The first version of Balto was not artificial intelligence. It was an Excel macro built by a young telesales rep who kept discovering an expensive truth: training is easy to remember in a quiet room and remarkably hard to retrieve when a stranger is objecting in your ear. In 2017, Marc Bernstein typed an objection into a spreadsheet and made the right response appear. His colleague Chris Kontes saw the trick and recognized a company hiding inside it.
That tiny contraption answered the question that still defines Balto: what if software helped before the conversation went wrong? Most conversation-intelligence tools were rearview mirrors. They recorded calls, generated transcripts and showed a manager the wreckage later. Balto put the instruction on screen while the customer was still there.
The five-second company
Balto Software, Inc. now sells enterprise software to contact centers. Its core Agent Assist product listens to both sides of a live conversation, recognizes what is being discussed and displays the next useful thing: a required disclosure, an objection response, a checklist item, a knowledge-base answer or a reminder to ask a question. The agent remains the speaker. The machine is the stagehand sliding a cue card into view.
Around that live cue, Balto has built a wider system. Notes summarizes the call. Quality Assurance scores interactions against configurable criteria. Compliance flags language that may deserve review. Coaching finds useful moments and prepares material for supervisors. Insights lets leaders ask questions across the conversation pile. Newer voice-agent and omnichannel products extend the system beyond its original human-on-a-phone frame.
The customers are businesses where one sentence can change revenue, compliance or trust: insurers, health systems, banks and credit unions, collections firms, business-process outsourcers, retailers and home-services operators. Balto publicly displays names including Staples Canada, Chubb, RingCentral, Empire Today, Watches of Switzerland and Integris Health. It says the platform serves more than 300 organizations and has guided more than 500 million conversations.
The product is built for the five seconds between a customer's objection and an agent's answer.The Balto thesis, reduced to its useful core
What failed first
Traditional training failed first. Bernstein and Kontes had heard the lessons. Under pressure, they forgot them. Later customers described versions of the same problem at industrial scale: supervisors sampling a sliver of calls, scripts going stale, new hires asking managers for help and post-call analytics diagnosing a mistake when nothing could be done about it.
ONLINE Information Services, a collections business, had a post-call system that was complicated enough to require technical certification. It could explain calls, but not guide a collector through one. During the pandemic, the company also needed to change its language quickly as people faced financial stress. Balto's playbooks could be revised and delivered to remote agents. The case study reports compliance-score averages rising from below 40 percent to above 86 percent in three months, plus more than 15 fewer hours of weekly management work.
A flooring company gave Balto a cleaner test. It put a random subset of sales reps on the product and kept the rest off, then watched save rate. Balto says the guided group improved that rate by 26.1 percent after one month and preserved $3.2 million in revenue in three months. That changed the customer's mind. It expanded Balto from sales into customer service and quality assurance, where every call could be scored instead of roughly one percent.
These are vendor case studies, not universal promises. But the shape of the rollout matters more than the glossy number. Balto did not ask the customer to believe in a sweeping AI transformation. It started with one group, one live workflow and one metric connected to money. The broader platform earned its way in.
What it cost - and what it sells
The founding trio - Bernstein, Kontes and Davidson Girard - pooled $210,000 of personal savings, according to Bernstein's account, and began in a 75-square-foot coworking office in downtown St. Louis. They kept enough money aside to imagine two years without salaries. Their stake lasted 18 months, long enough to produce a market-ready product and a handful of customers.
Outside capital then arrived in layers: roughly $1.2 million to $1.3 million in 2018, $3 million in 2019, a $10 million Series A led by Sierra Ventures in 2020 and a $37.5 million Series B led by Stripes in 2021. RingCentral Ventures joined that last round. Balto announced approximately $52 million raised in total. No public valuation is available.
Customers do not get a public menu with a tidy price per seat. Balto's terms describe enterprise SaaS subscriptions purchased through order forms, support included with the service, and optional professional work covered separately. In plain English: custom contracts, likely shaped by products, scale and deployment. This is not a self-serve browser tab for a three-person support desk. It is software that must connect to telephony, workflows, scorecards, policies and the politics of how people are evaluated.
Balto calls its employees Baltonians and the group a pack, a bit of canine branding that could have become unbearable but at least commits to the joke. Its public values favor gratitude, growth and ownership. The company has described 72-hour engineering hackathons, a women-focused employee group, a diversity council and a mentorship program for colleagues learning to code. More consequentially, Balto says more than half the team has worked in contact centers, sales, service or collections. That is useful scar tissue for anyone designing a prompt an agent must tolerate during a difficult call.
St. Louis is part of the pitch, too. Balto has said it wants to do for the city what Dell did for Austin: build a technology company that helps thicken the local ecosystem around it. The workforce is distributed, but the address and ambition remain in Missouri. For an enterprise AI market often narrated from San Francisco, that choice makes Balto an instructive regional story as well as a product one.
The moat is not the transcript
Speech-to-text is widely available. So are large language models. Balto's differentiation has to live elsewhere: low-latency delivery, the ability to embed in major contact-center systems, configuration that reflects a company's own policies, and the operational memory built from seeing which prompts help. Its listed integrations include RingCentral, Five9, Genesys, Amazon Connect, Talkdesk, Twilio, Salesforce and others.
Timing remains the sharpest distinction from post-call revenue-intelligence tools such as Gong or Chorus. The closer alternatives - Observe.AI, Cresta, NICE, Verint, CallMiner and Level AI - increasingly offer overlapping QA, analytics and agent-assist features. Balto's market is no longer a category of one. Its response has been to join the functions around the call in one license and to keep expanding from advice into completed work.
That expansion brings tension. Balto's founding metaphor is the sled dog that guides while a human remains in charge. The company now also sells Togo voice AI agents. Helping a worker and replacing a defined slice of work can coexist, but buyers should ask where the handoff sits, who reviews exceptions and whether customers know when they are talking to software.
When this does not work
Real-time guidance is a poor fit when calls are rare, entirely bespoke or impossible to reduce to a repeatable process. It also disappoints when source material is stale, integrations lag, leaders treat prompts as surveillance, or nobody owns the playbook after launch. A wrong cue delivered instantly is still wrong - only faster.
The part worth stealing
Start with retrieval, not intelligence. Find the precise moment a capable person knows something but cannot access it quickly enough. Build the smallest intervention that changes that moment. Then choose a metric close to the outcome - save rate, required disclosure, handle time, first-call resolution - and run a pilot with a comparison group.
Next, use the exhaust from that workflow to earn adjacent products. Balto's live transcript could become a score, a note, a coaching clip, a compliance alert and a trend line because each was another interpretation of the same interaction. That is a more disciplined platform strategy than launching six unrelated AI features and hoping customers assemble the story.
Finally, preserve the human factors. Balto says more than half its employees have experience in contact centers, sales, service or collections. That proximity matters because the enemy is not merely model error. It is the prompt that blocks the screen, the checklist that rewards robotic phrasing, the score nobody trusts and the supervisor who uses a coaching tool as a cudgel.
Balto's best idea remains its first one: help while help can still change the outcome. The next 500 million conversations will test a harder idea - whether a growing AI workforce can recognize the equally valuable moment to stay quiet.