Observable, auditable, defensible: How FloQast builds AI that accountants trust with Grafana Cloud

When you’re trying to close out your organization’s monthly finances, the last thing you want is uncertainty. Accounting teams at more than 3,500 companies, including Lululemon, Chipotle, and Shopify, rely on FloQast, an AI-powered accounting platform, to reconcile accounts, match transactions, and keep the close on schedule. Increasingly, AI agents do the prep work of drafting journal entries, tying out balances, and flagging exceptions for review. But when an AI agent touches a company’s general ledger, accountants can’t just assume it was all done properly.

“Our AI story is fundamentally a trust story,” says Ankush Sharma, Head of Infrastructure (SRE & AI Engineering) at FloQast. “Agents that touch financial workflows have to be observable, auditable, and defensible. If we can’t reconstruct what an agent did and why, we haven’t earned the right to put it in front of an accountant.”

FloQast holds its infrastructure to the same standard an auditor holds an accountant, and that standard shapes how Sharma’s team builds and operates its platform. Today, FloQast runs its reliability practice on Grafana Cloud, with more than 250 SLOs anchored to the accounting calendar, full-stack visibility from the accountant’s browser to the AI agent’s decision trace, and incident investigations that resolve 10% faster than they did before adopting Grafana Assistant.

Built by accountants, held to an auditor’s standard

FloQast was founded in 2013 in Los Angeles by former accountants who wanted a better way to manage the monthly process of finalizing a company’s financials—known in finance as the “month-end close,” or simply “close.” The company has since grown into an AI-powered accounting transformation platform that helps accounting teams manage and automate the critical work involved in closing the books, ensuring financial accuracy, and preparing reliable financial reporting. Its pitch to customers is blunt: AI you’d stake your license on. Every output is explainable, every decision is logged, and sign-off always stays with a human. The platform is ISO 42001 certified and SOC 2 Type II compliant.

The accounting calendar gives Sharma’s team an unusual reliability profile. Thousands of companies close their books during the same handful of days at the start of each month, so FloQast has to be prepared to handle that spike in usage at a time that’s critical for its users. A degraded Enterprise Resource Planning (ERP) sync on day two of the monthly close is not an inconvenience; it is a controller’s missed deadline.

“Our customers work against a hard calendar,” says Sharma. “So our reliability targets follow that calendar. The platform has to be at its best precisely when it is under the most stress.”

SLOs that speak the language of the ‘close’

Accountants feel the close. FloQast builds its SLOs around exactly those moments— when financial data needs to sync, calculations need to finish, transactions need to post, and supporting documents need to be uploaded. Grafana SLO turns those journeys into SLIs with error budgets, burn-rate alerts, and dashboards the whole engineering organization shares.

The team currently tracks 250+ SLOs across 90+ services, covering 70% of critical user journeys.

“A wall of green infrastructure metrics means nothing if a reconciliation hangs on day three of the close,” says Sharma. “We define SLIs around what our customers experience, and we spend error budget the way accountants spend time during close week: deliberately.”

Error budgets also settle a perennial argument about resource consumption. When a service has budget to spare, teams ship. When the budget burns down, they harden. “It takes the debate out of the room,” Sharma says. “The data decides.”

Shifting from reactive to proactive

As organizations close out their monthly accounting, the real danger is the problems that accountants don’t report. When a controller is three days into the close-out process and hits a slow page, she works around it. But because a ticket is never  filed, FloQast never knows the experience was degraded. Multiply that pattern across thousands of concurrent users at the start of every month, and you get a reliability blind spot.

Grafana Cloud Frontend Observability changed what FloQast can see. The team now monitors errors, page loads, and Core Web Vitals across every real user session. A bad release or a degraded endpoint surfaces as a signal within minutes, typically scoped to a specific browser, region, or feature, while the impact is still small and the on-call engineer still has time to contain it.

“The biggest change is that support tickets stopped being our detection mechanism,” says Sharma. “By the time a ticket arrives, it’s usually confirming something we’re already on. The quiet failures nobody reports are exactly the ones we catch now.”

That shift in posture changed the question FloQast asks about its own platform. “It moved our definition of reliability out of the data center and into the accountant’s browser,” says Sharma. “We stopped asking ‘are the services up?’ and started asking ‘is the close fast?’”

Answering that question accurately requires a population view, not a case file. A single session tells Sharma’s team what happened to one controller. The aggregate—how many customers are affected, what the p95 experience looks like at peak close, which browser or region is carrying the most pain—tells them whether it’s an outlier or the leading edge of something systemic. It also corrects the bias built into every ticket-driven system: the customers who file reports aren’t a representative sample; during close week, when everyone is racing a deadline, they’re the least likely to stop and write one up.

“One user’s bug is an anecdote,” says Sharma. “The population view is epidemiology. We prioritize based on what moves the real experience for thousands of users, not on whoever reported loudest.”

When a signal warrants a deeper look, the evidence chain is continuous. Engineers move from a performance alert to the session where it happened. From there they can find the exact failing request, as well as the corresponding backend trace and the service behind it—same trace IDs, same timeline, one platform. For a company whose entire value proposition rests on auditability, a broken chain of custody is a structural problem.

“With a separate frontend tool, that chain breaks at the tool boundary,” says Sharma. “You’re stitching timestamps across systems, and every incident starts with a debate about whose data is right. For a company whose whole story is about auditability, one unbroken chain of evidence from click to database isn’t a convenience—it is a must-have.”

Since adopting Frontend Observability, FloQast has reduced triage time and escalation round-trips by 25%.

Root cause at 2 a.m., with an agent on call

FloQast ships AI agents to accountants. Its SRE team runs one of its own. Grafana Assistant lets engineers across the company query telemetry, build dashboards, and explore incidents in natural language—pulling observability out of the specialist tier and into the hands of product engineers who can now self-serve answers that used to require a ticket to the platform team.

During incidents, Grafana Assistant Investigations does the first pass. The SRE agent correlates metrics, logs, and traces, surfaces suspect signals, and proposes a root cause that the on-call engineer verifies rather than hunts for. FloQast has seen investigation time drop by 10% since adopting it.

“We hold Grafana’s agent to the same standard we hold our own,” says Sharma. “It shows its work. I can see the queries it ran and the evidence behind its hypothesis. That’s what makes it usable at 2 a.m., when trust is the scarcest resource in the room.”

An audit trail for the agents themselves

The deepest layer of FloQast’s trust story is observability into the AI itself. At the product level, FloQast already logs every agent decision so customers can defend their close to auditors. Grafana Cloud Agent Observability extends the same defensibility to the engineering layer.

With Agent Observability, Sharma’s team traces agent workflows end to end: prompts, completions, tool calls, retrievals, latency, token spend, and error rates. It’s all broken down by model and by agent, in the same Grafana Cloud stack that holds the rest of the company’s telemetry. When an agent’s output drifts, when latency creeps, or when the cost per workflow moves, the team sees it against the same SLO discipline it applies everywhere else.

“Traditional observability tells you whether the system is up. Agent Observability has to tell you whether the system is right, and what it costs to be right,” says Sharma. “For an agent drafting a journal entry, we want the whole chain: which model, which prompt version, what it retrieved, how long it took, and what it did when it was uncertain. Auditors ask accountants to show their work. We ask the same of our agents.”

This is also where FloQast’s engineering roadmap and Grafana Labs’ product direction converge. Sharma wrote the book on the subject, literally: “Observability for Large Language Models: Site Reliability and Chaos Engineering for AI at Scale.” In it, he argues that SRE discipline must extend beyond infrastructure into model behavior: SLOs for latency and quality, error budgets that balance model accuracy against shipping velocity, and chaos experiments that probe how agentic systems fail. “Grafana Cloud is where those ideas stop being theory for us,” he says. “The direction they are taking with Assistant, Investigations, and Agent Observability maps almost one to one onto how we think reliability for AI should work.”

Giving agents eyes on production

The newest addition to the stack is gcx, Grafana’s CLI built for agentic workflows. FloQast engineers use it to pull dashboards, alert rules, and SLOs into code, and to give their AI coding agents structured access to production telemetry.

“An agent fixing a bug without production context is pattern-matching on source files and hoping,” says Sharma. “With gcx, the same agent can read the actual state of the running system before it proposes a change. That is the difference between an assistant and a liability.”

What’s next

The next frontier for Sharma’s team is deliberately breaking their AI agents before customers find the failures—chaos experiments against agent workflows that test failure modes traditional resilience testing doesn’t surface. Alongside that, FloQast plans to wire evaluation pipelines directly into SLOs so model quality burns error budget the same way downtime does, treating a degraded agent response the same as a missed SLA. Broader adoption of Agent Observability as it matures is also on the roadmap.

“The financial close is a monthly exam with no retakes,” says Sharma. “Trust isn’t something you ship once. It’s an SLO you defend every month, and Grafana is how we defend it.”

Instrument your AI stack. Start free at grafana.com/products/cloud