Connected workflows

The AI assistant: ask, don't type

Choose AI Assistant from the Work menu to open its full workspace. It keeps all of your conversations in the signed-in company together; use Conversations on a phone or the left-hand list on a computer to resume an older chat. The address of a selected conversation can be copied safely: only you can open it in your signed-in company.

On a computer, a small chat button can also sit in the bottom-right corner of the Hub for staff accounts. Click it for a quick recent conversation, or use Open in Assistant to continue it in the workspace. The workspace has a Show assistant popup switch; it only controls that browser's launcher, not the Assistant page or your saved conversations. The preference is saved separately for each account and company on this browser and stays in sync across tabs. Recent in the popup shows your latest three chats; View all conversations opens the full history. Send or clear an unsent popup message before moving to the workspace. Unsent drafts are kept while switching chats, not after leaving or reloading the page.

On a phone or tablet, tap More, then AI Assistant. The conversation fills the screen, under a short one-line "AI Assistant" heading. The conversation list opens as a labelled drawer and closes after you choose one; the Show assistant popup switch and the Assistant guide link live at the bottom of that same drawer, not above the conversation as they do on a computer. Attaching a PDF uses a small 📎 button next to where you type, instead of a separate row. Your typed message and chosen PDFs stay with that conversation if you rotate, resize or switch chats before sending.

The assistant is included on the Company plan. On Solo and Crew it doesn't appear; an owner can upgrade any time under Settings → Billing. If you're on Company and still don't see it, an administrator hasn't turned the feature on yet. Making changes through the assistant is a second, separate switch an administrator turns on: some organisations may have the assistant itself on but actions still off.

What you can ask it

The assistant can look up, for the company you're currently signed into:

Deals: search by name, or list everything in the pipeline with its stage.
Estimates: the line items, categories, tax rate and markup on a deal's estimate, and any variations.
Estimate categories: what category names are already in use on a deal, so you can keep naming consistent.
Tasks: filtered by deal, by a search term, or including completed ones.
The job schedule: one deal's schedule tasks, or a company-wide list of every job with an active schedule.
Pipelines and stages: so it can resolve where a new deal should go.
Templates: their names and the items in each, so you can ask for "the items from the kitchen template" by name rather than by id.
Your staff: who a to-do can be assigned to, so "give it to Mike" works without you looking anyone up.
A deal's own fields: the custom properties your organisation has set up on a deal (consent number, site address, and so on).

Open the assistant from a deal's page and it already knows which deal you mean: ask "what's on this estimate?" without naming it.

What it can do (where your organisation has turned this on)

Where an administrator has turned on assistant actions, you can also ask it to:

Add or edit a draft estimate's line items and categories: "add 10 nails to the Roofing category of this estimate" adds exactly one correctly categorised line. It reuses an existing category name if one close enough already exists, rather than creating a near-duplicate.
Duplicate a line item, and rename a category (every line in it follows the rename).
Import selected items from a template into a draft estimate.
Create a new estimate on a deal, and change an estimate's name, GST rate or default markup. A rename works at any time; the GST rate and the default markup are priced settings, so they only change while it's a draft.
Create a to-do, with a due date, a priority and the staff member it's for: assigning notifies them, exactly as assigning on the board does.
Edit an existing to-do: reassign it, unassign it, retitle it, or change its due date, priority or linked deal. "Move that one onto Sarah and make it due Friday" is one sentence.
Fill in a deal's custom fields. Only the fields you name change, and a field with a set type is checked the same way the deal panel checks it: a number field won't take words, and a drop-down won't take an option nobody configured.
Delete a line item, a category or a whole draft estimate; mark a draft estimate as sent; revert a sent estimate back to draft; create a new deal; or move a deal to another stage.

A sent or accepted estimate stays locked, the same as everywhere else in the app. The assistant can't work around that; it will tell you the estimate needs to be reverted to draft first (itself an action you approve).

When an action is interrupted

If a change is cut short (the connection drops, or the server restarts mid-write), the assistant will not quietly try it again. It cannot always tell whether the change went through, so it says so, tells you to check, and if you ask for the same change again it stops and asks you to confirm first, warning that approving may apply it twice. Different changes are unaffected; only an identical repeat is held back.

Approving a change

The last group above (delete a line item, a category or a whole estimate; mark an estimate sent; revert one to draft; create a deal; move a deal to another stage) are destructive or hard to undo, so the assistant always stops and asks first. You'll see a card in the chat naming the action and what it will affect (the actual line item, category or estimate, by name, looked up from your data rather than from what the assistant typed) with Approve and Decline buttons. Nothing happens until you choose:

Approve runs exactly the action shown: never something re-typed or re-interpreted at that point.
Decline cancels it; nothing is changed, and you can keep chatting.

For a new deal the card also names the pipeline and stage it would land in, not just the deal name, so a wrong pipeline is something you can decline rather than something you discover later. Deleting an estimate names the estimate and how many line items go with it; moving a deal names the destination stage and warns that stage automations can fire.

Adding, editing or duplicating a line item, renaming a category, importing template items, changing an estimate's settings, filling in a deal's fields, and creating or editing a to-do are lower-risk and run straight away: you'll see them appear in the estimate grid live, without needing to reload the page.

There is one exception, where your organisation has turned it on. Once the assistant has read a deal's fields in a conversation, every one of those lower-risk changes also stops for an approval card, for the rest of that chat. A deal's fields can hold answers a website visitor typed into your lead form, and a visitor could write something that tries to tell the assistant what to do. The assistant is told to treat that text as information, never as instructions, and the card makes sure nothing changes unless you say so. The card starts with a short note explaining why it appeared. To make direct changes without the cards again, start a new chat.

If a change takes too long and the assistant stops waiting on it, it says so rather than pretending it failed: the change may still have gone through. It then refuses to make any further changes until it has learnt what actually happened, so nothing gets applied twice. That hold covers everyone in your organisation, not just the one conversation: starting a new chat won't get around it, because asking twice is exactly how a change lands twice. Reading is still fine, so you can ask it to check.

When the change does finish, the assistant updates that line in the chat to say it went through, and the estimate grid refreshes on its own. If nobody can ever work out what happened (the assistant was interrupted and stopped before it could record the outcome), the hold lifts after about fifteen minutes and you get a notification asking you to check that one action before repeating it.

What it still can't do

The assistant only ever sees what's true today: it won't guess at a price or a date it doesn't have an answer for. If it has nothing to tell you, it says so rather than making something up. It can't touch billing, user roles, organisation settings, connected-app credentials, or anything outside the tools listed above; those stay staff-only screens, by design, not by choice of prompt.

Your organisation also has a daily limit on how much you can ask the assistant to do in one day. If you hit it, it'll say so plainly rather than quietly stalling.

There's also a second kind of daily allowance, measured by how much the assistant actually reads and writes rather than by how many messages you send. That distinction matters: asking "what's on today?" is a tiny job, while feeding it a sixty-page supplier quote is an enormous one, so counting messages alone doesn't reflect what the assistant is really doing. It's switched off unless whoever runs your Summit Hub deployment has set one up. There's no screen for it yet, so if you want one, that's a conversation with them rather than a setting you'll find. When it's on and your organisation has used it up, the assistant says so and picks up again the next day. Work your team has running *right now* counts toward it too, so on a busy morning you can see that message while a few big jobs are still finishing, and sending again a moment later usually goes through.

Usage limits

On plans that include the assistant, it also comes with a monthly allowance: a set amount of assistant use each month, included in your plan. Big jobs, like reading a long supplier quote, use more of it than quick questions do.

When the allowance runs out, the assistant stops taking new messages and tells you the date it starts again. Nothing you've already done is lost, and the rest of Summit Hub carries on as normal.
If the assistant is busy partway through an answer, it waits and tries again by itself. If the allowance runs out before that retry, the answer stops there with the same allowance notice.
When the month starts again depends on your plan. If you pay for Summit Hub, it's your billing date. Otherwise it's the 1st of each month.
One person can't use the whole month's allowance in a day. Each person has a daily share. If you reach yours, the assistant tells you so, and the rest of your team can keep using it.
If the assistant runs on your own AI account (see "Which AI account and model the assistant runs on" below), that use is billed to your account and doesn't count toward the allowance.
If the assistant says it couldn't check the allowance, that's a brief problem on our side, not your allowance running out. Try again in a moment.

Seeing how much is left

Account owners can see the team's usage in Settings, Billing, under AI assistant usage. It shows how much of this month's allowance is used, as a percentage and in credits, and the date it resets. Credits measure assistant use; they aren't a dollar amount.

At 80%, the assistant shows a short note above the conversation saying how much is used and when it resets. If you dismiss it, it stays hidden for you until next month. Other people on your team still see it until they dismiss it themselves, and so do you in any other organisation you belong to. Owners also get one email and a notification the first time the team passes 80% in a month. Owners can turn that email off in My Account, like the other notification emails.
At the limit, sending a message shows a note in place of the usual error. It says the allowance is used up and when it resets. Owners get a link to buy more credits; everyone else is asked to check with their account owner.
If you've used your own daily share, the note says so, and the rest of your team can carry on.

Buying more credits

Account owners can top up with a pack of extra credits when the monthly allowance isn't enough.

Go to Settings, Billing and find AI assistant usage. (If the assistant has stopped because the allowance is used up, the Buy more credits link in the assistant takes you straight there.)
Select Buy more credits, pick a pack, then Continue to payment.
Pay on Stripe's checkout page. You come back to Settings, Billing when you're done.

The credits appear on the usage card once the payment clears. For a card payment that's usually within a minute.

Bought credits are used only after the monthly allowance runs out, so the allowance included in your plan always goes first.
They carry over from month to month, and each pack expires a set time after you buy it. The usage card shows how many bought credits you have left and when the next ones expire. The oldest credits are used first.
Only account owners can buy credits. Anyone else who hits the limit is asked to check with their account owner.

For a refund on a credit pack, contact whoever looks after your Summit Hub account. Once a refund goes through, the refunded share of the pack comes off your bought credits by itself: a full refund removes what's left of that pack, and a part refund removes the same share of its credits. Credits you've already used from it stay used. If your team needs a bigger monthly allowance rather than a top-up, talk to them too.

Where the numbers are, and what the terms say

The size of each plan's monthly allowance is on the pricing page, in the plan comparison, and on the usage card in Settings, Billing. The top-up pack's size and price are on the pricing page too, in the assistant section and the plan comparison, and how long bought credits last is shown when you buy one. The Terms of Service (clause 9) explain how the allowance and top-ups work but don't give any of those numbers, so we can change them without changing the terms.

What clause 9 says, in short:

The allowance resets each billing period. Anything you don't use doesn't carry over.
Every request the assistant makes counts, including one that fails or that you stop partway through.
Bought credits work only with the assistant. They have no cash value, you can't transfer them to another account, and they expire after the period shown when you buy them. We refund them only where the law says we have to.
If we cut the allowance on a plan you already pay for, you get at least 30 days' notice by email, and the change starts from your first billing period after that notice ends. A bigger allowance, or a new price for packs you haven't bought yet, can happen without notice.

Using your own AI tools with Summit Hub

*(Only where your organisation has turned this on; it's off by default.)*

If your team already works with a coding-agent tool like Claude Code, Codex or opencode, Summit Hub can hand one of those a temporary, tightly-limited pass to do the same jobs the chat dock does, so you can ask for estimate work from the tool you're already in.

What "tightly-limited" means, concretely:

The pass is tied to one chat conversation, to you, and to your company: it can't reach another company's data, and it can't do anything your own account couldn't already do.
It covers only the specific set of jobs it was issued for. Asking for anything else is refused and written into your organisation's audit trail.
It expires within the hour, usually within twenty minutes, and your owner can cancel any outstanding pass at once from the settings page.
Anything that would normally ask for your approval in the chat is refused outright: an outside tool can't approve on your behalf, and nothing it asks for turns up as an approval card for you to tick. It's told to hand the action back to you, and you run it in the chat like any other.

If instead you'd like Summit Hub to run one of those agent tools *for* you, it runs in a sealed container: no access to any file on our servers, and no way to reach anything on the internet except Summit Hub itself. Every attempt to reach somewhere else is blocked and logged. That option runs on the AI provider key Summit Hub is configured with, set once by the operator, so the spending, and the ability to switch it off, sit in one place.

Note on AI "gateways": some providers (for example OpenRouter, or opencode's Zen) are resellers that pass requests on to the underlying model company. If Summit Hub is pointed at one, assistant traffic goes to that company as well as to the model provider, so the operator has to tick a box confirming that before it's used. The key is stored encrypted and is never shown back in full.

Which AI account and model the assistant runs on

*(Platform operators only, at Overseer → AI assistant model. Configured once for the whole of Summit Hub; there is no per-organisation copy of this.)*

By default the assistant runs on Summit Hub's own AI account, on a model we choose. The operator can point it at a different AI account (a different key, a different bill, a different model) and every organisation's assistant follows. It is one setting because it is one runtime: the model, the address and who processes the traffic are the same for everybody, so they are chosen in one place by the person who runs the platform rather than N times by N owners.

What gets filled in:

Runtime. *Built-in assistant* is the normal choice and works everywhere. *External agent runtime* is the sealed-container option described above, and needs a server set up for it.
Provider. *Anthropic Messages API* is the format the built-in assistant speaks. *OpenAI-compatible endpoint* is only usable by the external runtime: if you pick it for the built-in assistant, it will tell you so rather than quietly falling back.
API key. Use a dedicated key with a spend cap set at the provider. It is stored encrypted; the field shows a masked version (sk-a****wxyz) and leaving it untouched keeps the key already saved. There's a separate tick box to remove it and go back to the built-in model.
Base URL. Blank means the provider's own address. For OpenRouter the address depends on the provider setting above, and the two are genuinely different: https://openrouter.ai/api for the Anthropic format, https://openrouter.ai/api/v1 for the OpenAI-compatible one. The box suggests the right one; if you put the *other* one in, it says so and refuses to save it, rather than letting every answer fail later.
Model. Start typing and the box offers the models that endpoint accepts. A name that endpoint doesn't serve is refused when you save, with the valid names listed, rather than working right up until someone asks the assistant a question. Blank uses that endpoint's default, shown greyed out in the box. With Anthropic directly that's Claude Opus 5.5 (claude-opus-5-5), and Claude Opus 5 (claude-opus-5) is still in the list if you want to stay on it. Through OpenRouter the default is still anthropic/claude-opus-5, and it's spelled differently because the two addresses name models differently. Opus 5.5 isn't offered through OpenRouter yet. If your organisation uses the external runtime, a blank model still means Claude Opus 5 for now, until that runtime has been checked on Opus 5.5. You can pick Opus 5.5 there yourself.

We only hold a model list for the addresses the built-in assistant talks to. For an OpenAI-compatible address (which only the external runtime uses), any name is accepted, the box says so, and a wrong one shows up the next time the assistant runs. Leaving it blank there hands the choice to the runtime's own default rather than to us.

Two things worth knowing:

Model names are spelled differently depending on how you reach them. The same model is claude-opus-4-8 through Anthropic directly and anthropic/claude-opus-4.8 through OpenRouter. The picker handles this; pasting a name from elsewhere often doesn't.
Conversations already open keep the model they started on. A change applies to new conversations, across every organisation, so nobody's half-finished question changes character mid-way.
If someone else saved while your page was open, your save is refused. There is one set of these settings for the whole application, and the panel saves all of it at once, so a page opened an hour ago would otherwise put back the address, model and consent as they were then. Instead it tells you to reload and make the change again.

The panel lists the model addresses this server is allowed to reach: one list for the built-in assistant, another for the external runtime. Adding an address to either is a server change, not a settings change, so if the address you want isn't listed, saving tells you rather than letting the assistant fail later. The built-in assistant runs inside Summit's own servers, so it only ever connects to addresses approved in the server's own configuration.

Watching it work

While the assistant is working, the bar at the bottom of the chat says what it's actually doing right now ("Reading the estimates", "Adding a line item", "Thinking" between steps), with a counter once it's been going a few seconds, so a slow answer reads as slow rather than as stuck. The step in progress pulses in the conversation until it finishes. Press Stop at any point.

The conversation also follows itself down: when the assistant posts, the chat scrolls to the newest message on its own. If you've scrolled up to re-read something, it leaves you where you are; scroll back to the bottom and it picks the follow up again.

When something goes wrong mid-answer

If the assistant runs out of time, or hits a problem, it tells you, and once it can make changes for you, it will also tell you what it had already done before it stopped. Three different things can happen, and the assistant says which one it was:

*"it already completed: …"* That change is saved. Don't ask for it again.
*"these may or may not have gone through: …"* The assistant stopped waiting for something that was still running. Check whether it landed before asking again.
Nothing mentioned: nothing was changed, so it's safe to just ask again.

If you do ask again for something that already went through within the last few minutes, the assistant recognises it and tells you it's already done rather than doing it twice.

Turning a supplier's PDF into estimate lines

On a deal with a draft estimate open, click Attach PDF in the assistant and ask for something like "add these to the estimate". Up to three PDFs, 10 MB each and 15 MB in total per message. If your documents come to more than that, send them across separate messages.

File size isn't the only limit: length is too. Megabytes say very little about how long a document is: a short scanned quote can be far bigger than a long typed one, and it's the *length* that decides how much work reading it is. So there's a second cap on how much the assistant will take in one message. Go over it and the message doesn't run: you get a reply saying roughly how big it was, how big it's allowed to be, and what to do: send a shorter document (or just the pages that matter), or start a new chat to drop the earlier conversation from the message. Nothing is charged for a message that's too big, so trimming and re-sending is safe. (Measuring the length does involve sending the document to the AI provider, the same as any other message; it just isn't answered or billed.)

A long document also limits how many steps one message can take. Answering you often means several rounds (look something up, read the result, look up the next thing) and the assistant re-reads your whole message on every round. A short question has room for all of them; a message carrying a sixty-page quote uses up the same room in about three. If it runs out mid-way it stops and tells you the conversation got too big to keep re-reading, and starting a new chat with just the document and the question is the fix.

The same cap applies whichever runtime your organisation is on. On the external agent runtime the length is *estimated* rather than measured: that setup does its own talking to the AI provider, so Summit Hub never sees the provider's own count, which means the figure quoted in a refusal there is approximate.

The assistant reads the document and shows you a proposal card: every line it found, with its category, quantity, unit cost and markup. Nothing has been added yet. Untick anything you don't want, then press Add N lines: that click is what writes them, and the estimate grid updates immediately.

A few things worth knowing:

The assistant can't change anything in a chat that has a document in it. That's deliberate: a document from outside your business could contain text trying to instruct the assistant, so in these chats it can only ever show you a proposal to approve. For direct edits, start a new chat.
A proposal can only be applied once. Applying it again does nothing.
The estimate must be a draft. A sent or accepted estimate is refused; revert it to draft first.
A file isn't sent until you send the message. Attaching just queues it, so removing one before you press Send leaves no trace. If the send itself is refused (the chat is still working on your last message, or you've hit the daily limit), the documents are discarded too, so the chat isn't left holding a file it never read. Your files stay on the composer, so retrying is one click away.
You can't attach while the chat is working. Wait for the current message to finish, or press Stop.
Prices come from the proposal, not from the card. You tick lines; the figures that get written are the ones the server recorded when it read the document, so what you approved is exactly what lands.
Privacy: an attached document is uploaded in full to Anthropic's API when the message runs, the same as your chat messages are.

What it can see

The assistant only ever sees the company you're currently signed into, never another company's data, even if you belong to more than one. It works from the exact same permissions your account already has; it can't reach anything you couldn't already open yourself.

Privacy

When you use the assistant, your message and whatever data it looks up to answer you are sent to Anthropic's API (the company behind Claude, the model that powers it), the same way your message would leave the building if you emailed a question to a consultant. This only happens when you actively send a message; the assistant never runs in the background.

Every conversation is kept, so your organisation's owner can review what was asked, what the assistant answered, which tools ran, and which changes you approved, from Settings → AI assistant audit. The page shows the newest 50 events; Older events → at the bottom walks back through the full history, and your filters stay applied as you page.

Platform operators can also review assistant conversations across every organisation, for abuse investigation and support, the same disclosed oversight that already applies to how we operate the platform. This access is restricted to a small operator role that itself requires two-factor authentication.

If Summit Hub has been configured with a different AI provider key or an AI gateway (see *Which AI account and model the assistant runs on* above), assistant messages go to that provider instead of, or in addition to, Anthropic, including the reseller in the middle if it is a gateway such as OpenRouter or opencode's Zen. That is a platform-wide choice the operator makes explicitly, with a consent tick, and it can be undone at any time by removing the key.

Be aware that some gateways are routers, not a single company: OpenRouter in particular forwards each request on to whichever model company serves the model you picked, so your assistant traffic is seen by OpenRouter *and* by that company. The consent tick covers both, and the model you choose determines who the second one is.