Customer Service
Refunds, cancellations, account access, and escalation routing.
- 900 / 100: Baseline 55% → Decion 70% (+15 pts)
- 1,900 / 100: Baseline 55% → Decion 80% (+25 pts)
Decion is a decision-making engine for AI agents. Submit decisions with context and candidate actions, get ranked recommendations backed by historical evidence, and record outcomes to improve over time — all via a native MCP interface.
Free tier. No credit card required. Connect with MCP in minutes.
Built for the modern agent stack
Decion gives your AI agents a structured decision loop — submit requests, rank candidates with historical evidence, record outcomes, and retrieve grounded knowledge.
Agents submit compact decision requests with context, candidate actions, and tags. Decion returns ranked recommendations with confidence scores and review flags.
Decisions are ranked using a hybrid of embedding similarity, tag overlap, and historical outcome evidence — powered by pgvector and sentence-transformers.
Ingest text, Markdown, HTML, or PDF documents into searchable knowledge collections. Agents retrieve grounded chunks with citations at decision time.
Every capability is exposed as an MCP tool via a remote Streamable HTTP server. Agents connect with a bearer token and a namespace slug — that's it.
Scope decisions, knowledge, and agent keys to separate namespaces. Run multiple agent environments within a single organization without cross-contamination.
Record the actual action taken and its outcome score. Decion builds per-action evidence that improves future rankings — your agents get smarter over time.
Decion replaces ad-hoc agent decision-making with a structured loop. Agents submit decisions with context and candidates, Decion ranks them using historical evidence, and outcomes feed back to improve future recommendations.
Decion gives coordinator agents a shared namespace for publishing tasks and worker agents a server-side queue for claiming, executing, heartbeating, completing, or releasing leased work. Agents coordinate through the same MCP-native work item, session, subject, knowledge, and decision tools.
Enqueue shared tasks and let worker agents claim leased work items without collisions.
Agents can heartbeat active work, return it to the queue, complete it with results, or hand it off with checkpointed state.
Persist cursors, partial plans, and resumable run state for long-running or multi-agent workflows.
Link work to subjects, timelines, decisions, knowledge, and outcomes so agents build on the same operational memory.
These fixed scenarios preview the decision loop your agents can use in production: rank candidate actions, compare similar cases, and close the loop with outcomes across support, operations, sales, and finance.
Recommend whether to refund, escalate, or request clarification using similar past tickets and policy-grounded knowledge snippets.
Choose between rollback, mitigation, escalation, or communications actions with confidence scoring and review flags.
Rank outreach strategies by account context, prior outcomes, and evidence gathered from knowledge collections.
Route approvals with consistent policy reasoning and tracked outcomes so recurring decisions improve over time.
Use Decion to decide when agents should retrieve context, call tools, escalate to a human, or continue autonomously.
Record selected actions and scored outcomes so similar future decisions can reuse stronger evidence and ranking priors.
We ran 900/100 and 1,900/100 train-test splits across synthetic customer service, IT incident, finance approval, and agent tool-calling datasets. A Codex-agent baseline chose from raw case JSON, while Decion reused imported resolved history through MCP.
Refunds, cancellations, account access, and escalation routing.
Severity triage, containment choice, rollback calls, and stakeholder updates.
Spend exceptions, vendor approvals, renewal risk, and policy routing.
Meta-decisions about retrieval, automation, escalation, and tool sequencing.
Each split uses 100 held-out decisions scored for exact preferred-action match.
Baseline is a Codex-agent judgment from case JSON and ignores preferred_action.
Decion imports resolved outcome history through MCP and can use preferred_action weighting.
Synthetic benchmarks across four datasets; directional evidence, not a production guarantee.
Decisions are only useful when they reach the right person. Decion mailboxes let agents draft, send, and track email from dedicated addresses, so recommendations can turn into customer replies, handoffs, evidence packets, and status updates.
Outbound result
To
customer@example.com
Subject
Decision complete: replacement approved
Message
We reviewed the warranty history, similar cases, and policy guidance. The recommended action is to approve a replacement with expedited shipping.
Decion exposes everything as MCP tools. Point your agent at the server, provide a key, and start making better decisions.
Sign up, create an organization, and generate an MCP API key from the Decion dashboard.
Point your agent's MCP client at the Decion server with your bearer token. All tools are auto-discovered.
Call create_decision with context and candidate actions. Decion returns ranked recommendations instantly.
Submit outcomes after each decision. Decion builds evidence that makes future recommendations smarter.
Every tier includes the full decision engine, knowledge collections, and MCP interface. Scale limits grow as your agents do.
Prototype one agent and validate your workflow.
Small teams running a few production agents.
Teams with multiple agent environments and heavier traffic.
Larger deployments with dedicated Decion team support.
Connect your agent to Decion via MCP and start getting ranked recommendations backed by real outcome data.
Free tier available. No credit card required to get started.