Agentic AI platform for enterprises

Put AI agents to work on the systems you already run.

Your ERP, core platforms and in-house applications stay where they are. Terrapin's agents work on top of them, with each user's own permissions, a person's approval before risky actions and a full record of every step.

Live execution · EX-2041800:00
Priya S. · Credit controller · Ask AI panel in the ERP

“Send reminders to every customer more than 60 days overdue.”

  1. CHChannelAsk AI panel
  2. WSWorkspaceFinance · EU
  3. OROrchestratorPlans 3 steps
  4. AGAgentCollections
  5. TLToolSend reminder
  6. ERPYour ERPSAP S/4HANA
    Acted asPriya
    Approvals0 of 1
    TraceWriting
    • No rip-and-replaceAgents work on top of SAP, Oracle and your own applications
    • Your permissionsAgents act as each user, inside the access they already have
    • Approval firstRisky actions wait for a named person to approve
    • Every step tracedEach decision, tool call, approval and cost on record
    • 10 channelsChat, API, events, WhatsApp, email, voice and more
    • 1 platformEvery team's agents, governed in one place

    The problem

    Why enterprise AI stalls after the pilot

    The demo answers questions from documents. The value is in reading orders, updating records and starting processes in the systems that run the business, and that is where most programmes stop.

    Reach

    Pilots cannot safely touch core systems

    To act in the ERP or a custom app, a bot needs credentials and broad access. Security says no, and the pilot never leaves the sandbox.

    With TerrapinSystems become governed business capabilities. Agents use them without ever holding a key or seeing an endpoint.

    Sprawl

    Every team builds its own bot

    Sales, service and finance each buy or build an assistant with its own prompts, keys and vendor. Nobody can see what runs where.

    With TerrapinOne platform and many Workspaces: shared tools and knowledge, with each team's context and policy kept apart.

    Control

    No approvals, no audit trail

    An agent sends a message or changes a record. Nobody approved it, and nobody can show afterwards how it decided.

    With TerrapinApprovals bound to the exact action, and a trace of every step, policy check, tool call and cost.

    Security

    A prompt is not a security control

    “Never show another customer's data” in a prompt can be talked around. Limits that matter have to sit outside the model.

    With TerrapinIdentity, data scope and hard limits are enforced by the platform on every step, whatever the model says.

    How it works

    One rule keeps every agent in its lane

    Each part of the platform answers one question, so responsibilities never blur. Select a layer to see what it does and where it sits in the collections request from the top of the page.

    Your systemsSAPOracleCore platformsCustom appsDatabases

    Connect anything you already run

    Your systems become capabilities agents can use, but never hold the keys to

    Tool Studio wraps an API, a workflow or a database procedure as a named business capability with a schema, a risk level and a policy. The agent sees the capability. The platform keeps the endpoint and the credentials.

    • REST / OpenAPIImport Swagger or OpenAPIPhase 1
    • WorkflowsYour existing enginePhase 1
    • GraphQLQueries and mutationsPhase 2
    • MCP serversReviewed before usePhase 2
    • FunctionsReviewed server-side codePhase 2
    • DB proceduresAllow-listed, least privilegePhase 3
    • Separate connections for DEV, TEST and PROD, so tests never touch live data
    • Secrets in a KMS-backed vault: revealed once, rotated, never in a prompt
    • Presets for OAuth2, API keys, mTLS, SAP and Oracle, plus custom connector adapters
    Tool Studio

    Governed by design

    Agents act with the right identity, and stop when a person must decide

    Who the agent acts as

    • Approvalsbound to the exact action, with four-eyes and MFA for critical steps
    • Human handofflive takeover with an AI summary, a ticket or a call back
    • Testing gaterequired tests must pass before any agent is published
    • Trace and auditevery step, policy decision, token and cost, in a tamper-evident log
    Execution EX-20418 · Collections agent for Priya S.Paused
    1. 1
      Get overdue invoices14 invoices found in SAP, read with Priya's access
    2. 2
      Draft reminders14 drafts in English, German and French
    3. 3
      Send remindersWaiting for approval · SLA 4 h, then escalates
    4. 4
      Log on the customer accountsDunning note posted through the finance workflow
    5. 5
      Write the traceSteps, approvals, tokens and cost
    Approval requiredMedium risk

    Send 14 payment reminders

    Requested by
    Collections agent, for Priya S.
    Why
    3 reminders exceed €10,000 (Workspace policy)
    Bound to
    args 7f3a…c91 · 14 recipients

    Try it: approve, reject, or change the action after it was requested.

    Channels

    One agent, ten ways in, the same rules everywhere

    Publish an agent to any channel from one builder. Identity checks, rate limits and policies follow it, and rich replies fall back to plain text so nothing is lost. Delivery is phased; labels show when each channel arrives.

    • ChatWebsite widget, hosted page or an Ask AI panel inside your own screensPhase 1
    • FormA form submission becomes an agent task, answered asyncPhase 1
    • APIA documented endpoint your own apps call, with OpenAPI generatedPhase 1
    • WebhookAny system can push an event, with signature checksPhase 1
    • SchedulerRecurring runs, such as every Monday at 9:00Phase 1
    • EventReact to a change, such as lead score above 80 or order createdPhase 1
    • WhatsAppTwo-way conversations, with approval for bulk sendsPhase 2
    • EmailRead, sort and answer inbound mailPhase 2
    • VoicePhone conversations with the same agents and rulesPhase 3
    • Mobile SDKAgents inside your own mobile appsPhase 3
    • Multilingual agents, starting with English, Hindi and Gujarati, with cross-language search
    • Identity checks and one-time-code step-up before sensitive actions
    • Anonymous visitors get no write actions unless you allow them

    Use cases by team

    Start from a Solution Pack, or put an agent on your own process

    Solution Packs are ready-made agents, tools and channels you install and adapt. The enterprise examples come from the platform specification's worked scenarios.

    Security, compliance and operations

    Built for the questions your CISO and auditors will ask

    Security is part of the design, not a setting. These controls are in the platform specification, and the figures below are design targets for the first release.

    Threat model

    • Prompt injection
    • Privilege escalation
    • Cross-tenant leakage
    • Credential theft
    • Denial of wallet
    • Webhook spoofing
    • SSRF
    • Malicious files
    • Insider misuse
    • MCP supply chain
    • Knowledge poisoning

    Tenant isolation

    • Row-level security in the database
    • Per-tenant encryption keys
    • Secrets in a KMS-backed vault, rotated
    • Append-only, hash-chained audit log

    LLM governance

    • Model providers allow-listed per tenant
    • Providers barred from training on your data
    • Optional PII redaction before model calls
    • Region-pinned storage and inference

    Data protection

    • Consent, purpose and retention per scope
    • Sensitive data never stored automatically
    • Erasure cascades through memory
    • Law checklist, e.g. India's DPDP Act

    AI Operations

    • Resolution, containment and handoff rates
    • Quality, safety and approval-time KPIs
    • 7-day before and after view per publish
    • Daily cost caps for each Workspace

    Roles and duties

    • 10 roles, scoped to tenant, unit or Workspace
    • No role gets AI rights automatically
    • No one approves their own request
    • Audit kept from 13 months to 7 years
    • Design targets
    • 99.9% runtime availability
    • ≤ 2.5 s first chat token, p95
    • ≤ 250 ms platform overhead
    • 15 min / 1 h RPO / RTO
    • WCAG 2.1 AA accessibility

    Delivery

    A phased rollout, shaped with pilot customers

    The platform is specified and being delivered in phases. Durations are indicative and count from programme kick-off. Phase 1 is complete when two pilot customers are live.

    NowSpecification in engineering review · pilot places open
    1. Phase 04–6 weeks
      Foundation
      • Tenancy, SSO and access
      • Audit, trace store, vault
      • Model Gateway
    2. Phase 1 · MVP10–14 weeks
      First agents in production
      • Agent Studio, Tool Studio (REST, workflows), Knowledge
      • Chat, form, API, webhook, scheduler and event channels
      • Approvals, traces, roles; Lead Qualification and Digital Receptionist packs
      Exit: two pilot customers live
    3. Phase 210–12 weeks
      Scale-out
      • WhatsApp, email, handoff inbox
      • MCP, GraphQL, Functions
      • Evaluation, replay, metering; four more packs
    4. Phase 310+ weeks
      Differentiate
      • Voice and Mobile SDK
      • Parallel multi-agent plans
      • Staged rollouts, DB procedures

    Pilot and design partners

    Bring one process and one system. Shape the platform with us.

    Pilot customers work with the product team on a real use case through Phase 1. Already in use today: Terrapin's Digital Marketeer, Digital Receptionist and Customer Service digital workers, which move onto the platform as Solution Packs.

    How to buy

    From first look to agents in production, in ten clear steps

    You always know what happens next, who owns it and what you get at each step. Start with step one or talk to us today.

    1. 1Understand the productDeck, brochure, videos and an interactive walkthrough
    2. 2Contact salesShare your systems and goals; we reply within one business day
    3. 3Pre-sales callWe map your processes, systems and risk controls, and pick a first use case
    4. 4ProposalScope, connectors, channels, rollout plan and total cost, in writing
    5. 5Agreement and paymentSign off and pay securelyPurchase complete
    6. 6OnboardingTenant, SSO, roles, Workspaces and environments set up
    7. 7Implementation and trainingSystems connected, agents built and tested, your team trained
    8. 8Go-live and handoverAgents published after their tests pass and handed to your team
    9. 9HypercareClose watch on traces, quality and cost in the first weeks
    10. 10Ongoing supportHelp when you need it, upgrades and regular KPI reviews

    Pick one process. See an agent run on it.

    In a working session we map one of your processes, the systems it touches and the controls it needs, and show how an agent would run it.

    Company Logo