Sign in to continue

Performance

How Percival is doing β€” volume, outcomes, confidence, and what your reviews say.

Daily activity

Outcomes

Review verdicts

Confidence distribution

Self-assessed confidence of each draft in this period

Agent status

Agent settings

Automation

Percival can sweep existing open tickets (not just new ones) and act per the master workflow, after a short hold. These options also control how it summarises handled tickets and reads a customer's prior tickets for context.

When a human agent later replies publicly to a ticket Percival drafted for, Percival compares the two. If they broadly agree, the entry is auto-rated πŸ‘ and removed from your review queue β€” so you won't hand-review it (it does not create a training example; Percival still learns from the human's reply via the normal history mining). Divergences stay queued for you. Tickets where the customer had other open tickets are never auto-rated. Off by default β€” turn on only if you want fewer entries to review by hand and are happy for agreed drafts to skip human sign-off.

Access

Customer access control

Restrict which customers Percival will engage. A blocked customer means Percival stays completely silent on their tickets (no draft, no note, no reply).

Hand-off agents

Percival leaves a ticket alone if one of these human agents has already replied to it.

Ignored Tickets

Specific ticket IDs Percival will never handle, even if reopened. Added automatically when a customer refuses to cooperate or a claim needs human-only handling β€” or add one by hand.

TicketReasonAdded

Knowledge base

Everything Percival learns from, organised into mini headers. Files & canned responses are used as reference (retrieval); tickets are used as worked examples.

Upload a file

Add text snippet

Real ticket history Percival learns from as worked examples (customer β†’ human reply, with context). Percival keeps learning automatically from the tickets it handles and your feedback in the Activity log; hypothetical conversations can be generated in the Playground.

Review examples

Master workflow

Percival follows one grand flowchart meant to cover every situation you can think of. Click a step to edit its context and branches; drag the canvas to pan and use the zoom controls (or Ctrl/⌘ + scroll) for clarity. Each box is a step in plain English; arrows (optionally labelled with a condition) lead to the next step. Reference a Freshdesk field β€” built-in or custom β€” with the field picker (it inserts a token like {{Product Category}}), and Percival evaluates that branch against the live ticket when drafting.

Actions

Reusable multi-step procedures β€” e.g. a Lidl refund: create a supplier ticket with the customer's receipt as evidence, close it with a reason, send the customer the refund canned response, close their ticket. Auto = Percival executes it in deployed mode; Manual = Percival proposes it, a human presses run; Inactive = off.

1 Β· Build the stepscreate the supplier ticket, attach the customer's evidence, send canned responses, set fields, close tickets β€” in order.
2 Β· Say when it runsa plain-English trigger ("Lidl + Scheppach + wants a refund") plus the fields that must be filled in (IBAN, BIC…) β€” it refuses to run without them.
3 Β· Drop it into the workflowpress Copy on the card and paste ACTION: run [Name] into a workflow step β€” Percival triggers it there (or wherever the trigger matches).

Advanced Actions

Bespoke, email-scoped handling for tricky senders. When a requester's email contains your match value (e.g. @lidl), Percival ignores the master workflow and house rules for that ticket and follows ONLY your instructions. Sending still respects the current deployment mode.

⚠ Advanced Actions bypass the customer blacklist. A matching email is engaged even if it's on the block list β€” so an over-broad match (e.g. @ or .com) can make Percival reply to everyone, including bounce, no-reply, and retailer addresses. Match as narrowly as you can, and use Test before you rely on one.

House rules

Standing instructions Percival applies to every reply β€” e.g. Never say good morning or good afternoon β€” always say hello. Before a rule is accepted it's checked against the existing ones: a rule that contradicts or duplicates another is rejected with the reason. Accepted rules override the persona, examples, and workflow wording wherever they disagree.

Add a rule

SLA Plan

For overdue tickets, gradually lower the auto-send confidence bar so a customer who's been waiting still gets a reply. It reuses the draft Percival already prepared (no re-drafting), only lowers the send threshold for that ticket, and adds an apology line at the start. Off until you press Get Overdue Tickets; after that it keeps pulling in newly-overdue tickets automatically.

Failed Tickets

Tickets the scanner tried to answer but gave up on after repeated errors (usually a bad attachment). Each is listed with the error it failed on. Retry re-runs it β€” if it succeeds it drops off the list; if it fails again the new error is recorded.

Recover stranded tickets

Tickets the scanner saw but never engaged β€” no error, no draft. These arrived with a blank subject and the complaint only in the description, which the ticket-list feed drops, so Percival wrongly read them as empty. The list feed now includes the description; this pass re-checks each stranded ticket against the normal gates and treats it like a fresh arrival β€” on deployed it auto-sends when the draft clears the threshold, otherwise it holds a draft for review. Bounded per run.

Activity log

Every ticket Percival has processed, with the response it gave or recommended. Mark each good or bad, and write what it should have said β€” your corrections become worked examples that improve future replies.

Playground

Try Percival on a customer message, or have it generate hypothetical conversations. Everything you create here is saved as a thread below for you to review β€” nothing is sent to a customer.

Respond to real tickets

Enter real ticket numbers β€” separated by spaces, commas, or new lines. Percival drafts a reply to each; you review the full thread and choose which ones actually get sent to the customer.

Draft a reply

Type a customer message and see what Percival would draft. Saved as a thread.

Generate conversations

Claude invents realistic multi-turn threads (customer ↔ Percival) for you to review.

Conversations

Accounts

Create dashboard logins for your team and reset their passwords. The admin account (Daniel) is built in β€” its password cannot be changed and it cannot be removed.

Create an account

UsernameRoleCreated