Crescendo Bought a 3,000-Person BPO

Key points

• AI scores every support conversation, and low-rated cases are taken off the bill.
• Fixes must be verified by AI and approved by a person before they reach the system.
• New roles handle AI deployment, incident response, and tuning.

An AI company that began by buying a call center

A typical enterprise software startup sells a system to customers and collects usage fees. A traditional outsourcing firm does the opposite. It hires people to do the work for a company and charges by hours worked or headcount. Crescendo, a customer support company that launched in January 2024, took a path that is neither. A small artificial intelligence (AI) software company of about 20 people bought a huge outsourcing firm with about 3,000 support agents.

Venture capital firm General Catalyst announced Crescendo’s launch on January 25, 2024. It described the company as “one of the first vertically integrated, AI-native contact center businesses.” The founders’ backgrounds are unusual too. Co-founder Andy Lee founded Alorica, a large contact center outsourcer. Matt Price ran product and sales at Zendesk, a customer management software company. Slava Zhakov was chief technology officer of Genesys, a contact center software company. They had deep experience of how both the software industry and the support outsourcing industry work from the inside.

The first plan was an AI layer that writes playbooks for handling repetitive requests. The company says its platform has been in production use since late 2023. On October 2, 2024, it raised $50 million at a $500 million valuation. On the same day, it bought PartnerHero, a customer support outsourcing company (a BPO) headquartered in Boise, Idaho. A BPO is an outsourcing service that takes a company’s phone, chat, and email inquiries and handles them for it.

The deal gave Crescendo about 3,000 support staff and more than 200 customers, including Airtable, Khan Academy, and Udemy, all at once. Matt Price later told Pulse 2.0 that Crescendo was then “a startup of twenty people.” It was not an existing call center adopting AI. It was a company designed around AI from the start that swallowed a whole human operation. In their joint letter, the CEOs of the two companies described the combination as “neither a stand-alone AI SaaS tool nor a BPO, but the best of both.”

Crescendo and the staff it bought

The first response comes from a machine, and people back it up

In Crescendo’s actual operation, an agent called Concierge stands at the very front. Concierge takes customers’ first questions directly, whether they come by voice, chat, email, or text. Its goal is not just to give guidance but to solve the problem itself.

When the agent cannot find an answer, it hands the case to a human agent “with the full thread intact.” The people waiting behind are support staff Crescendo hired itself. The company calls them Superhumans. According to the pricing terms published in October 2025, this human handoff team was by default shared across customers and supported English and Spanish. A team dedicated to one customer, or staff for other languages or countries, cost extra.

Human agents do not start from a blank screen. A tool called Expert Assist reads the customer’s procedures, internal policies, and past case history and writes a finished draft that can go straight to the customer. The company’s product page describes the agent’s role in one plain sentence: “Agents review, send, and move on.” Agents polish the replies AI wrote and take on hard, complex problems. They also shape the overall customer experience and keep up to date the internal knowledge base that the AI reads and learns from.

Every conversation is checked, and fixes need a person’s permission

After the work is done, quality control runs as a chain of four agents, one after another.

First, a Quality Agent scores every conversation with customers, without exception. It does not pick a sample to listen to. It grades them all. Next, an insights agent collects the conversations with low satisfaction or failed resolution, groups them by common type, and traces the causes. An optimization agent drafts fixes based on the causes it finds. Finally, a simulation agent tests the fixes in advance against “1K customer days” of support data.

But passing the test does not let the AI change the system on its own. According to IT Brief, any change to the system must be approved by a person before it is actually deployed. Chief Product Officer Tod Famous put the principle this way: “Self-improving doesn’t mean self-authorising.” On its website, the company calls the standard that separates what AI may do alone from what needs a person’s permission the Safe Action Framework. Think of a self-driving car. It may simulate its route thousands of times, but a person still presses the switch to enter the real highway.

The company’s job postings show three roles that stand out, and their descriptions show the division of labor between people and agents.

The first is the AI Platform Support Specialist, hired for remote work in Honduras and for US-based staff. This is the incident-response role that steps in first when an AI error affects customers. These people look across AI settings, workflows, routing rules, and links to outside tools like Zendesk and Salesforce to find why something failed. When they find the problem, they adjust the settings and add checks so the same problem does not happen again. The job posting calls the role “the control layer between production AI systems, partner environments, and internal teams.” It says that with experience, people in the role will manage “a portfolio of partner AI assistants” across several customers.

The second is the AI Deployment Specialist. After a contract is signed, this specialist builds and connects an AI assistant tailored to the customer. They run real tests with the customer and write the procedures and training materials for the human agents who work behind the AI.

The third is the Real-Time Coordinator. Based in Honduras, this person tracks agent schedules, breaks, and queue demand across chat, email, and voice channels in real time. This is typical workforce management in a traditional call center, and it is still an important part of how Crescendo runs.

The speed of getting started is standardized too. The company promises to launch the service within 30 days of the contract. First it sets the scope of use cases to handle, then it cleans up the customer’s knowledge documents, then it trains and tests the assistant. After that, it opens the system channel by channel. Even after launch, it keeps monitoring results and retraining.

It charges per resolved case, not per hour

Crescendo’s pricing differs from the old outsourcing habit of charging for time. It puts a price on results that were actually resolved, not on the hours agents worked.

In October 2025, the pricing page offered a single rate starting at $2.99 per resolution. It included 24-hour human agent handoff, quality checks on every conversation, and knowledge base management as standard. By May 2026, the pricing had split into two tiers. The “Managed AI” tier charged a base monthly service fee starting at $2,900, plus $1.25 per solve. For the “Managed AI & Superhumans” tier, which adds human agents, the company stopped publishing a price and asked customers to talk to sales. The current website has dropped all specific prices.

A quality guarantee underlies the pricing. If the customer satisfaction score (CSAT) on AI-resolved conversations is not higher than the score of the customer’s previous outsourcing provider within 30 days of launch, that month of service is free. Also, a conversation that Crescendo’s own model predicts will get a satisfaction score of 2 or lower (“Predictive CSAT”) is automatically credited on the next month’s bill. The company says the guarantee applies equally to conversations resolved by AI and those resolved by human agents.

It is not clear, however, how the overall work splits between people and AI. A testimonial from one customer on the company website says AI resolved 70% of support and specialist agents handled the other 30%. But that is the figure for one customer, not a company-wide average. How the roughly 3,000 PartnerHero employees were reassigned, and in what proportions, has not been made public either.

How this differs from the old outsourcing model

Crescendo works differently from a traditional customer center in four ways.

The first is how much quality is checked. In a traditional call center, a manager or quality staff member picks a few calls from each agent, listens, and scores them. In a 2022 report, market research firm ISG Research pointed out that this kind of sampling scores “as little as 1% to 3% of the total.” A 2016 survey by the call center trade publication Call Centre Helper found that about two-thirds of its 113 respondents evaluated 2% or less of all interactions. At Crescendo, machines evaluate every conversation, and people focus on looking at the common patterns the AI picks out and approving fixes. On top of that, the score ties directly to the bill, so dissatisfied conversations lead to credits.

How support conversations are quality-checked

The second is what the customer pays for. Many BPO contracts are still billed by the number of people put on the job (FTEs) or by labor hours. According to a survey by IT research firm HFS Research, reported by Analytics India Magazine in February 2026, reliance on contract models based on labor input stands at 42% today and is expected to fall to 28% within three years. The report noted that “procurement teams remain wedded to predictable, FTE-led pricing.” Crescendo bills per resolved case, not per labor hour, and bundles human agents into the package. The company calls this a break from “the multi-decade policy of charging for labor.” Under hourly billing, the outsourcer earns more when support work grows and drags on. Under Crescendo’s model, it earns only when it solves the problem, and a conversation the customer was unhappy with becomes a loss.

The third is the role of human agents. In a traditional BPO, an agent’s basic job is to work through inquiries in a queue that never stops filling. At Crescendo, AI takes the first pass on every inquiry. People solve the complex cases the AI could not handle and passed on. They polish and send the drafts the AI wrote, and they fix the knowledge documents. Going further, people take the system-management roles: deploying AI for each customer, responding to malfunctions, and tuning the AI. Crescendo says it is training its agents to become “skilled AI practitioners,” but it has not said how many have actually moved into such roles.

The fourth is how the organization was built. AI startups in this field usually either sell software to call centers or compete with outsourced centers. Crescendo, with about 20 people, instead bought a traditional company with about 3,000 agents and more than 200 customers. General Catalyst calls this the “Venture Buyout Model.” The AI startup got experienced front-line staff, customers, and a global operating base on the first day. The outsourced workforce got software infrastructure designed from the start with AI as the front door.

What this approach teaches if it shakes the industry

If this way of operating lasts, several pillars of the outsourcing industry could be shaken.

First, billing by time and headcount could come under heavy pressure. If the market accepts per-resolution pricing tied to customer satisfaction, existing vendors that earned revenue in proportion to agents’ working hours will have to explain why customers should pay for time and not for results. As the HFS Research data shows, the long habits of corporate buying departments slow the change, but the direction is moving.

Second, scattered software tools could merge into one. The founders called the setup in which an existing chatbot and a call center outsourcer coexist an “adversarial operating model.” The software vendor that builds the chatbot and the call center that answers the phone compete to take the same customer inquiry. Crescendo proposes to replace the chatbot, the call center phone system, the ticketing software, the quality evaluation tool, and the workforce management solution with a single platform. If the idea works, both software companies that charged per-seat license fees and traditional BPOs could lose part of their business.

Third, the trend of AI companies buying service firms, instead of selling them software, could spread. An AI company buys an outsourcing firm to gain customers and on-the-ground expertise, then rebuilds its internal work around AI.

Operators in other industries can also borrow a few principles from this division of labor.

First, to price by results, you must first score all of the work. Crescendo can offer outcome-based billing and credit back dissatisfied cases because it has an automated system that checks every conversation. An organization that looks at only 2% of its workload as a sample would find it hard to offer such a guarantee.

Second, keep final approval with people. Even if AI analyzes errors, drafts fixes, and runs large-scale simulations, people hold the authority to release them into the live system. Even when you build a cycle that learns and improves on its own, you still need a gatekeeper who decides what gets deployed.

Third, give specific names to the new roles of people who work with AI. Crescendo posted separate jobs for investigating agent failures and for putting AI into customer environments and testing it. A vague phrase like “a human steps in” does not set up responsibility or a division of labor on the ground.

Fourth, ask who holds the scoring standard. If fees depend on results or satisfaction, customers should be able to see clearly how the model that sets those scores evaluates conversations.

What is still unproven

Crescendo’s way of working is not fully proven. First, much of the performance data the company gives rests on its own announcements. The claims of a 70% resolution rate from day one, an average dissatisfaction rate of 0.68%, and 99.8% answer accuracy have no publicly disclosed, objective measurement method. Predictive CSAT, which sets the credit amounts, is not an externally recognized metric either. It is an algorithm the company built itself.

Also, the exact split between AI and people across the whole company, the current total headcount, and what the roughly 3,000 PartnerHero agents do now have not been made clear.

At a basic level, views also differ on whether to see Crescendo as a purely AI-native company or as a huge traditional BPO in the middle of being rebuilt. It did start with AI technology first. But much of its operating staff and its first large customer base came from the outsourcing company it bought. The statistic that the old quality-control sample is only 1% to 3% rests on a 2022 analyst report and a small 2016 survey, and the sample and method of the HFS Research survey were not disclosed.

---

Sources

1. General Catalyst — Our Creation of Crescendo — https://www.generalcatalyst.com/stories/our-creation-of-crescendo — 2024-01-25
2. Crescendo — Crescendo – the next episode! — https://www.crescendo.ai/news/crescendo-the-next-episode — 2024-10-02
3. Crescendo — Crescendo Closes $50M in Financing with Latest Round Led by General Catalyst (press release) — https://www.crescendo.ai/news/crescendo-closes-50-million-in-total-financing-with-latest-round-led-by-general-catalyst — 2024-10-02
4. Crescendo — PartnerHero + Crescendo = Augmented AI (Wayback Machine snapshot, 2025-02-16) — http://web.archive.org/web/20250216213912/https://www.crescendo.ai/news/partnerhero-crescendo-augmented-ai — 2024-10-02
5. Crescendo — Pricing (Wayback Machine snapshot) — http://web.archive.org/web/20251010003209/https://www.crescendo.ai/pricing — 2025-10-10
6. Crescendo — Pricing (Wayback Machine snapshot) — http://web.archive.org/web/20260510091409/https://www.crescendo.ai/pricing — 2026-05-10
7. Crescendo — Human in the loop (Wayback Machine snapshot) — http://web.archive.org/web/20260615023353/https://www.crescendo.ai/product/human-in-the-loop — 2026-06-15
8. Crescendo — AI Optimization (Wayback Machine snapshot) — http://web.archive.org/web/20260510074519/https://www.crescendo.ai/product/managed-service — 2026-05-10
9. Crescendo — Operational AI (Wayback Machine snapshot) — http://web.archive.org/web/20260510080817/https://www.crescendo.ai/product/operational-ai — 2026-05-10
10. Crescendo — homepage — https://www.crescendo.ai/ — accessed 2026-10-22
11. Crescendo — Total Outcome Guarantee — https://www.crescendo.ai/total-outcome-guarantee — accessed 2026-10-22
12. Crescendo — AI Platform Support Specialist (job posting, Honduras) — https://jobs.generalcatalyst.com/companies/crescendo-2-1fc09d6e-7427-4a85-8af7-a6d2cbfc5622/jobs/68610160-ai-platform-support-specialist — accessed 2026-10-22
13. Crescendo — AI Deployment Specialist (job posting) — https://jobs.generalcatalyst.com/companies/crescendo-2-1fc09d6e-7427-4a85-8af7-a6d2cbfc5622/jobs/52760092-ai-deployment-specialist — accessed 2026-10-22
14. Built In — AI Platform Support Specialist (Crescendo job posting, US, removed 2026-02-03) — https://builtin.com/job/ai-support-specialist/7722688 — accessed 2026-10-22
15. Crescendo — Real-Time Coordinator (job posting, Honduras) — https://jobs.generalcatalyst.com/companies/crescendo-2-1fc09d6e-7427-4a85-8af7-a6d2cbfc5622/jobs/72492268-real-time-coordinator — accessed 2026-10-22
16. Pulse 2.0 — Crescendo: Interview With CEO Matt Price About The Full-Service CX Company — https://pulse2.com/crescendo-profile-matt-price-interview/ — 2025-06-03
17. SiliconANGLE — Crescendo raises $50M for its AI-powered customer service platform — https://siliconangle.com/2024/10/02/crescendo-raises-50m-ai-powered-customer-service-platform/ — 2024-10-02
18. Outsource Accelerator — Crescendo acquires PartnerHero, raises $50Mn for AI-powered CX — https://news.outsourceaccelerator.com/crescendo-acquires-partnerhero-ai-powered-cx/ — 2024-10-08
19. IT Brief UK — Crescendo launches AI-native customer experience platform — https://itbrief.co.uk/story/crescendo-launches-ai-native-customer-experience-platform — 2026-08-26
20. ISG Research — Contact Center Quality Processes Must Modernize for the Digital Era — https://research.isg-one.com/viewpoint/customer_experience/contact_center_quality_processes_must_modernize_for_the_digital_era — 2022-07-15
21. Call Centre Helper — Less than 2% of Interactions Evaluated for Quality Purposes (poll) — https://www.callcentrehelper.com/poll-percentage-of-the-interactions-evaluated-for-quality-purposes-94585.htm — 2016-11-04
22. Analytics India Magazine — BPOs are Ripe for AI Disruption—Their Contracts Aren’t (citing HFS Research) — https://analyticsindiamag.com/it-services/bpos-are-ripe-for-ai-disruptiontheir-contracts-arent — 2026-02-09