SoundHound AI brings voice agents and business actions together in OASYS

SoundHound AI builds conversational systems for interactions that happen on calls, inside vehicles and across business services. Its OASYS platform combines voice agents with tools for controlling actions, testing conversations and involving people when needed. The company is particularly relevant where speech is the main interface and the conversation must end in a useful business result. Evaluating it means looking beyond voice quality to the system that completes, records and supports the requested action.

01 / ProductA voice business with a broader agent platform

The SoundHound overview presents OASYS across contact centers, digital assistants, restaurant ordering, automotive experiences and employee support. Those settings share spoken interaction but differ substantially in surrounding systems. A restaurant agent must understand menu choices and submit the right order. An in-car assistant needs an interface appropriate to the vehicle. A service agent may need to identify a customer and change a business record.

The OASYS page describes an agent lifecycle covering creation, control, simulation, live operation and improvement. Prebuilt agents and configuration from existing interaction material sit alongside tools for inspecting results. The platform also describes optional human-assisted resolution, where a person can help an agent without necessarily taking over the entire conversation. Whether that produces a useful experience depends on how quickly assistance is available and what the person can see.

Ownership has recently changed the scope of the company. SoundHound’s 4 September 2026 announcement confirms the completed acquisition of LivePerson. It says LivePerson’s messaging platform will be integrated into OASYS. That wording matters: the transaction is completed, but it does not establish that every promised cross-channel integration is already available to a new customer.

02 / AudienceWhen speech is the central product requirement

SoundHound is worth evaluating when an organization already has substantial voice traffic or a physical environment in which typing is inconvenient. The relevant audience includes service operations, restaurant technology teams and manufacturers embedding assistants into products. These buyers need a conversation that can survive interruptions, imperfect audio and changes of mind while still respecting the rules of the transaction.

The voice capabilities page describes interruption handling, noise filtering, multilingual interaction and domain tuning. These are vendor-described capabilities, not results of our testing. A useful comparison is our Retell AI blueprint for teams considering a voice-agent building platform. Our Sierra blueprint examines service journeys and their operating model. Compare the whole route from incoming interaction to confirmed result.

A company with only a small number of simple support messages may not need a broad voice platform. Equally, an organization that mainly wants a transcription feed should distinguish that narrower requirement from an agent that takes actions. Speech recognition, conversation management and order execution can fail independently. A supplier should demonstrate the combination relevant to your channel rather than substitute a polished general conversation.

03 / WorkflowA proposed restaurant phone-order pilot

Consider a proposed evaluation for a restaurant receiving phone orders during busy periods. This is an illustrative workflow, not a SoundHound installation we tested. Restrict the initial scope to collection orders from one store and one current menu. The agent needs the right location, available items, permitted modifiers and a trustworthy way to submit the order to the restaurant’s system.

Begin with the menu source and its update process. A caller may use an old product name, combine options in an unusual way or request an unavailable item. The conversational layer can clarify what is wanted, but the ordering system should validate the resulting basket. The agent should offer a supported alternative or ask an employee for help when the available data cannot resolve the request.

Use interruptions as a primary test rather than an occasional edge case. Have a caller change the size after choosing a topping, remove an item while the basket is being repeated, and ask about collection time halfway through an order. The test should check the final basket after each change. Fluent acknowledgment is not enough if the wrong modifier remains in the underlying order.

The control and security description distinguishes autonomous reasoning, rule-based execution and human assistance. For the proposed pilot, make price calculation, promotion eligibility and final order submission controlled operations. The language model can help navigate the conversation, while the ordering system determines the valid price and accepted order. Repeat the confirmed basket before the final commitment.

Ambiguous submission is an especially valuable failure case. If the restaurant system accepts an order but the response to the agent times out, blindly submitting again could create a duplicate. Ask the implementation team to show how the agent resolves that state. The desired customer experience is an honest confirmation after checking the order record, or a supported handoff when its status remains uncertain.

Finally, test the restaurant’s side of the service. A successful phone call is not useful if employees receive an incomplete ticket or cannot find the order when the customer arrives. Include the visible kitchen or store record, the promised collection information and the procedure for a subsequent correction. These are proposed acceptance criteria, not claims that every OASYS configuration supplies the same integration automatically.

04 / PricingAn enterprise proposal needs a specific channel and workload

The public contact route invites a commercial discussion; the reviewed OASYS pages do not provide a universal public numeric tariff. This article therefore does not estimate a per-minute, per-order or per-seat rate. Ask for a proposal that names the products, channels, deployment support and ongoing services actually included.

For restaurant ordering, compare cost with accepted, usable orders and remaining staff effort. A low conversation cost can hide substantial manual correction if the basket needs rebuilding. At the same time, a carefully escalated unusual order may be the right result. The comparison should recognize correct completion, useful assistance and avoidable rework separately.

Clarify the commercial treatment of transferred calls, test traffic, seasonal peaks and human-assisted resolution. These are questions for the proposal, not published SoundHound billing rules. Following the LivePerson acquisition, buyers should also distinguish functionality in their present agreement from functionality described in a future integration plan.

05 / DistinctionsWhy voice controls and business controls must meet

SoundHound’s emphasis on spoken interaction is meaningful because conversational errors often occur before business logic is reached. A customer who interrupts or corrects themselves can change the meaning of an otherwise well-formed order. An evaluation that uses only clean recordings and complete sentences would miss the part of the offer most relevant to noisy, fast-moving environments.

OASYS also presents function-level and channel-level controls rather than treating every response as free-form generation. The practical opportunity is to preserve a natural conversation while limiting what can change in the business system. Ask the supplier to demonstrate the same requested action through several phrasings and verify that the underlying eligibility check remains identical.

The broader company portfolio may help businesses with both physical and digital touchpoints. However, portfolio breadth is not the same as a shared production deployment. The completed acquisition expands SoundHound’s corporate capabilities; whether a specific channel transition is supported should be demonstrated against the actual product version and integration in the proposal.

06 / QuestionsWhat needs independent deployment evidence

The vendor publishes strong accuracy, latency and scale claims. We did not run an audio benchmark or verify those figures for a customer environment, so they are not presented as expected results. Measure the actual phone path, background conditions, menu vocabulary and peak demand you intend to support. A model-level result does not automatically describe the full interaction.

The other unresolved boundary is assistance. Establish whether an employee can intervene, what context they receive and what happens when no one is available. A promise to keep the conversation moving should not conceal a stalled transaction. The pilot should expose incomplete actions clearly enough that staff can recover them without asking the customer to repeat everything.

07 / DecisionJudge the accepted action, not just the voice

SoundHound is a serious evaluation candidate when voice is central to a customer or product experience and business-system actions matter. Choose one channel and a bounded task, then test interruptions, system failures and staff recovery alongside successful calls. Expand only when the confirmed downstream record and the spoken customer understanding stay aligned.