Gartner predicts agentic AI will autonomously resolve 80% of common customer service issues by 2029. Today, only 46% of people trust AI — even as 66% use it regularly. PrimeAssist closes that gap.
Prefer concierge onboarding? Join the waitlist
78%
of organizations now use AI in at least one business function
Stanford HAI, AI Index Report (Apr 2025) →53%
of companies are deploying AI agents in customer service
BCG, Agentic AI in Customer Service (Dec 2025) →// the four pillars
Four constraints, enforced together — each one a promise the agent keeps on every answer.
Every answer cites a source. Confidence is first-class.
PrimeAssist answers only from documents you have uploaded and approved. Every claim carries a citation back to the originating page. When confidence drops below your threshold, the agent says I do not know instead of guessing.
Topic allowlists, policy packs, and refusal templates that explain why.
PrimeAssist enforces your topic boundaries before the model ever runs. Policy packs codify what the agent can say about pricing, claims, eligibility, refunds. Refusals explain to the customer why — never a blank stare.
Tools take real action through your APIs — with dry-run, confirmation, and audit.
PrimeAssist agents take real work through your APIs — book the appointment, refund the order, look up the claim, update the address. Every action passes through governance: dry-run by default, customer confirmation for anything irreversible, full audit trail for compliance.
The agent knows when to hand off. The human gets full context.
PrimeAssist watches its own confidence in real time. When it dips below your threshold, when a topic falls outside policy, when the customer explicitly asks for a human — the agent hands off with the full conversation, the relevant policies, and a clean summary. No please describe your issue again.
// how it works
A real PrimeAssist trace for one customer turn. Each step is a structured event the agent must explain — to your compliance team, to a human handoff, or back to the customer who asked.
Search your approved knowledge — never the open web.
The agent searches across the documents you have uploaded and approved (PDFs, Word, Markdown, scraped sites, API specs) and ranks the top relevant chunks. A confidence value rides every retrieval. Nothing the agent does not have grounding for can be claimed.
Draft an answer inside the topics and policies you set.
The model drafts an answer using only the retrieved chunks, inside your topic allowlist and your policy pack. Refusal templates fire when the question is out of scope; PII gets masked before it ever reaches the model; confidence is checked against your threshold.
Call your APIs — with dry-run, confirmation, and audit.
When the answer requires real work — a lookup, a refund, a booking — the agent calls your API. Destructive operations dry-run by default. The customer confirms anything irreversible. Every call is audited with the policy that authorized it, the inputs, and the result.
Hand off the moment confidence drops — never silently.
Below your confidence threshold? Outside your topics? Customer asks? The agent hands off to a human with the conversation, the relevant policies, the candidate answer it was about to send, and the confidence score. No please describe your issue again.
// what the platform does
The pillars are not slogans — they are the surfaces a tenant configures. Eight capabilities, each one a working part of the platform.
Every reply is drawn from retrieved passages of the tenant's own knowledge, with those sources surfaced inline. An answer that cannot be grounded is never shown as confident.
Each agent is bounded by configured topics and policy packs. Out-of-scope questions are declined or routed, and sensitive triggers force a human handoff.
Tools let the agent read and write through the tenant's existing APIs. Writes run behind dry-run, customer confirmation, and approval queues so every action is deliberate and audited.
Tenants bring their own handbooks, plan documents, and product docs. PrimeAssist chunks and embeds them into a per-tenant vector namespace, isolated from every other tenant.
Golden sets run against every agent version and A/B comparisons measure each change. Citation-regression alerts catch grounding drift before it reaches customers.
The platform-side surface for running PrimeAssist as a product: cross-tenant management, plan administration, and impersonation backed by a full audit trail.
The widget and dashboard take on each tenant's brand colours and radii. Theming is driven entirely by CSS custom properties, with no runtime style scripting.
Starter packs for insurance, banking, healthcare, SaaS, and telecom — each a curated knowledge base, policy pack, tool surface, and golden set. Day one is a configuration, not a build.
// one agent, every channel
Chat, voice, phone, SMS, WhatsApp, and email all run the same agent loop — same knowledge, same policies, same audit trail, same handoff path. Pick a channel to see how it answers.
An embeddable script drop-in that streams grounded answers with citations, on any site.
Realtime voice on Azure OpenAI Realtime, with barge-in, in the browser or over the phone.
Twilio inbound and outbound calling — the agent answers real phone numbers and warm-transfers to humans.
Two-way SMS via Twilio — the same grounded agent, answering over text.
WhatsApp Business messaging routed through the shared inbound pipeline.
Inbound email parsed into conversations, with threaded grounded replies.
// peer-reviewed productivity
+14% +34%
// what this looks like in production
Klarna also revised parts of its all-AI stance in 2025 and rehired support staff for complex cases. That tradeoff — fast where the agent is grounded, human where it is not — is precisely what PrimeAssist treats as the default, not the lesson learned the hard way.
Klarna press release, Feb 2024 →// regulator-grade by design
EU AI Act Article 13 requires high-risk AI systems to be sufficiently transparent for deployers to interpret output and use it appropriately. PrimeAssist makes citations and per-answer confidence first-class.
EU AI Act, Article 13 (in force Aug 2026) →NIST AI 600-1 names confabulation in current-generation LLMs as pervasive. PrimeAssist's grounded retrieval and confidence threshold are the prescribed mitigation.
NIST AI 600-1 (July 2024) →OECD AI Principle 1.3 requires plain, easy-to-understand information on the sources of data, inputs, and logic that led to a prediction. PrimeAssist surfaces them inline with the answer.
OECD AI Principles, updated May 2024 →// the four disclosures
Security
Per-tenant Row-Level Security, AES-GCM encryption with per-org data keys, MFA-enforced operator access, and an append-only audit log with tamper-evident export.
Read more
Compliance
Pre-GA. Controls already shipping — audit log, retention, encryption, RLS. Pending — SOC 2 Type I audit, ISO 27001. DPA available on request.
Read more
Sub-processors
Identity (Auth0), object storage (MinIO), LLM inference (Anthropic, OpenAI), SMS + voice (Twilio), email (Postmark). Region and purpose disclosed.
Read more
Operator access
Operator impersonation is time-bounded, reason-logged, and surfaces in your audit log. Tenants can end an active impersonation session from their dashboard.
Read more
// be among the first
Drop your details. We will reach out personally when we have a slot that fits your vertical.
The hosted waitlist form is not wired in yet. Reach out directly and we will add you to the list by hand.
Email to join
Operators: set PUBLIC_WAITLIST_PROVIDER and the matching form ID in .env to drop in a Tally or ConvertKit embed without touching code.
No backend yet — your details go straight to our inbox and stay there.