Cresta

What it is

Cresta sells both halves of the argument every contact-centre buyer has first: make the human agent better, or take the contact away from them. Agent Assist listens to the live call or chat and pushes the next-best action, the knowledge article, and the compliance prompt into the rep’s screen. Cresta AI Agent handles the conversation end to end with no rep on it. Both run off the same conversation models, the same policies, and the same quality rubrics, which is the whole pitch — you are not choosing a deflection vendor and an enablement vendor and hoping their definitions of “resolved” agree.

The company was founded in 2017 out of the Stanford AI Lab by Sebastian Thrun, Zayd Enam, and Tim Shi, and is run by CEO Ping Wu, who co-founded Google’s Contact Center AI before joining in May 2023. Headquarters is Palo Alto, headcount is above 500. Cresta closed a $125M Series D in November 2024 co-led by World Innovation Lab and the Qatar Investment Authority, taking total funding past $280M, and announced it had passed $100M ARR at the end of April 2026 — reported by Axios on 30 April 2026 — alongside naming Sequoia’s Doug Leone board chair.

Named customers skew large and voice-heavy: Verizon, United Airlines, Alaska Airlines, Cox Communications, Marriott, Hilton, CVS, Brinks Home, Snap Finance, Propel Holdings, Aqua Finance. Cresta states it has processed over 100 million enterprise conversations.

Why it shows up in CX stacks

• The supervisor layer is the part competitors do not sell. The Agent Operations Center, launched December 2025, puts human-handled and AI-handled conversations in one monitoring queue. A supervisor watches a live AI-led interaction, flags a high-risk one, steers the AI agent’s next response, messages through the AI’s interface, or takes the conversation over outright. If your rollout plan involves running humans and agents side by side for a year rather than flipping a switch, that queue is the operational artefact you actually need and the thing most AI-agent vendors leave you to build.
• Knowledge Agent reads the screen, not just the transcript. Launched 17 March 2026, it pairs ambient listening with live browser context — account status, order history, loyalty tier — so the answer it surfaces is scoped to the specific customer on the line, with source links. The problem it targets is the toggle tax between CRM, knowledge base, and billing console, which is where handle time actually goes in a tier-1 queue.
• Agent development got a testing loop in mid-2026. Conductor shipped 11 June 2026 as a natural-language build environment covering discovery, blueprint, build, test, and post-launch diagnosis, with Cresta claiming engineers ship production agents 2x faster. It runs against Synthetic Customers — personas generated from your own historical conversations rather than from a survey — and AI Agent Testing 2.0, which drafts evaluators from your documentation. Simulation against your own traffic shapes is the difference between a pilot that survives a Monday and one that survives Black Friday.
• It plugs into the CCaaS you already bought. Connectors cover NICE CXone, Genesys Cloud CX, Five9, Amazon Connect, Twilio, LivePerson, and Salesforce, with listings on AWS and Google Cloud Marketplace, so the platform decision does not force a telephony migration.
• The compliance set is built for regulated buyers. ISO/IEC 42001, SOC 2 Type II, HIPAA, GDPR, CCPA, TISAX, and stated PCI-DSS alignment. ISO 42001 in particular is the one procurement teams in financial services and healthcare have started asking for by name.

Pricing reality

No published rate card and no self-serve tier. The one hard public anchor is Cresta’s AWS Marketplace listing: Agent Assist for Chat at $150,000 for 12 months covering up to 125,000 chats, then $1.20 per additional chat, and Agent Assist for Voice at $150,000 for 12 months covering up to 100,000 calls, then $1.50 per additional call. Contracts there are noncancellable and nonrefundable, with infrastructure charges separate.

Two things fall straight out of that listing. First, the meter is linear — $150,000 over 125,000 chats is $1.20 a chat, exactly the overage rate, so there is no volume break waiting above the commitment and no reason to over-commit for one. Second, and more important at the negotiating table: the $1.50 you pay Cresta for a voice interaction buys assistance on a call a human still handles and still gets paid for. Intercom Fin bills $0.99 per resolution and Zendesk $1.50 per automated resolution, and those units retire the labour cost. Comparing the headline numbers without normalising the unit is the most common way this evaluation goes wrong.

Third-party 2026 estimates put whole Cresta contracts at roughly $60,000 to $150,000 a year and per-seat rates around $100 to $200 per agent per month on annual terms — treat both as estimates, not quotes; neither is published by Cresta. Take voice and chat assist together and the AWS anchor implies about $300,000 a year before overages or any additional module.

Best for

Enterprise CX and contact-centre leaders running roughly 300 seats or more, weighted to voice, in a regulated or high-value-per-contact operation — telco, airlines, insurance, home security, consumer lending, hospitality — whose next twelve months are about agent productivity, QA coverage across every interaction rather than a 2% sample, and a controlled hand-off of the easy contact volume to AI. It is the strongest pick when the CX organisation and the AI programme have to report against the same quality rubric.

Skip it if your queue is chat-dominant and low-complexity, if the goal is deflection with a shrinking headcount and no assist layer at all, if you need pricing you can approve without a procurement cycle, or if your volume sits below roughly 50,000 contacts a year where the fixed commitment cannot pay back.

Versus the alternatives

The incumbent volume sits with the CCaaS and workforce-engagement suites — NICE, Verint, and Genesys — whose native AI is bundled into a contract you already signed. Take the bundled option when your existing QM and WEM licences already cover the scoring use case and a second per-seat platform cannot clear the finance bar; that is the cheapest first test and the right answer for a large share of mid-market queues. Sierra is the fastest-growing entrant in the segment at roughly $200M ARR by May 2026, and it plus Decagon are the picks when the objective is containment and the human assist layer is not part of the plan. Parloa wins when telephony estate integration and a predictable per-minute meter decide the deal. Gladly is the pick for a person-centred model where the customer, not the ticket, is the record. Solidroad covers coaching and QA alone at a fraction of the cost if that is the only job. Cresta’s case against all of them is narrow: it is the one that runs the human queue and the AI queue under one supervisor console and one scoring rubric, and if you are not going to operate both, you are paying for half a platform.

Watch-outs

• The commercial model bills you for seats while the product removes them. Assist is priced against agent seats and assisted conversations; AI Agent success shrinks exactly that base. A three-year assist commitment signed the quarter before a containment push is a bill for headcount you no longer have. Guard: negotiate a blended pool at signature rather than two line items — put the conversion ratio between assist spend and AI-agent spend in the contract, cap the annual step-down penalty, and refuse a multi-year seat floor that your own automation roadmap is designed to breach.
• Every outcome number on this page is vendor-published. The 30-point NPS gain and 50% QM cost cut at Brinks Home, 20% revenue lift at Cox, 61% collections lift at Aqua Finance, 58% AI Agent containment at Propel Holdings — all Cresta customer stories, none independently audited, and each defines its own metric. Guard: run a two-week shadow period on your own traffic, score it against your existing FCR and containment definitions, and write that definition into the contract as the acceptance test rather than into the SOW as a goal.
• Real-time assist is continuous monitoring of employees, and Europe treats that differently. Ambient listening plus live browser context is screen and audio capture on every agent, every shift. The EU AI Act prohibits inferring emotional state of workers in the workplace outside narrow exceptions, and Cresta’s supervisor tooling surfaces emotionally charged conversations. Guard: complete a DPIA before the pilot, not before go-live; open works council consultation in DACH and France ahead of procurement, since that timeline is measured in months; and get the vendor’s written statement of which models infer agent emotional state versus customer sentiment, with the agent-facing ones disabled in EU deployments.
• Three of the products you will be shown are under a year old. Agent Operations Center is from December 2025, Knowledge Agent from March 2026, Conductor and Synthetic Customers from May and June 2026. The demo is current; the production track record at your scale is not. Guard: buy only the modules with reference customers at your seat count and in your vertical, and take the rest as a no-cost pilot addendum with a named exit rather than as contracted line items.
• Cresta rides on your CCaaS and inherits its fate. The platform integrates with NICE CXone, Genesys, Five9, and Amazon Connect rather than replacing them, so a telephony migration mid-term can invalidate the integration you bought against. Guard: confirm the connector supports your specific platform version and deployment topology in writing, make working integration an acceptance criterion rather than an implementation assumption, and check what the contract says about your own platform change before you sign a term longer than your CCaaS contract.