition of ‘resolved’.
The company and the problem
Xero is a cloud accounting platform used by small businesses, accountants and bookkeepers. Salesforce’s customer story describes a company that has grown to serve five million small businesses and their advisors across more than 180 countries — a support surface area that scales with subscriber growth, not with headcount budgets.
Two things make Xero’s support load unusual. First, the questions are consequential: bank feeds, payroll runs, tax filings and account security are not casual retail queries, and a confidently wrong answer about a payroll cutoff has real financial consequences for the person asking. Second, the user is rarely a finance professional. Many are tradespeople or hospitality owners doing bookkeeping at 10pm, and the help they need is procedural rather than conceptual.
Xero’s answer since 2019 has been Xero Central, a consolidated self-service knowledge base. By the time agentic AI arrived, that estate was already doing heavy lifting: the pre-agent stack, built on Experience Cloud, Salesforce Knowledge and a Coveo-powered generative model, resolved roughly 97% of help sessions with personalised content. That is a crucial piece of context that most write-ups skip. Xero was not automating a broken support function. It was replacing a search-and-retrieve experience that already worked with a conversational one that could hold a multi-turn thread and hand off cleanly.
Providing customers with a digital experience that connects them to a human at the right time helps them unlock value from Xero faster.
— Nigel Piper, Executive General Manager, Customer Experience, Xero, Salesforce customer story
What was actually built
Three distinct systems are in production, and they are worth separating because they have different risk profiles and different evidence behind them.
1. The customer-facing service agent
Salesforce describes a conversational service agent deployed across the in-product help menu for Xero’s five million customers, built on Agentforce. It holds natural, multi-turn conversations, and when the customer needs a person it hands off via chat, callback or a logged case. Salesforce’s Agentforce metrics page adds a structural detail: the agent coordinates specialised agents for banking, payroll and security. In other words this is not one monolithic prompt with a knowledge base bolted on; it is a routing layer over domain-scoped sub-agents.
2. The SDR agent
On the commercial side, Xero built an agent that behaves like a sales development rep for accountant and bookkeeper partners. Per Salesforce, it uses firmographic data captured at sign-up, writes personalised outreach, answers product questions, books meetings with human SDRs, and keeps CRM fields current — lead classification, assignment, prioritisation, outreach and follow-up.
so people can spend more time in conversations that actually move the needle
— Ben Gilmore, Head of Customer AI, Xero, Salesforce customer story
3. The real-time engagement layer
Separately from the Salesforce estate, Xero’s customer engagement team rebuilt onboarding intervention on Twilio Segment. Speaking at Twilio’s SIGNAL 2026 event in Sydney and reported by CXFOCUS, Xero’s Vladka Kazda described roughly 50 live real-time activations triggered off product behaviour and models including churn propensity and next best action, plus a self-optimising content agent built in 90 days that generates email variations and scores them against downstream product adoption rather than opens and clicks.
simply shifting our abandonment comms from a 24-hour delay to real time delivered an immediate three to six percentage point lift in conversion
— Vladka Kazda, Global EGM and SVP Customer Engagement, Xero, CXFOCUS
How the support path runs
The entry point is the help menu inside the Xero product, not a separate chat website, so the user is already authenticated and in context.
Grounded in the Xero Central knowledge estate built since 2019.
Salesforce describes specialised agents for banking, payroll and security coordinated by the service agent.
Xero aggregates customer experience data in Snowflake and federates it into Data 360 with a zero-copy approach, giving agents a unified customer view without duplicating records.
When the customer needs a person, the agent hands off by chat, callback or a logged case rather than dead-ending.
The measured outcome
Two of those numbers deserve scrutiny rather than applause. The 62% figure appears on Salesforce’s Agentforce metrics page; the Xero customer story page carries a differently-worded stat card reading 65%+ online queries resolved autonomously. Those are almost certainly different denominators — cases versus online queries — measured at different dates. Anyone citing ‘62%’ as a benchmark should note that the vendor itself publishes two adjacent figures with different definitions.
The second is the comparison with the old system. A naive reading says Xero went from 97% to 62% and got worse. It did not. The 97% described help sessions resolved with personalised content in a self-service knowledge experience; the 62% describes support cases resolved autonomously in a conversational agent that can take actions and escalate. Different funnel stage, different unit, different bar.
| Dimension | Pre-agent stack | Agentforce service agent |
|---|---|---|
| Interaction model | Search and retrieve, generated answers over knowledge base | Multi-turn conversation inside the product help menu |
| Reported outcome | ~97% of help sessions resolved with personalised content | 62% of support cases resolved autonomously |
| Architecture | Experience Cloud, Salesforce Knowledge, Coveo-powered generative model | Agentforce with domain sub-agents, grounded via Data 360 |
| Escalation | Customer finds a contact path | Handoff via chat, callback or logged case |
What the public record does not establish
This is a vendor-published deployment. Salesforce has a commercial interest in the numbers, Xero has approved them, and no independent audit exists. Specifically:
No CSAT or quality figure is published for the service agent. Salesforce’s story says the handoff is seamless; it does not say what customers scored the experience, or how often an ‘autonomous resolution’ was followed by the same customer opening a case about the same problem a day later. Resolution rate without a reopen rate is half a metric.
No cost, headcount or handle-time figure appears anywhere in the public record. Xero has not said the agent saved X dollars or Y FTE-equivalents, and it would be irresponsible to infer one from 800,000 cases.
The 3x open rate is an early experiment. Salesforce frames it as the first six weeks of experimenting with the SDR agent. Open rates are the weakest signal in email marketing, inflated by image-proxy prefetching, and no downstream figure — meetings booked, partners onboarded, revenue — has been published.
The engagement-layer gains are not obviously LLM gains. Kazda attributes the three-to-six-point conversion lift to latency: moving abandonment comms from a 24-hour batch to real time. That is an event-streaming and orchestration win, which any team could bank without generative AI in the loop.
The model stack is undisclosed. Agentforce is the platform; which underlying foundation models serve which turns is not stated. Xero separately ships customer-facing AI with other partners, so the vendor landscape inside the company is plural.
What an integration team should learn
Buying the agent before you own the knowledge
The single most transferable fact in this case is chronological. Xero Central launched in 2019. The agent arrived in 2026. Salesforce explicitly credits years of investment in a centralised knowledge base as the repository of trusted proprietary content the agents now draw on. Teams that skip that step ship an agent that hallucinates confidently over a fragmented wiki.
Copying customer data into the agent platform
Xero keeps its customer experience data in Snowflake and federates it into Data 360 with a zero-copy approach, so the agent reads a single source of truth rather than a stale replica. The common failure is the opposite: a nightly ETL into the CRM, which guarantees that the agent occasionally tells a customer something that stopped being true yesterday.
One giant agent instead of a routed team
Salesforce describes Xero’s agent coordinating specialised sub-agents for banking, payroll and security. Those domains have different data, different compliance exposure and different escalation rules. A single prompt trying to cover all three gets long, gets contradictory, and becomes impossible to regression-test.
The sequencing is the strategy
Salesforce’s account of how Xero did it is a test-and-learn pattern: targeted experiments to validate agent performance before wider rollouts, and internal pilots — an employee assistant in Slack — before customer-facing use cases like the SDR agent. That order matters. Internal users tolerate a bad answer and report it; five million small businesses asking about payroll do not. Teams building similar workflow automation or agentic support systems should read the Slack assistant as the dress rehearsal that made the customer-facing rollout defensible.
Xero Central launches as the consolidated self-service knowledge base.
Generative answers on Experience Cloud, Salesforce Knowledge and a Coveo-powered model resolve about 97% of help sessions with personalised content.
Salesforce publishes the Xero customer story: SDR agent, Slack employee agent, and a conversational service agent rolled out across the in-product help menu for five million customers.
Salesforce reports 62% of support cases resolved autonomously across more than 800,000 cases a year; Xero’s CEO tells analysts one million customers used its generative AI features over 12 months, with monthly users up ninefold.
The wider context at Xero
The support agent is one slice of a much larger AI programme. On the product side, Xero ships JAX, its AI companion, and has been running XeroForce, a natural-language agent builder for accounting practices, in alpha. CEO Sukhinder Cassidy put adoption numbers on the record in November 2026.
One million customers have used our generative AI features over the last 12 months and monthly users are up ninefold year-on-year
— Sukhinder Cassidy, CEO, Xero, diginomica
For an integration team the relevant read is that Xero is running the same playbook in two directions at once: agents that reduce the cost of serving customers, and agents sold as product capability to those same customers. The internal deployment is the proving ground for the external one. If your own roadmap has an AI product on it, the support function is the cheapest place to learn what grounding, routing and escalation actually cost you. More deployment write-ups of this kind are collected in our case studies, and the model-selection trade-offs behind them in our models notes.
Autonomous resolution
DefinitionA conversation closed by the agent without a human joining or a case being routed to a queue. Note what it does not guarantee: that the customer was satisfied, that the answer was correct, or that they did not return through another channel. Vendors rarely publish the reopen-rate denominator, and Xero’s published figures are no exception.
Frequently Asked Questions
What did Xero actually deploy?
Three things. A conversational service agent in the in-product help menu for its five million customers, built on Agentforce and coordinating sub-agents for banking, payroll and security. An SDR agent that nurtures accountant and bookkeeper leads and books meetings. And a real-time engagement layer on Twilio Segment that triggers onboarding interventions off live product behaviour rather than batched exports.
How many support cases does the agent resolve on its own?
Salesforce reports 62% of support cases resolved autonomously, across more than 800,000 cases a year. The Xero customer story page separately shows 65%+ of online queries resolved autonomously and roughly 500,000 service agent conversations since launch. The two percentages use different units and were likely measured at different times, so treat them as indicative rather than precise.
Did the agent replace support staff?
Nothing in the public record says so. No headcount, cost-saving or handle-time figure has been published by Xero or Salesforce. The stated framing is channel optionality and immediacy, with escalation to a human by chat, callback or logged case. Any claim that the deployment cut jobs or saved a specific sum would be invention, not reporting.
Why did the resolution rate appear to fall from 97% to 62%?
It did not fall; the units changed. The 97% referred to help sessions resolved with personalised content in the previous self-service search experience. The 62% refers to support cases resolved autonomously by a conversational agent that handles multi-turn problems and escalates. Comparing them directly is a category error, and it is a common one when reading vendor stat cards.
What is the most transferable lesson?
The knowledge base predates the agent by seven years. Xero’s agents are grounded in a curated repository the company had already invested in heavily, with a zero-copy data layer supplying live customer context. Most organisations attempting the same deployment are actually attempting two projects: a content consolidation project and an agent project. Budget for both.
Twarx analysis
Original interpretationThe agent is the last 10% of the project. Xero Central launched in 2019; the conversational agent arrived in 2026 and is grounded in that curated knowledge estate plus a zero-copy live data layer. Organisations attempting the same deployment are usually attempting two projects — content consolidation and agent build — and only budgeting for one.
Read the numbers honestly
Analysis by
Rushil Shah · AI Systems Builder & Founder, Twarx