Tidewell

An AI support agent for online stores: it answers from your help center and order data, shows its sources, and hands off to your team when it should.

Made withtw-answer 3.2 modelRAGConfidence-based handoffModel-judged test sets

Watch the film Build something like this

Facts

Status: Live
Platform: Web app + mobile · site chat widget
Channels: Web chat · Email · App
Flows: Inbox · Help center · AI training · Reports
Models: tw-answer 3.2 · tw-retrieve 2.1 · tw-judge 1.4
Handoff: Below the confidence threshold, on a guidance rule, or on request
Year: 2026

FIG. 01 / PRODUCT FILM

Muted preview · play with sound for the full cut

Who it's for

Support leads and agents at online stores whose customers ask about orders, returns, refunds and shipping in web chat, email or the app. For a team that wants the AI to answer what the store's own sources cover, show where every answer came from, and hand everything else to a person with the story already written.

The problem

A store's support questions — where is my order, where is my refund, do you ship there — already have answers in its help center and its order data. A chatbot that guesses sounds just as sure when it is wrong, and when it gives up, the customer explains everything again to a person who starts from zero. The questions it cannot answer disappear instead of turning into better articles.

The product

Tidewell answers web chat, email and in-app messages from your help center, a read-only order lookup and your store policies, and every sentence in an answer carries a numbered citation to the source it came from, with a confidence score beside the reply. When that score falls below the threshold you set, a guidance rule applies or the customer asks for a person, the AI pauses and hands the chat to an agent with a summary, the order timeline and suggested replies. Questions it cannot answer become knowledge gaps in the help center, where it drafts the missing section from your agents' own replies, and every change to its guidance runs against a test set of real, anonymized chats, judged answer by answer, before it goes live. Reports shows the week: AI resolution, handoffs and why they happened, customer ratings and the most-cited articles.

FIG. 02 / HOW IT WORKS

How it works

Step 1: Inbox
Every answer is built from sources. Tidewell reads the order and the help center, and each sentence settles in only once its source is attached, numbered to a citation under the reply, with a confidence score beside it. When the score falls under the threshold, a guidance rule applies or the customer asks for a person, the AI pauses and the chat passes to an agent with a summary, the order timeline and three suggested replies.
Step 2: Help center
Questions the AI cannot answer are grouped into knowledge gaps, with how many chats asked each one and how many were handed off. For a gap, Tidewell drafts the missing article section from your agents' own replies, citing the reply behind every sentence, and previews how the AI answers with the section in place — its confidence against the handoff line. The draft is replayed on the past chats that asked it, and nothing goes live until someone accepts and publishes it.
Step 3: AI training
Changes to the AI's guidance — its rules for tone, handoffs and what it may promise — run against a test set of real, anonymized chats before they go live. Each answer is judged against a reference: is it grounded, does it follow the rules, is it within the word limit. The run shows every question's confidence against the handoff line, the live version's level beside the new one and the failures to review; publishing unlocks when the run is done.

FIG. 03 / SCREENS

The product, screen by screen

Tidewell Inbox on desktop: open conversations marked Needs you or AI answering with confidence meters, a refund conversation in which the AI's answer cites the order lookup and a help-center article, an amber banner saying the AI paused because the customer asked for a person, the customer's view of the chat widget, and a side panel with the handoff summary, the order timeline and three suggested replies.

Tidewell Help center on desktop: five knowledge gaps ranked by chats, an article with a new AI-suggested section in a dashed outline and every sentence numbered to its source, the agent replies it was drafted from with the lifted phrases highlighted, and a preview whose confidence rises from 0.52 to 0.91, past the 0.75 handoff line, above a replay of 18 past chats.

Tidewell AI training on desktop: the guidance rules of a draft version with one edited and one new rule, the knowledge sources and the handoff threshold slider, a finished 40-question test run as confidence bars against the 0.75 line, resolved, handed-off and failed counts compared with the live version, two failed cases to review, and one case open with its cited answer and the judge's checks.

Tidewell Reports on desktop: a week of tiles for conversations, AI resolution, handoffs, customer rating and first response, daily bars split into AI-only and handed-off chats against last week's marks with today's bar hatched as in progress, the reasons chats were handed off, the most-cited articles and the latest customer ratings.

Tidewell Inbox on a phone: the chats that need a person as avatars along the top, a refund conversation with the paused-AI banner and its reply-due time, the AI's cited answer, the handoff line from the AI to the agent, a suggested reply and the reply field, above the tab bar.

Tidewell Help center on a phone: the knowledge gaps with their chat counts, the AI-suggested section with its numbered sources, the AI's confidence rising past the handoff line to 0.91, the replay of 18 past chats, and Accept section and Redraft buttons.

Tidewell AI training on a phone: the finished test run as confidence bars against the handoff line, resolved, handed-off and failed counts beside the live version's, one case with its cited answer, and a Publish button.

Tidewell Reports on a phone: four tiles for conversations, AI resolution, handoffs and customer rating, the week's conversations by day against last week's marks with today's bar hatched as in progress, and the reasons chats were handed off.

FIG. 04 / DESIGN DIRECTION

Design direction

Warm and calm: a paper-warm background, white cards on hairline rules, and no gradients, glows or heavy shadows. Bricolage Grotesque carries the wordmark, page titles, names and big numbers; Figtree carries everything else, with tabular figures on every number that changes. One deep teal belongs to the AI — its avatar, primary actions, citation markers and resolved bars — and amber marks what needs a person: handoffs, due times, low confidence. A dashed teal outline appears only around text the AI suggests, until someone accepts it.

Palette

Paper#FAF7F2
Ink#231F1C
Teal#0F766E
Citation tint#DDF0EC
Handoff amber#C27A2C

Type

Bricolage Grotesque: Wordmark, page titles, names, big numbers
Figtree: Conversations, articles, tables, labels

FIG. 05 / MOTION LANGUAGE

Motion language

“Grounding tide”: Tidewell only says what it can source, and it knows when to let go — the motion shows exactly that. A sentence appears as a faint draft and inks in only when its source docks: a quote chip lifts from the order card or an agent's reply, glides to the answer and becomes its numbered citation. Every level rises from its baseline like water — confidence, test results, a day's conversations — against a visible tide mark, the handoff line or last week's level. A handoff passes as a small baton from the AI to the agent, whose ring then closes. Text never types itself out, and levels settle without bouncing.

Relay

Live

A control room for multi-agent AI workflows: agents pass work along a visible canvas, every step and cost is traced, and nothing ships until someone signs off.

Claude Opus 5.5Claude Sonnet 5Claude Haiku 4.5