Decagon Prompts

Write a Decagon Agent Operating Procedure (AOP) in plain language for this refund workflow: when a customer asks for a refund on [PRODUCT], first verify the order in [BACKEND SYSTEM], then check that it falls within the [NUMBER]-day return window, then confirm the refund amount of [AMOUNT], and only then execute the refund via [PAYMENT TOOL]. Include the exact escalation rule: if the refund exceeds [THRESHOLD AMOUNT], hand the conversation to a human agent with a summary of everything checked. Add the confirmation message the agent should send after the refund succeeds.

Design a Watchtower QA criteria set for our Decagon AI agent. I run a [COMPANY TYPE] support operation and need Watchtower to flag any agent response that: promises a refund above [AMOUNT] without escalation, gives [COMPLIANCE AREA] advice without a disclaimer, uses a tone that is not [TONE WORDS], or skips the order verification step. For each criterion write the pass rule, the fail example, and the severity (critical, warning, info). End with the weekly report format I should review: volume flagged per criterion and the top 3 failure themes.

Plan a Decagon Experiments A/B test to lift our chatbot resolution rate from [CURRENT]% to [TARGET]% over [NUMBER] weeks. Define variant A (current AOP set) and variant B (revised greeting, [NUMBER] extra clarification questions, and a new fallback that offers a callback). Specify the traffic split ([X]% / [Y]%), the primary metric (autonomous resolution rate), secondary metrics (CSAT, escalation rate, average handle time), the minimum sample size of conversations, and the decision rule for promoting the winner. Include the exact prompt I should give the team when launching the experiment.

Build a Decagon proactive agent campaign for [EVENT], e.g. a known outage, a delayed shipment batch, or a billing change. Write the proactive message for [CHANNEL: voice, SMS, or chat] that goes to [SEGMENT] customers affected by [ISSUE], using a [TONE] tone, acknowledging the issue, giving the fix status from [STATUS SOURCE], and offering one clear next action. Then write the AOP branch that handles the two most likely replies: 'this does not fix my problem' and 'I want to speak to a person'. Include the suppression rule so no customer gets the message twice.

Draft the knowledge base tuning task for our Decagon agent covering [PRODUCT LINE]. List the [NUMBER] highest-volume intents from [TIME PERIOD], and for each intent write: the 3 most common customer phrasings, the correct resolution path with the backend system involved ([SYSTEM NAME]), the answer the agent should give when the data is missing, and 2 edge cases that must escalate. Add a maintenance checklist for reviewing these articles every [FREQUENCY] and the signal that tells us an article needs a rewrite (for example, Watchtower flags or repeat contacts above [X]%).

Create a Decagon Agent Assist setup guide for our human agents on the [TEAM NAME] team. Describe how the AI copilot should surface during a live chat: show suggested replies based on [KNOWLEDGE SOURCE], auto-summarize the conversation so far when an agent joins mid-thread, and translate messages between [LANGUAGE A] and [LANGUAGE B] in real time. Write the exact agent workflow: when to accept a suggestion, when to edit it, and when to ignore it. Include the 3 training exercises each agent completes in week one and the CSAT target for the pilot ([X]% on assisted conversations).

Write the Voice of Customer analytics brief I should run monthly on our Decagon Insights data. Ask for: the top [NUMBER] emerging complaint themes with example phrasings, the sentiment trend by [CHANNEL] over [TIME PERIOD], the [NUMBER] intents with the highest escalation rate and their likely root cause, and one product or policy change recommendation per theme with an owner ([TEAM]) and expected impact. Format the output as a one-page exec summary plus a table of themes with volume, sentiment delta, and recommended action. Include the prompt to export this to [REPORTING TOOL].

Design an onboarding plan for launching a new Decagon AI agent for our [BRAND] in [REGION]. Week 1: connect [TICKETING SYSTEM] and [KNOWLEDGE SOURCE], import [NUMBER] historical conversations for training, and draft the first [NUMBER] AOPs covering our top intents. Week 2: run [NUMBER] simulator tests per AOP, set Watchtower criteria, and configure the [ESCALATION RULE]. Week 3: soft launch to [X]% of [CHANNEL] traffic with daily Watchtower review. Week 4: full launch criteria (autonomous resolution above [TARGET]%, CSAT above [TARGET]%, zero critical Watchtower fails for [NUMBER] days). Include the go/no-go checklist.

Create a CSAT recovery playbook for conversations our Decagon agent escalated with a low score. For a [RATING]-star rating on a [ISSUE TYPE] case, write the follow-up message a human agent should send within [NUMBER] hours: acknowledge the specific failure from the transcript summary, apologize without deflecting, offer the concrete fix ([COMPENSATION OR ACTION]), and invite the customer to reopen the case via [CHANNEL]. Add the internal step: log the root cause in [SYSTEM] tagged as [AI FAILURE TYPE], and the rule for when the case should feed back into an AOP rewrite (for example, same failure [X] times in [TIME PERIOD]).

Write the integration setup checklist for letting our Decagon AI agent take real actions in [BACKEND SYSTEM], e.g. [SALESFORCE, ZENDESK, SHOPIFY, or STRIPE]. Cover: the API credentials and permissions to create ([PERMISSION SCOPE]), the read-only calls the agent may use freely (lookup order, account, subscription status), the write calls that need an explicit customer confirmation first (refund, cancel, plan change), the sandbox test sequence of [NUMBER] transactions before going live, and the audit log format every action must write ([TIMESTAMP, AGENT ID, ACTION, CUSTOMER ID]). Add the rollback rule: any action that fails twice in a row must escalate with the full error context instead of retrying.