Skip to content
SAVE4545% off every plan.See the offer

Integrations & Setup GuidesAI Provider Configuration

AI Provider Configuration

AutomateNexus CRM includes AI-powered features such as email drafting, lead scoring, conversation summarization, sentiment analysis, and intelligent…

On this page· 45

Overview#

AutomateNexus CRM's AI features, such as Karrie, your AI agents, email and reply drafts, meeting summaries, the Messenger's answers, AI text replies and the phone receptionist, run on an AI provider you connect with your own API key, or on AI credits from AutomateNexus CRM ($5 a month per 5 million tokens) when you have not connected one. No plan includes AI usage: with neither a provider nor credits, these features are off. You can connect OpenRouter (one key for hundreds of models), the model makers OpenAI, Anthropic, Google Gemini, DeepSeek, Mistral AI, z.ai (GLM), Cohere, AI21 Labs and Perplexity, the open-model hosts Groq, Together AI, DeepInfra and Fireworks AI, and servers you run yourself (Ollama, LM Studio, vLLM or any OpenAI-compatible API). Connect several, make one the default, and choose which model runs each kind of work, or let the app choose with Automatic. This guide walks through the setup and explains model selection.

Prerequisites#

Before configuring AI providers, ensure you have:

  • An owner or admin role in your AutomateNexus CRM organization. Other members can see the AI settings but not change them.
  • An account with at least one supported AI provider.
  • A valid API key from your chosen provider with sufficient credits or an active billing plan.
  • An understanding of the AI features you plan to use (this affects which models you should select).
  1. Log in to your AutomateNexus CRM dashboard.
  2. In the sidebar, under Administration, click Integrations.
  3. Choose AI in the left menu of the Integrations page.

The AI page is arranged from top to bottom:

  • AI credits — the tokens your organization has used this month on AI credits, and where an admin adds more.
  • Providers — the providers you have connected, the default one first with a Default badge. The default answers first; when it fails, the others are tried in the order shown. Below them is the Use AI credits when your providers fail switch.
  • Spending limit — once you use a provider of your own, a monthly limit on what your own keys spend.
  • Models — which model runs each kind of work.
  • History — the Keep what the AI writes for 30 days switch. Off, nothing the AI writes is kept, only counted. On, each request and its text are kept for 30 days and listed under AI → AI Reports; people see their own pieces, and owners and admins see everyone's.
  • On this computer — the Ollama on my computer card.

Everyone in the organization can see these settings; owners and admins change them.

Connecting a Provider#

The steps are the same for every hosted provider:

  1. Under Providers, with no provider yet, click Connect OpenRouter, or Other providers for any other one. Once a provider is connected, use Add provider. The Add a provider list has a search box and groups the providers under Many models, one key, Model makers, Open-model hosts and Your own server.
  2. In the Connect dialog, paste your key into API key. The Get a key at link beside it opens the page where the provider makes keys. The key is encrypted when it is saved and never shown again, here or anywhere else in the app.
  3. Choose a Default model. The picker suggests models for the provider; once a key is typed, the Load … models link (named for the provider, for example Load OpenAI's models) fills it with the list your key can use. OpenRouter's catalog loads on its own and can be searched. The default model runs unless you choose another model for a kind of work under Models.
  4. If you already have a provider, switch on Default provider to make this one answer first; your other providers step in when it cannot. The first provider you connect becomes the default.
  5. Click Check connection to send the provider one short request with this key and model, or click Connect, which runs the same check before it saves. When the check fails for a reason other than a rejected key, the button becomes Save anyway.

Each connected provider shows its default model and the kinds of work set to it. Use Edit to change its key or model, and its three-dot menu for Check connection, Make default and Remove…. A provider that no longer lists the model it is set to shows Model gone: choose another model.

Anthropic (Claude) Setup#

About Anthropic#

Anthropic offers the Claude family of models, known for their strong reasoning capabilities, safety features, and ability to handle long documents. Claude models are excellent for email composition, document analysis, and complex conversational tasks.

Getting Your Anthropic API Key#

  1. Go to platform.claude.com (the dialog's Get a key at link opens its keys page) and sign in or create an account.
  2. Click Create Key.
  3. Give the key a descriptive name (e.g., "AutomateNexus CRM").
  4. Copy the API key. It will start with sk-ant-.
  5. Ensure you have added a payment method and set a usage limit under Plans & Billing in the Anthropic console.

Then connect it as described under Connecting a Provider, choosing Anthropic Claude in the Add a provider list.

OpenAI (GPT) Setup#

About OpenAI#

OpenAI provides the GPT family of models, widely adopted and well-suited for general-purpose text generation, summarization, and classification tasks within AutomateNexus CRM.

Getting Your OpenAI API Key#

  1. Go to platform.openai.com and sign in.
  2. Navigate to API Keys in the sidebar (under your profile settings).
  3. Click Create new secret key.
  4. Name the key (e.g., "AutomateNexus CRM") and set permissions to All or restrict to specific capabilities.
  5. Copy the key. It will start with sk-.
  6. Ensure billing is set up in your OpenAI account.

Then connect it as described under Connecting a Provider, choosing OpenAI in the Add a provider list.

Google (Gemini) Setup#

About Google Gemini#

Google offers the Gemini family of models through Google AI Studio or Vertex AI. Gemini models support multimodal inputs and are well-integrated with the Google ecosystem.

Getting Your Google AI API Key#

  1. Go to aistudio.google.com and sign in with your Google account.
  2. Click Get API Key in the top navigation.
  3. Click Create API Key and select a Google Cloud project (or create a new one).
  4. Copy the generated API key.

Then connect it as described under Connecting a Provider, choosing Google Gemini in the Add a provider list. In a workspace connected to Google, a key whose project has billing turned on can be confirmed not to train (see A Google Gemini, Mistral AI or Cohere Key That Does Not Train).

Mistral Setup#

About Mistral#

Mistral AI offers high-performance open-weight models that are efficient and cost-effective. They are a strong choice for organizations that value the flexibility of open models or need European data processing.

Getting Your Mistral API Key#

  1. Go to console.mistral.ai and sign in or create an account.
  2. Navigate to API Keys.
  3. Click Create new key.
  4. Name the key and copy it.
  5. Ensure billing is configured in your Mistral account.

Then connect it as described under Connecting a Provider, choosing Mistral AI in the Add a provider list. In a workspace connected to Google, a key on the Scale plan, or with "Anonymous improvement data" off, can be confirmed not to train (see A Google Gemini, Mistral AI or Cohere Key That Does Not Train).

Groq Setup#

About Groq#

Groq provides ultra-fast inference through custom hardware (LPU). It hosts open models such as Llama 3.3 70B and delivers responses with very low latency, which suits features where speed matters most, such as the phone receptionist and live chat.

Getting Your Groq API Key#

  1. Go to console.groq.com and sign in or create an account.
  2. Navigate to API Keys.
  3. Click Create API Key.
  4. Name the key and copy it.

Then connect it as described under Connecting a Provider, choosing Groq in the Add a provider list.

Other Hosted Providers#

DeepSeek, z.ai (GLM), Cohere, AI21 Labs, Perplexity, Together AI, OpenRouter, DeepInfra and Fireworks AI connect the same way, with an API key from the provider's own console; for most of them, the dialog's Get a key at link opens the right page. OpenRouter is one key for hundreds of models from OpenAI, Anthropic, Google, DeepSeek, Qwen and others, billed by OpenRouter. A Perplexity key also gives your agents web research, and a Cohere key can also do search by meaning (in a workspace that has had Google connected, only once you confirm the key does not train). DeepInfra and Fireworks AI host open models at a low price per token.

Replicate and Straico are not offered, because the app's AI features cannot call them. A key saved for one earlier is shown as never used, with a way to remove it.

Self-Hosted and Local Models#

Your Own Server#

Choose Ollama (self-hosted), LM Studio (self-hosted), vLLM / OpenAI-compatible server or Custom OpenAI-compatible API under Your own server in the Add a provider list. Instead of a key, the dialog asks for:

  • Server address — An address the internet can reach, because the app runs in the cloud; a localhost address is refused. The dialog suggests a free Cloudflare Tunnel (cloudflared tunnel --url http://localhost:11434 for Ollama; port 1234 for LM Studio; port 8000 for vLLM), and Ollama must be started with OLLAMA_HOST=0.0.0.0. /v1 is added for you when it is missing.
  • API key — Only if your server asks for one.
  • Default model — The model your server runs (for Ollama, the name you pulled). Suggested models are offered as chips, or use the Load … models link once the address is typed.

Nothing is billed per token; the server runs on your hardware.

Ollama on My Computer#

Under On this computer, the Ollama on my computer card runs open models on the machine you are using, with nothing to expose and nothing billed per token. It looks for Ollama from your browser and, when Ollama is not running or does not allow the app's address yet, shows what to install or set. Choose an installed model and click Use this computer for AI. It works while a tab of the app is open on that machine.

When You Connect Google#

Once you or anyone in your workspace connects Google Calendar, Google Meet or Google Drive, Karrie, the agents and the other AI that reads your records use only these:

  • OpenAI, with an API key from platform.openai.com.
  • Anthropic (Claude), with an API key from platform.claude.com.
  • OpenRouter. The app asks it to send your requests only to hosts that neither keep nor train on them, and its one key still reaches models from many makers.
  • Groq, Fireworks AI, DeepInfra, Perplexity and AI21 Labs, whose terms say nothing sent through their API trains a model, whatever the key.
  • Ollama on my computer, which runs on the machine you are using, so nothing leaves it.
  • AI credits from AutomateNexus CRM, if you have them.
  • Google Gemini, Mistral AI and Cohere, with a key an owner or admin confirms does not train (see A Google Gemini, Mistral AI or Cohere Key That Does Not Train below).

Every other provider is passed over for that AI: DeepSeek, z.ai (GLM), Together AI, a Google Gemini, Mistral AI or Cohere key nobody has confirmed, and every server of your own: Ollama (self-hosted), LM Studio (self-hosted), vLLM / OpenAI-compatible server and Custom OpenAI-compatible API. Their keys and addresses stay saved, and they are still used for the AI that reads nothing from Google. A Gmail account connected with a Google sign-in before counts as a Google connection too.

Search by meaning keeps the rule too. In a workspace that has had Google connected, finding your records, help articles, agents' memories and Knowledge by meaning is done by your own OpenAI or OpenRouter key, or a Google Gemini or Cohere key you have confirmed does not train, and then by AutomateNexus CRM's OpenAI or OpenRouter. A Gemini or Cohere key nobody has confirmed is passed over for it and never used, and OpenRouter is asked to send that work only to hosts that neither keep nor train on it. A workspace whose search was set up with Gemini's model stays searchable, as OpenRouter runs that model too.

A Google Gemini, Mistral AI or Cohere Key That Does Not Train#

Google Gemini, Mistral AI and Cohere promise not to train on what they are sent only for some keys, and a key does not show which it is. So in a workspace that has had Google connected, an owner or admin confirms it for each key:

  1. On the AI page, open the provider.
  2. Check the key in the provider's own account, as the dialog says, and tick that provider's sentence:
    • Google Gemini: "This key is in a Google Cloud project with billing turned on, so Google does not use what it is sent to improve its products." Google AI Studio shows a key as Paid once its project has billing.
    • Mistral AI: "This key is on Mistral's Scale plan, or training is turned off for it, so Mistral does not train on what it is sent." In Mistral's Admin Console, under Privacy, "Anonymous improvement data" is off (the Scale plan has it off already).
    • Cohere: "This is a production key (a card is on file) with training turned off, so Cohere does not train on what it is sent." In Cohere's dashboard, under Settings, Data Controls, training is off.
  3. Click Save.

The dialog links to the provider's own terms. Once ticked, the key is used for all of the workspace's AI, Google data included: Karrie's chat, the agents, inbox drafts, meeting notes and, for Gemini and Cohere, search by meaning, as an OpenAI or Anthropic key is, and it no longer shows Not for Google data. The tick is saved with your name and the date, and the dialog then says Confirmed on that date. Pasting a new key clears it, so tick it again for the new key. You can untick it at any time, and the key is passed over again for AI that reads Google data.

In the Add a provider list, these three say Used for AI that reads Google data only with a key you confirm does not train. DeepSeek, z.ai (GLM), Together AI and servers of your own have no tick, so they stay passed over.

AI That Reads Nothing From Google#

AI that works only from what you type or upload, your business profile or a website reads nothing from Google, so the rule does not cover it. It uses any of your providers, including the ones passed over above. That covers websites and site copy, Studio's brand kit, the business profile, social posts, form insights, building workflows and apps, What's new posts, card and paper scans, dictation, guide translation and SEO.

On the AI page, a provider that is passed over shows Not for Google data and says what it is still used for. It keeps its Default badge, and you can still make it the default.

When Only Google Calendar, Google Meet or Google Drive Is Connected#

If your workspace's only Google connections are Google Calendar, Google Meet and Google Drive, or some of them, Karrie's chat and the agents can still use the providers that are passed over, with the meetings from Google Calendar left out:

  • Google Meet counts here while it only makes meeting links, which go on Google Calendar.
  • Google Drive counts here while its files are only attached, to a task, a record, an invoice, a ticket, a proof, a deliverable, a document or a chat message. No AI reads an attached file.
  • Karrie's chat does when the provider she asks first is one of them: the one chosen for Assistant under Models, or else your default. The agents do when theirs is: the one chosen for Agents under Models, or else your default. An agent set to a model of its own asks that model's provider first.
  • Karrie then answers without those meetings, and says so when a question needs them. She can still move or cancel one when you ask: tell her the day, the time it starts or who it is with. A card under her answer shows which meeting she found and the change, and nothing happens until you approve it there. The AI provider is never told which meeting it was.
  • The agents then work without those meetings too. When one looks for free times, it says to check them against your Google Calendar.
  • Work that carries on from Google data keeps the rule: a job Karrie hands to an agent while she runs on one of the providers above, or work an agent hands on after reading the Google meetings.
  • The other AI that reads your records, such as inbox drafts and meeting notes, still uses only the providers above.
  • This ends once anyone in the workspace connects Gmail, and while the workspace keeps notes made from Google Meet transcripts.
  • It also ends, from then on, once AI reads something from Google Drive: when records are imported from a Google Drive file, a Google Drive file is added to an agent's Knowledge, or a workflow is started by new rows in a Google Sheet. Before you pick a file from Google Drive to import or to add to Knowledge, the menu says so. Imports and Knowledge files from before November 13, 2026 count too when they could have come from Google Drive: anything since Drive was first connected, and nothing before October 20, 2026.

On the AI page, those providers show Not for Google data and say they are used for Karrie's chat and the agents, without the Google Calendar meetings. When Models puts one of the providers above first for Assistant or for Agents, they name only the other. With Google Drive connected, the page also says that its files are only attached here and no AI reads them. Once one of the uses above has ended this, the page says which one and when, for example "You imported records from Google Drive on Nov 3."

Why#

Google's rules for apps that receive Google data (its Limited Use requirements) say that data must never be used to train an AI model. Calendar events and meeting transcripts become the workspace's meetings, and Drive files you import or add to Knowledge become its records and Knowledge. Karrie, the agents and the other AI that reads your records see them, so the rule covers all of that AI, not only the Google parts.

The providers above promise in their terms not to train on what is sent to them through their API, and Ollama on my computer sends nothing off your machine. Google Gemini, Mistral AI and Cohere promise it only for some keys: Gemini for a key whose Google Cloud project has billing turned on (its free tier may use what it is sent to improve Google's products), Mistral AI on its Scale plan or with training turned off, and Cohere for a production key with training turned off. A key does not show which it is, so you confirm it for each key. DeepSeek, z.ai (GLM) and Together AI have not promised the same for every model and key. A server's address could be any AI service at all, including one that trains on what it is sent, so a server of your own is passed over too. If you run open models on a server, OpenRouter runs the same models and is asked not to keep or train on anything.

Switching a Google connection off does not lift the rule, and neither does the person who connected it leaving the workspace, because the meetings, files and mail it brought in stay. The AI providers page then says which kind of Google connection the workspace had, and when.

A ChatGPT or Claude Plan Is Not an API Key#

ChatGPT Plus, Pro and Team plans work only inside ChatGPT, and Claude Pro and Max plans only inside Claude: neither company lets other apps use them. To run the CRM's AI on OpenAI or Anthropic, create an API key at platform.openai.com or platform.claude.com and add billing there. The provider bills API use per use, separately from any ChatGPT or Claude plan.

Before You Connect Google#

If none of your providers is on the list above, add one first, such as OpenAI, Anthropic or OpenRouter, as described under Connecting a Provider. Then connect Google, and Karrie and the agents keep running on a provider of your own instead of AI credits.

Choosing Models#

The Default Provider and Model#

The default provider answers first, and its default model runs every kind of work that has no model of its own. When it fails, your other providers are tried in the order shown under Providers. Change the default with Make default in a provider's menu, or on the first row of Models, Default model.

A Model for Each Kind of Work#

Under Models, each kind of work can run on a provider and model of its own, or stay on Default model:

  • Assistant — Karrie's chat, and questions asked of a record, a table, the knowledge base or a meeting.
  • Agents — Your agents, the SDR and the AI steps in workflows.
  • Writing — Emails, reply drafts, social posts, SEO and knowledge-base articles.
  • Summaries and insights — Meeting summaries, activity and project analysis, and forecasts.
  • Builders — The website builder and the workflow builder.
  • Customer conversations — Messenger answers, AI text replies and phone calls. The phone receptionist can also name its own model in its settings.

Pick a provider for a row, then its model, and click Save model choices; changes apply to the next request. Assistant, Agents and Customer conversations call tools, so a model that cannot is flagged there. With no provider of your own there is nothing to choose: on AI credits, everything runs on one included model.

Automatic#

Choose Automatic on the Default model row to let the app pick among your own providers' models for every kind of work that has no model of its own, or choose it on one row for just that work. How Automatic chooses sets the approach: Save money, Balanced or Best results. Each row then names the model chosen, and Why explains the choice. A model you chose for a kind of work stays chosen, and Check your providers’ models again refreshes the lists Automatic picks from. An organization set up before Automatic existed is asked once, with Turn on Automatic or Keep my choices.

Cost Optimization Tips#

  • Put cheaper models on high-volume work such as Summaries and insights, and stronger ones on Agents and Writing.
  • Use Set a limit under Spending limit for a monthly limit, in US dollars, on what your own provider keys spend, estimated from the tokens each provider counted at the model's listed price. When it is reached, AI stops using your own keys until the month turns, and owners and admins are told. AI credits are not counted: the credits you buy are their limit.
  • Monitor usage under AI → AI Usage: input, output and total tokens with an estimated cost, by provider, by model and by tool or agent.
  • Give each AI agent a Monthly budget in its Limits section; the agent pauses itself when it reaches it.
  • The AI credits card shows the tokens your organization has used this month on AI credits. Credits come in blocks of 5 million tokens at $5 a month. An admin adds them with Add AI credits on the card, or under Administration → Plans, and an agency shares them with the client organizations on its plan. Calls on your own provider key use no credits.
  • Use AI credits when your providers fail is on unless you switch it off: when none of your providers answers, the included model answers on your AI credits. Switched off, such a request fails instead.

Agent Model Settings#

Response controls live on each agent rather than on the provider. In AI → AI Agents, an agent's Model section has the model choice (Organization default follows the model set for Agents under Models), a Response length in tokens and a Creativity setting, and its Instructions section holds the standing brief that shapes its output.

Testing AI Features#

  1. Test the key: click Check connection in a provider's dialog or its three-dot menu. It sends one short request and says which model answered and how long it took, or why it failed: a key the provider did not accept, an account that cannot pay for a request, a model the provider does not offer to the key, or a provider that did not answer.
  2. Test an agent: open an agent under AI → AI Agents and start a run by hand from its page.
  3. Test the Messenger: with Answer with AI on under Inbox → Support Hub → Messenger, ask a question in the chat on your website.
  4. Test meeting notes: open a recorded meeting under Scheduling → AI Notetaker and ask a question on its Ask tab.

Troubleshooting#

API Key Rejected#

  • Run Check connection: it says whether the provider refused the key, the account cannot pay, or the model is not offered to the key.
  • Verify the API key is correct and complete. A key with a space or a line break in it is refused before it is saved: copy it again.
  • Check that your AI provider account is active and has billing configured.
  • Ensure the API key has not been revoked or expired.

AI Features Are Off#

  • The AI credits card says when Karrie and the agents are off (the organization has no AI provider of its own and no AI credits) or paused (this month's credits are used up). Connect a provider of your own, or add AI credits.
  • If Spending limit shows Reached, AI has stopped using your own keys until the month turns. Raise the limit with Change.
  • With Use AI credits when your providers fail off, a request fails when none of your providers answers.

"This Answer Skipped the Providers That Have Not Promised That"#

Someone in the workspace has connected Google, so Karrie, the agents and the other AI that reads your records use only providers that promise not to train on what they are sent (see When You Connect Google). Add one, such as OpenAI, Anthropic or OpenRouter, or add AI credits.

Slow AI Responses#

  • Put the kind of work, or the agent, on a faster model, or on a Groq-hosted model.
  • Shorten an agent's Response length.
  • Check the AI provider status page for any ongoing service issues.

AI Output Quality Issues#

  • Tighten an agent's Instructions, or the Messenger's Guidance, to give better context.
  • Choose a stronger model for that kind of work under Models, or set How Automatic chooses to Best results. Larger models generally write better, at a higher cost and latency.
  • Provide more context in the records the AI works from (detailed notes, conversation history).

Usage Costs Higher Than Expected#

  • Review AI → AI Usage to see which provider, model and agent use the most tokens.
  • Move high-volume kinds of work to smaller, cheaper models, or set How Automatic chooses to Save money.
  • Set a Spending limit and agent monthly budgets.

Was this page helpful?

Last updated