Dispatch
Night Shift · New

Set it up once. It checks its own work every night.

Most AI support bots get set up once and never looked at again. Night Shift reads your assistant's conversations every night, works out what went wrong, and proposes specific fixes with the evidence attached. You approve them in one click, or let Autopilot handle the safe ones.

Bring your own AI key and pay your provider at cost — no per-resolution fees. Dispatch plans: $9/month on your own AI key, or $49 and $149/month with Dispatch AI included. 14-day free trial, billed through Shopify.

Your assistant's morning report New

Each morning you get a short report by email or Slack, and as a card in Dispatch and your Shopify admin. Every item says what to change, why, and which conversations prove it.

  • Setting changes you can apply in one click, like handing off sooner or adding a starter prompt
  • Plain-language advice for things only you can change, like your return policy
  • Evidence on every item: links to the exact conversations behind it
  • Anything Autopilot changed overnight, with Undo

Night Shift report

Tuesday, 6:00 am

Example
142conversations read
3suggestions
1applied by Autopilot
  1. SettingLow risk

    Offer a person sooner on damaged-item chats

    Hand off: When it's stuck → Early

    Shoppers with damaged orders asked for a person several times before being offered one.

    9 conversations as evidence
  2. SettingLow risk

    Add a starter prompt: “Do you ship to Canada?”

    Asked more than any other question last night, usually as the first message.

    12 conversations as evidence
  3. Advice

    Say when refunds arrive in your return policy

    Your policy doesn't say. Consider adding: “Refunds post 5–7 business days after we receive your return.”

    14 conversations as evidence
Autopilot shortened replies (Standard → Short), a low-risk setting.
An example report. The store, numbers and findings are made up to show the format.

AI support drifts. The transcripts go unread.

An AI assistant is only as good as its last setup. Products change, policies change, and shoppers ask things nobody planned for. The answers are all in the transcripts, but few teams have time to read them.

Set up once, then left alone

Most stores tune their bot during setup and move on. Settings that made sense at launch quietly stop fitting.

Hundreds of chats, no reviewer

The questions it fumbled, the handoffs that came too late, the shoppers who left after a recommendation: it's all logged, and rarely read.

Small gaps, real costs

One missing line in a policy can mean the same question answered badly every day, or handed to your team every day.

How it works

  1. It reads last night's chatsA sample of the period's conversations: transcripts, tool results, handoffs, thumbs, refusals and sales from chat.

  2. It proposes fixes, with evidenceConcrete setting changes and plain-language advice, each linked to the conversations behind it.

  3. You approveApply, dismiss or snooze each one. Or switch on Autopilot and let it apply the low-risk settings itself.

What it looks at in each run:

  • Resolution rate
  • Handoff rate and timing
  • Unanswered or “I don't know” replies
  • Policy gaps
  • Failed tool calls
  • Repeated questions
  • Drop-offs after a recommendation
  • Offer results

Then the loop closes: the next run checks whether each change helped, and suggests reverting it if it didn't.

What a suggestion looks like

Two kinds of output. These are examples to show the format, not results from a real store.

Setting changes

Typed, checked, one click to apply

Each one changes a single setting, shows the current and proposed value, why, the evidence, the expected effect and how confident it is.

Examples
  • Hand off sooner: When it's stuck → Early, for damaged-item chats.
  • Add a starter prompt: “Do you ship to Canada?”
  • Add an announcement: “Holiday shipping cutoff is Dec 18”, because shoppers keep asking about delivery by the holidays.
  • Shorter replies: Standard → Short, because shoppers on mobile drop off mid-answer.
Advice

For the things only you can change

Free-form suggestions about your store, written by the AI, with the conversations behind them.

Example Your return policy doesn't say when refunds arrive, and shoppers asked about it several times this week. Consider adding “Refunds post 5–7 business days after we receive your return.” Example Three shoppers asked whether the Trail Jacket is waterproof. The product description doesn't say. Add it, and the assistant can answer.

Smart defaults. Tune what matters. Night Shift tunes it for you. New

Dispatch works well out of the box. When you want it to sound and behave like your store, the settings that matter are a click away, and Night Shift watches the conversations and suggests which ones to change. Every setting Night Shift can tune is on a fixed list with safe limits and a risk level.

One-click presets

Concierge, Efficient, Sales-forward or Hands-off. Preview exactly what changes before you apply.

Scheduled presets

A Black Friday tone or weekend handoff rules between two dates, then back to normal by itself, with history kept.

Brand vocabulary

Words to prefer and words to avoid: “say ‘bralette’, not ‘bra top’”, “never say ‘cheap’”.

Off-limits topics

Topics the assistant won't discuss, like medical advice, answered with your own deflection line.

Sensitive-issue escalation

An allergic reaction, a chargeback threat or harassment goes to a person, per category, right away or after one reply.

Announcements

Time-boxed facts like “holiday shipping cutoff is Dec 18”. The assistant knows them only while they're active.

Your own FAQ

Question-and-answer pairs that take priority over everything else, for the answers you want word for word.

VIP routing

Signed-in customers with tags you choose get a person offered sooner.

You stay in charge

Night Shift suggests; you decide. It's built so a bad night can't break your store's assistant.

You approve by default

Nothing changes until you press Apply. Dismiss or snooze anything you disagree with.

Every change versioned

Each change records who made it (you, Night Shift or Autopilot), what changed and why. Roll back any version in one click.

Autopilot stays in its lane

Opt-in, and only for low-risk settings within safe limits. It never touches money, offers, models, billing, policies or security.

Checked by the server

The AI only proposes. Every proposal is validated against a fixed list of tunable settings and their limits before it can apply.

Shoppers can't steer it

Transcripts are untrusted data. A shopper's message can't turn itself into a settings change.

Personal data minimised

Emails and phone numbers are redacted before analysis, and each run reads a capped sample.

What it costs

Night Shift is opt-in on every plan, with no add-on fee. You see the cost in settings before you switch it on.

How oftenRuns a monthOn Dispatch AIOn your own AI key (runs billed by your provider)
Nightly (default)about 30about 60 conversationsabout 30 runs
Every 3 daysabout 10about 20 conversationsabout 10 runs
Weekly4 or 5about 8 to 10 conversations4 or 5 runs
Run nowwhen you press itabout 2 conversations1 run

Each run is capped at a fixed size, so it counts as about 2 Dispatch AI conversations. A run is skipped if your conversations for the month are nearly used up, so it never takes the last ones from shoppers. On your own key, each run is capped at about 60k input and 4k output tokens: typically under $0.10 a run on Haiku-class models, and up to about $0.45 on the largest models, at list prices. Your provider's invoice has the real number.

How it compares

Most AI support bots never check their own work. Some helpdesks now suggest knowledge-base fixes from conversations their AI couldn't resolve: Gorgias has Opportunities, Intercom has content recommendations for Fin (with its Pro add-on), and Zendesk has an automation potential report. Night Shift does a different job:

It tunes behaviour, not just content

Handoff timing, starter prompts, reply length and which features are on, as well as advice on your policies and product pages.

Nightly, on every plan, with undo

No add-on to buy. Every change is versioned and one click to roll back, and the next run checks whether it helped.

Competitor features as documented by each vendor, checked 19 October 2026: Gorgias Opportunities, Intercom content recommendations, Zendesk automation potential report. Full comparisons.

Questions

Does Night Shift change my settings without asking?

Not unless you turn on Autopilot. By default every suggestion waits for you: Apply, Dismiss or Snooze. With Autopilot on, only low-risk settings inside safe limits apply themselves, and each one is listed in the morning report with an Undo button.

What will Autopilot never touch?

Money (offer amounts and caps), which AI model you use, billing, your legal and policy text, and anything security-related. Those are always suggestions you decide on, or advice for you to act on.

Can I undo a change?

Yes. Every change, whether you applied it or Autopilot did, is saved as a new version of your settings with who made it and why. Roll back to any earlier version with one click.

What does it cost?

Night Shift is opt-in on every plan. On Dispatch AI, each run counts as about 2 conversations. That's about 60 a month if it runs nightly. On your own AI key, it runs on your key instead, capped at about 60k input and 4k output tokens a run: typically under $0.10 a run on Haiku-class models, and up to about $0.45 on the largest models. You see the cost in settings before you turn it on, and a run is skipped if your Dispatch AI conversations for the month are nearly used up.

How often does it run?

Nightly by default. You can switch to every 3 days or weekly, or press Run now on the dashboard whenever you want a fresh report.

Can a shopper trick it into changing my settings?

No. Transcripts are treated as untrusted data: a shopper's message can never become a suggestion on its own. The model only proposes; the server checks every proposal against a fixed list of tunable settings and their safe limits before it can be applied.

What happens to shopper data in the audit?

Emails and phone numbers are redacted before conversations are analysed, and the audit reads a capped sample, not your whole history. It uses the same AI setup as your assistant: Dispatch AI or your own key.

How do I know a change actually helped?

The next run checks. If a change didn't improve the numbers it was meant to move, Night Shift suggests reverting it.

The AI front door to your helpdesk

Answers shoppers instantly, sells while it helps, and hands the hard ones to your team in the tools they already use.