Operate the whole platform by chat.

Most "AI" features in invoicing tools mean one thing: a chat box that drafts an invoice you have to retype anyway. Guliel's AI Companion isn't that. It operates the platform.

Ask it to issue an invoice — it issues the invoice. Ask it to build an automation that sends your accountant the monthly expense bundle — it builds the automation, names it, turns it on. Ask it to run last quarter's revenue by customer — it runs the report and shows the result. Approval gates exist where the action is material; routine operations execute. The AI is the interface layer, not a chatbot bolted onto the side of the product.

Status — the AI Companion is in development alongside the REST API and MCP server that give it actions to call. The design below is what we're shipping. The changelog flags general-availability when it lands. Early-access is available on request — hello@guliel.com.

What the AI can actually do

The action surface is the same one an API caller has and the same one MCP agents have. Three different interfaces, one platform underneath.

Issue invoices

"Invoice Acme for 12 hours at our standard rate, payable in SGD, due net 30." The AI checks the customer record exists, pulls your standard rate from the catalog, picks the SGD currency, sets the due date, issues the document under the appropriate country standard. The PDF lands. If the customer record didn't exist, the AI asks once, creates it on confirmation, and continues.

Log expenses

"That AWS bill that came in yesterday — code it to infrastructure and apply our usual GST treatment." Done. The expense is approved, categorized, and tagged. If the bill hasn't been scanned yet, the AI prompts you to forward it or upload it.

Build automations

This is the part most products won't let AI do. The AI generates an automation definition — trigger, filter, action, action parameters — and offers to enable it. "Whenever stock on Widget A drops below 50, draft a supplier order for 200 units and send it." The AI assembles that into the trigger/action format the automations engine expects, names it, turns it on. You can see and edit the rule like any other automation. Disable it from the automations page. The AI didn't write code you can't read.

Run reports

"Show me revenue by customer for the last quarter, top 10." The result renders in chat. "Export it as PDF and email it to my accountant." The PDF generates and sends. "Schedule that for the first of every quarter." A new automation appears in the list.

Multi-step workflows

"Reconcile last month's supplier invoices against open orders and flag anything where the supplier billed more than 5% over the ordered quantity." Multi-step, with a structured response. The AI walks the data, identifies the variances, presents them in a table you can act on.

Approval gates

The AI doesn't ask for permission to read your data or to format an output. It does ask for permission for actions that are material in the real world:

• Sending an invoice to a customer
• Sending an email to an external recipient
• Drafting and sending a supplier order
• Posting to a webhook
• Changing pricing or canceling a document

Per-action gates are configurable per org. You can set the gate to "always confirm," "confirm above $X," or "trust the AI." For multi-step workflows the AI surfaces the action plan first and asks one confirmation for the batch rather than nagging at every step.

The default for a brand-new org is conservative. The default for a power user who's set their preferences is to execute.

Plan limits

The AI Companion runs on real LLM tokens, which is a real marginal cost. The pricing reflects that:

Plan	AI messages / month
Free	10
Standard ($20 / org / month)	500
Premium ($99 / org / month)	Unlimited

A planned add-on at ~$5-10 / month lifts the cap for power users who don't want a full tier upgrade. Full pricing on /pricing.

A "message" is one conversation turn, not one tool call. Asking the AI to issue an invoice and then asking a follow-up about the result is two messages. The AI making four internal tool calls to fulfill one of those messages doesn't count as four — that's part of one turn.

How it composes with the rest

The AI is the most ergonomic interface, not the only one. The same actions exist on the REST API for code, on the MCP server for external agents (Claude, ChatGPT, whatever you run), and in the automations engine for triggered workflows. A common pattern: you build an automation in chat, the AI generates it through the same interface a developer would use, you edit it later from the automations page. Three faces, one machine.

FAQ

Can the AI delete documents or override compliance?

No. Material deletes go to a confirmation gate every time, regardless of preference, and country-compliance rules (which document types are allowed in which jurisdictions, what numbering sequence is enforced) aren't bypassable from chat. The AI surfaces an error explaining the rule rather than silently doing the wrong thing.

What model does the AI use?

Claude, currently. The model is abstracted behind the platform — we may swap or route based on cost and quality. The user-facing behavior is the same regardless. We don't advertise specific model versions because that's not the unit of capability that matters.

Does the AI have access to my data forever?

The AI sees the data it needs to fulfill the active conversation. We don't train models on your data. Per-organization data isolation applies to AI requests the same way it does to every other API call.

Can the AI build complex automations or just simple ones?

Complex. A 5-step sequenced action chain with branching filters is well within what the AI generates. The constraint is the automation engine's expressiveness, not the AI's — anything that's valid as a hand-built automation is something the AI can produce.

Will the AI ask before it does something destructive?

Yes, by default. Sending invoices, sending external email, mutating pricing, and canceling documents all hit the approval gate unless you've explicitly relaxed the gate in your org settings. The gates are visible and editable. We err on the side of confirming.

What if I want an external AI agent (Claude desktop, custom code) to operate Guliel?

Use the MCP server. The same action surface the in-app AI uses is exposed as MCP tools. Point your agent at our MCP endpoint, authenticate, and it can do everything the in-app AI can do.

Start free at /pricing.