Forethought vs SoundHound (2026)

Forethought connects to helpdesks, knowledge bases, CRM and API management platforms, with more than seventy integrations spanning contact centers and connector platforms including Airtable and Snowflake, and Autoflows reason through business policies to take authenticated action, not hand a suggestion back. Where no API exists at all, Browser Agent completes the workflow directly, so the action surface is not bounded by what partners expose.

Sourceforethought.ai/platformread 2026-09-05

Amelia 7 ships an expanded library of third-party integrations and pre-built connectors that let agents retrieve enterprise application data via APIs and execute actions. Banking deployments integrate with core banking systems for account inquiries, loan applications, transfers and card management, and healthcare deployments integrate with Epic.

In voice commerce, drivers order takeout, reserve restaurants through OpenTable, pay for parking through Parkopedia, and reach calendar, email and messaging over MCP to book meetings, so the agent completes transactions in outside systems. Agents consume MCP servers outward as tool calls, and the platform also hosts MCP and A2A agents inside it.

Sourcesoundhound.com Amelia 7.0 release, CES 2026 announcement and solution pagesread 2026-09-06

Autoflows is Forethought's agentic reasoning engine. It takes business logic in natural language, not a decision tree, and reasons through complex policies to resolve an issue end to end. Four agents hand off to each other in a set sequence. Discover generates the Autoflows, Solve detects intent and resolves, Triage tags and routes what needs a person, and Assist surfaces the right Autoflow to the human agent who picks it up. Browser Agent executes the steps no API can reach.

Sourceforethought.ai/platformread 2026-09-05

Agentic+ is a proprietary multi-process framework at the core of Amelia 7. It lets businesses deploy fleets of goal-oriented agents and handles a complex query by orchestrating multiple agents, each with a specific function, to resolve it by text or voice without human intervention.

It moves between generative flows, where a probabilistic engine synthesizes from data, and deterministic rules-based flows that guarantee consistent results, which SoundHound describes as essential in financial, legal and other high-precision transactions, so a deterministic path can sit inside an agentic run.

Workflows carry a human-in-the-loop mechanism as a defined step, and the framework orchestrates across models, deterministic flows and integrations together.

Sourcesoundhound.com Amelia 7.0 blog and press releaseread 2026-09-06

Work reaches the agent from any channel. Solve resolves across chat, email, voice, SMS, Slack and in-app, with email answered instantly, phone calls handled end to end, embedded chat, and internal Slack support, and Triage picks up tickets arriving in the helpdesk queue and routes them onward.

Sourceforethought.ai/platformread 2026-09-05

A single agent configuration serves phone, web chat, SMS and messaging apps from the same knowledge base in more than a hundred languages, and one hotel agent, for example, runs across marketing, reservations and on-property concierge over all four. Beyond the contact center, the same platform reaches into vehicles, TVs and smart devices, so work arrives from an embedded assistant in a car as readily as from a web widget, each an independent inbound queue. No event catalog, schedule or webhook trigger is published, and work starts when an inbound contact arrives on one of these paths.

Sourcesoundhound.com Amelia 7.0 release, platform overview and CES 2026 announcementread 2026-09-06

Forethought connects to a brand's helpdesk, knowledge base and CRM and learns from its historical support data from day one, with Discover generating Autoflow policies and help center articles from that corpus and Solve answering with the right Autoflow or knowledge base article. There is no stated index, embeddings layer, retrieval architecture, refresh behavior or source-level citation, so a maintained retrieval structure is claimed but not shown. The per-customer fine-tuned models are training, not retrieval.

Sourceforethought.ai/platformread 2026-09-05

Tailored answer libraries can be built for an agent, either through a Q&A format that returns precise answers or by uploading files, websites and other knowledge sources to support model-generated answers, so agents respond dynamically to questions nobody scripted.

Amelia 7.1 lists improved knowledge retrieval and fine-tuned matching algorithms among its accuracy gains and improved retrieval again among its latency gains. Businesses define the specific knowledge and functions needed to execute a task, domain-specific content ships per industry, and the same knowledge base serves web chat, SMS and voice.

Sourcesoundhound.com Amelia 7.0 blog and Amelia 7.1 releaseread 2026-09-06

There is no stated memory layer with a scope or lifetime. The platform is trained on a brand's historical support data from day one to generate Autoflow policies and surface knowledge gaps, which is training and retrieval over a corpus, not state the agent retains between sessions. The published agent set names Solve, Triage, Assist and Discover with no memory component among them.

Sourceforethought.ai/platformread 2026-09-05

Agents handle complex multi-turn interactions with context held across the exchange, and Amelia 7.1 makes agents more context-aware through process-specific instructions and improved retrieval. The platform maintains a conversation across phone, web, text and messaging, so a single interaction can move channel without restarting. No memory layer with a stated scope and lifetime is described.

Nothing says what an agent retains about a person between separate contacts, whether memory is scoped per user, per account or per workflow, how long it lives, or how an operator inspects or clears it. The answer libraries and uploaded knowledge sources are a corpus, and the execution histories are a log.

Sourcesoundhound.com Amelia 7.1 release and platform pagesread 2026-09-06

The gate before an action sits in Forethought's own product, in two forms. Browser Agent runs only inside pre-approved policies with encrypted credentials and human-in-the-loop checkpoints, with approval steps configurable per workflow, so the customer decides which browser actions the agent may take unattended. Alongside that sit configurable escalation paths and an approval workflow in which Assist drafts a response and a person approves before it is sent.

Sourceforethought.ai/webinars/ai-that-clicks-for-you-automate-76-percent-of-cx-work-apis-cant-reachread 2026-09-05

Layered guardrails, confidence checks, safe fallbacks and human escalation work together, and agents are grounded in customer system data and never guess. Agentic+ workflows include a human-in-the-loop mechanism to ensure personal oversight when required.

Escalation teams and queues are configured, human representatives handling escalated conversations are managed, canned response templates are prepared, and an Agent Console gives the person a companion AI that listens to live conversations and offers responses, offers or guidance. For telecom, a smart human-in-the-loop escalates across the full spectrum from simple queries to complex multi-step resolutions.

Sourcesoundhound.com Amelia 7.0 blog, platform and telecom solution pagesread 2026-09-06

The platform is independently audited and certified for security, availability and confidentiality, with annual SOC 2 Type II audits, controls aligned to NIST 800-53, HIPAA, GDPR and CCPA, encryption at rest and TLS in transit, production secrets managed in AWS tooling, a DPA reflecting GDPR and CCPA, security-design reviews, threat models, regular third-party penetration tests, a named Data Protection Officer, a HackerOne bug bounty, a trust center at trust.forethought.ai, and a fully published subprocessor table naming each processor's identity, purpose and location.

Automated redaction removes PII, PHI and financial data at ingestion with the original securely deleted within 24 hours. Forethought reports ISO 27001 compliance and SOC 2 certification, and it is not FedRAMP certified. Role based access is claimed on the platform page but not repeated on the security page or in the legal documents.

Sourceforethought.ai/platform/securityread 2026-09-05

Customers manage their own password complexity and SSO/SAML 2.0 requirements, so identity integrates through the customer's provider, and Amelia follows set practices for identifying and authenticating users accessing its systems.

Amelia states compliance with ISO/IEC 27001, SOC 2 Type II, PCI DSS 3.2.1 and the HIPAA Safeguard Rule, with independent auditors reviewing HIPAA and GDPR compliance and a certifications page on the Amelia domain. Healthcare deployments are positioned as HIPAA-compliant agents integrated with Epic.

Data moves over TLS 1.2 or higher and rests under AWS KMS with FIPS 140-2 validated modules, customer data is returned or destroyed on request, and no customer data is shared across clients.

Sourcesoundhound.com Amelia data processing agreement, platform and solution pagesread 2026-09-06

Discover surfaces insights across all support interactions and detects where the agent is failing, Triage gives CX leaders real-time visibility into agent performance and coaching opportunities, and audit trails and automatic redaction of personal, health and financial data at ingestion are in place. This reports on outcomes and content gaps, not on the run. There is no stated step-level decision trace, tool-call log, Autoflow execution record or retrieval log, which matters more on a product whose Browser Agent clicks and types inside third-party systems.

Sourceforethought.ai/platform/triageread 2026-09-05

Amelia 7.1 adds scripting tools to access AI agent execution histories, under transparency, one of its four headline improvement areas, to give enterprises deeper control, visibility and compliance in complex deployments. An execution history records what a specific agent did on a specific run and in what order. The build interface also has analytics tooling and an AI companion that answers questions while building. SoundHound describes no record of why an agent chose one path over another. No retention period or export format is published.

Sourcesoundhound.com Amelia 7.1 release and Amelia 7.0 releaseread 2026-09-06

Every core processor that hosts customer data sits in the United States. Amazon Web Services, MongoDB, Microsoft Hosting Services and Google are all listed as web hosting in the United States, with Cloudflare, OpenAI, Groq, Mistral and Pusher likewise, and only Datadog for logging and Snowflake for warehousing add Germany. The subprocessor table names each location.

There is no stated region menu, residency commitment, VPC, single-tenant or self-hosted option, and the security page, which covers compliance, encryption, product design, training, oversight, PII handling, FedRAMP scope and HIPAA in detail, does not address residency. A customer with an EU or UK residency requirement can see the answer and cannot change it.

Sourceforethought.ai/data-subprocessorsread 2026-09-05

Customers can host Amelia in their own private cloud as an alternative to the SaaS model, and SoundHound has customers doing this today.

Amelia 7.1 adds support for custom on-premise LLMs, so the inference layer can also sit inside the customer's estate, and the agentic stack has been shown running entirely on the edge inside a vehicle with NVIDIA, alongside deployment embedded in TVs and smart devices.

The default remains SoundHound-managed hosting on AWS, with Amelia holding administrative control of its virtual servers. No region selector or named residency jurisdiction is published for the SaaS mode.

Sourcesoundhound.com financial services solution FAQ, Amelia 7.1 release, Amelia DPAread 2026-09-06

Prebuilt agents come as a named set the customer adopts as units, each with its own product page. They are Solve for end-to-end customer resolution, Triage for classification and routing, Assist as an in-helpdesk copilot, Discover for gap detection and content generation, plus Browser Agent and Voice. Discover also generates Autoflows from a brand's own history, so the customer starts with populated workflows, not a blank builder.

Sourceforethought.ai/platformread 2026-09-05

Each member of the named product line does its own job. Smart Answering handles inbound calls, Smart Ordering and Dynamic Drive-Thru automate restaurant ordering across channels, Employee Assist gives staff hands-free access to manuals and training, Autonomics automates IT operations, and Sales Assist supports retail floor staff during live conversations.

Each remains a whole product without the others. Beneath them sits a library of business-ready presets for building agents that fully automate familiar tasks such as password resets and reorders, plus domain-specific content per industry and canned response templates for human representatives.

Sourcesoundhound.com Amelia 7.0 blog, platform overview and solution pagesread 2026-09-06

OpenAI, Groq and Mistral are listed as subprocessors under AI Models, alongside Gladia and Deepgram for speech recognition and ElevenLabs and Cartesia for text to speech on the voice product. The platform runs across several providers, not one, and its per-customer fine-tuned models sit on top of that mix. Cloudflare, AWS, MongoDB, Microsoft and Google on the same table are infrastructure, not model providers. The routing is vendor-internal and no model selection, override or bring-your-own-model control is exposed to the customer.

Sourceforethought.ai/data-subprocessorsread 2026-09-05

The customer chooses the model. Amelia's LLM-agnostic architecture lets the customer use the best model for the job, whether a public model, a private model or a combination, which SoundHound presents as a key strategic advantage that prevents vendor lock-in.

Amelia 7.1 extends the choice to models the customer hosts, adding support for custom on-premise LLMs, and where fine-tuning is pursued it runs through the LLM provider's own infrastructure with the configuration dedicated to the customer's account. LLM agnostic is a platform pillar of Amelia 7.0, repeated in earnings commentary, and Agentic+ orchestrates across models rather than binding to one. The choice spans public, private and self-hosted models.

Sourcesoundhound.com financial services and telecom solution FAQs, Amelia 7.0 and 7.1 releasesread 2026-09-06

A Headless channel ships as a named product surface for embedding the agent into a customer's own systems, alongside a mobile SDK for in-app chat, an Assist browser extension delivered into the helpdesk, and connector integrations reaching API management platforms, with code and no-code deployment paths. There is no published API reference, SDK documentation site or MCP server for Forethought's own platform, so the headless surface is named as a capability but not documented as one.

Sourceforethought.ai/channels/headlessread 2026-09-05

Amelia 7 can host any agent optimized for MCP and A2A, so enterprises and OEMs run self-built agents, SoundHound's prebuilt agents or agents from outside the SoundHound ecosystem together in one environment, and Amelia 7.1 adds scripting tools that reach agent execution histories. Agents also consume MCP servers outward for calendar, email and messaging. No outside caller is described driving the platform itself.

Consuming MCP servers is Amelia reaching out to other systems, and hosting third-party agents means Amelia runs somebody else's agent inside itself. A2A works in both directions, but there is no agent card for SoundHound's own agents, no endpoint or auth scheme, and no API reference, SDK or developer portal.

Sourcesoundhound.com Amelia 7.0 and 7.1 releasesread 2026-09-06

The Discover agent runs an improvement loop after deployment. It analyzes a brand's support data, surfaces knowledge gaps, and generates both help center content and new Autoflows to close them, so live performance feeds back into what the agent can resolve. The QA Agent, a named product at forethought.ai/platform/agent-qa, scores 100 percent of human agent tickets, not the AI agent's own behavior, and there is no stated simulation environment, scored test suite or evaluation harness for the customer's own agent.

Sourceforethought.ai/platform/agent-qaread 2026-09-05

There is no evaluation harness, no scored test cases, no quality gate in a release path and no controlled post-deployment experiment. Nothing describes previewing an agent against defined conversations before it goes live, versioning a configuration, rolling one back, or comparing a change against a control. The platform offers build-time analytics and an AI companion answering questions while building, and execution histories reachable by scripting form the run record. Amelia 7.1 reports accuracy and latency improvements as SoundHound's own results, not as something a customer can run.

Sourcesoundhound.com Amelia 7.0 and 7.1 releases and platform pagesread 2026-09-06

Browser Agent drives a real interface itself. It logs in, clicks, types and completes the same browser-based workflows human agents run today, using pre-approved policies and encrypted credentials, and Forethought aims it at the work that APIs cannot reach.

Sourceforethought.ai/webinars/ai-that-clicks-for-you-automate-76-percent-of-cx-work-apis-cant-reachread 2026-09-05

No browser, desktop or remote computer control is described. Agents reach outside systems through pre-built connectors, enterprise APIs and MCP servers, and their surfaces are voice, chat, SMS and messaging plus embedded assistants in vehicles, TVs and devices, all channels the agent speaks through. Vision AI, shown at CES, pairs a vehicle camera's real-time perception with the voice stack so the assistant can see and interpret the driver's surroundings, which is perception of the physical world, not operation of a software interface.

Sourcesoundhound.com Amelia 7.0 release and CES 2026 announcementread 2026-09-06