Autopilot Direct: Connecting Your LLM Provider

Autopilot Direct is the Autopilot subscription that runs on your organization's own LLM provider account. Instead of ACCELQ supplying the model capacity, you connect a provider such as Anthropic, OpenAI, Gemini, AWS Bedrock or Google Vertex AI with your own credentials, and every Autopilot interaction in your tenant is served through that account. This is the model often referred to as bring your own key (BYOK).

The connection is made once, at the tenant level, by a tenant admin under Tenant Admin > Autopilot > LLM Configuration. Until it is in place, Autopilot features stay unavailable to every user on an Autopilot Direct subscription.

This article covers:

• How Autopilot Direct works, and what stays with ACCELQ
• What to have ready before you connect a provider
• Connecting a provider and choosing models by task complexity
• What admins and users see once the provider is live
• Managing the configuration over time, including tenants that mix ACCELQ-managed and Autopilot Direct subscriptions

How Autopilot Direct Works

With Autopilot Direct, ACCELQ is the orchestrator and your provider is the engine. ACCELQ still builds the prompts, routes each task to a model, applies its guardrails and renders the results in the product. What changes is who the model calls are billed to and where the model capacity comes from.

Your provider: Serves all model calls

Every LLM interaction behind Autopilot, from generating a Scenario in Converse to Test Run Triage and Element Healing, is sent to the provider account you connect. Usage appears on that provider's own console and invoice, under your organization's agreements with the provider.

ACCELQ: Makes the calls on your behalf

The calls to your provider originate from ACCELQ servers, not from the user's browser. Your API key or credentials are encrypted at rest and stored within your tenant. If your provider account restricts access by network, contact ACCELQ Support for the addresses to allow.

Embeddings: Stay on ACCELQ's managed service

A small set of Autopilot features rely on embeddings, for example finding similar Scenarios or Actions. The chat and reasoning models offered by most providers do not produce embeddings. When the models you select cannot serve embeddings, ACCELQ runs those specific features on its own managed service and tells you so on the configuration card. No test content leaves ACCELQ for an unknown third party in this case; the managed service is the same one that serves ACCELQ-managed subscriptions.

Credits: Still apply

Autopilot credits still govern usage within ACCELQ, whichever provider serves the model calls. The Usage tab in Tenant Admin > Autopilot continues to report consumption as before.

Keep in mind: Only a tenant admin can view or change the LLM Configuration. Project admins and regular users see whether Autopilot is available, but not the provider, the key or the selected models.

Before You Connect

The setup itself takes a couple of minutes. Most of the effort is on the provider side, in getting a key with the right permissions and spend limits.

An active Autopilot Direct subscription. The LLM Configuration tab is available only when at least one Autopilot subscription in your tenant is Autopilot Direct. Check the Plan Information tab if you are unsure which type you hold. On a tenant with ACCELQ-managed subscriptions only, the tab shows no configuration because there is nothing to connect.

Tenant admin access. The configuration lives in Tenant Admin and cannot be delegated to a project role.

Credentials from one of the supported providers. ACCELQ currently supports:

Generate these in the provider's own console, with permission to invoke models and nothing more. The provider's documentation covers the steps.

Connecting a Provider

Connecting is a single form: pick the provider, enter credentials, confirm the models, test, save. Nothing is applied to the tenant until you save, so you can test as many times as you need.

Select provider: One tile per supported provider

Open Tenant Admin > Autopilot > LLM Configuration and click Connect LLM Provider. The provider grid appears. Select the provider your organization's key belongs to; the credential fields below the grid change to match. A tenant connects one provider, and all Autopilot Direct subscriptions in the tenant share it.

Enter credentials: Fields depend on the provider

For Anthropic, OpenAI and Gemini, paste the API key. For AWS Bedrock and Google Vertex AI, supply the cloud credentials and region the form asks for. The key is masked as soon as it is entered and is never displayed again, including to other tenant admins.

Confirm models: One per complexity level

ACCELQ preselects a sensible model for each of three complexity levels, Light, Standard and Complex, from the models your provider offers. Most admins keep the defaults. The next section explains what the levels mean and when to change them.

Test connection: Validates every selected model

Click Test Connection. ACCELQ makes a short call to each selected model using your credentials and reports the result per model. A test that passes for one model and fails for another usually means the credentials are valid but that model is not enabled on your account or is not available in the chosen region.

Save configuration: Applies to the whole tenant

Click Save Configuration. Autopilot becomes available immediately to every user on an Autopilot Direct subscription; no re-login is needed. ACCELQ does not allow saving while any selected model is failing its test, so fix the model selection or the provider-side access first.

Keep in mind: Test Connection confirms that ACCELQ can reach the model with your credentials. It does not confirm that your provider account has budget remaining. If the provider later rejects calls because a spend limit was reached, Autopilot features start failing for users even though the saved configuration is unchanged. The provider card in Tenant Admin reflects this the next time Test Connection is run.

Choosing Models by Complexity Level

ACCELQ does not send every task to one model. Each Autopilot task is classified by how much reasoning it needs, and the request goes to the model you assigned to that level. Assigning a small model to light work and a large one to complex work keeps quality where it matters and cost down everywhere else.

Light: Fast, high-volume tasks

Short, high-volume tasks where speed matters more than depth, such as classifying, labeling or matching. Default: the provider's smallest current model.

Standard: Everyday generation

Everyday generation and editing, which makes up most Autopilot work. Default: the provider's mid-tier model.

Complex: Deep reasoning

Tasks that span many artifacts or need multi-step planning and deep reasoning. Default: the provider's most capable model.

Changing the defaults

Each dropdown lists the models ACCELQ has validated for the selected provider. Two common reasons to change a default:

• Model approval policy. Your organization has approved specific models for use with internal data. Set every level to an approved model; the same model can be used at more than one level.
• Cost control. You want the Complex level on the mid-tier model to cap per-task cost, accepting lower quality on the hardest tasks.

Use Reset to return all three levels to ACCELQ's recommended defaults for the provider.

Keep in mind: ACCELQ warns when a lightweight model is assigned to Standard or Complex work. You can still save, but expect lower-quality output on those tasks and more retries. Assigning the largest model to every level is safe but makes every light task as slow and as expensive as a complex one.

Keep in mind: Model names in the dropdowns are the current versions ACCELQ has validated for that provider. When a provider releases a new version, it appears in the list after ACCELQ validates it; existing configurations keep their selected versions until an admin changes them.

After the Provider Is Connected

Once saved, the form is replaced by a provider card that shows the state of the connection at a glance.

Provider card: What is connected

The card shows the provider, the model assigned to each level, and when the connection was last tested. Credentials are never shown; the card confirms only that a key is on file.

Model status: Per-model health

After a test, each model chip carries a check or a cross. A cross means ACCELQ could not get a response from that model with the saved credentials, and the reason is shown under the chips, for example a model that is no longer enabled on the provider account or a region that does not serve it. Tasks at that level fail until the model is fixed or reassigned.

Embeddings notice: Which features run on ACCELQ

When the selected models do not produce embeddings, the card notes that ACCELQ uses its managed service for those features. This is informational and does not need action. The What does this mean? link opens the explanation.

Tab indicator: Something needs attention

The LLM Configuration tab shows a warning marker whenever no provider is connected or the last test reported a failure. The marker clears once a test passes.

What users see

When no provider is connected, the Autopilot workspace shows Autopilot isn't set up yet and directs users to their tenant admin. Autopilot entry points in the rest of the product, such as Converse and the AI actions in Design Studio, behave the same way. Once the configuration is saved, these surfaces become active without any action from the user.

Managing the Configuration

The provider card carries a Test Connection button and a menu with the remaining actions. Every change is applied tenant-wide the moment it is saved.

Test Connection: Re-check at any time

Runs the same per-model test used during setup and updates the Last tested time and the model status. Run it after rotating a key on the provider side, after changing model access on the provider account, or when users report Autopilot errors.

Change API Key / Change Provider: Rotate or switch

Opens the setup form with the current provider preselected. Paste the new key to rotate credentials, or pick a different provider tile to switch. Switching providers resets the model selection to that provider's defaults. Autopilot keeps working on the old configuration until the new one is tested and saved, so a rotation causes no downtime.

Example: Your security policy rotates API keys every 90 days. Generate the new key on the provider console, choose Change API Key, paste it, Test Connection, Save Configuration, then revoke the old key on the provider console. Revoke the old key only after the save, otherwise Autopilot fails in the gap.

Edit Models: Reassign complexity levels

Changes the model for one or more levels without touching credentials. Use it when a new model version becomes available, or when your organization approves or withdraws a model. The change applies to new tasks immediately; tasks already in progress complete on the model they started with.

Delete Provider Config: Disconnect entirely

Removes the provider, the stored credentials and the model selection. Autopilot features become unavailable to every Autopilot Direct user immediately, and the tab returns to the No LLM provider connected state. ACCELQ asks for confirmation. Use this when offboarding a provider for good; to switch providers, use Change Provider instead so there is no gap.

Keep in mind: ACCELQ does not revoke keys on the provider side. After deleting a configuration or rotating a key, revoke the old key in the provider console so it cannot be used elsewhere.

Keep in mind: Changes to the LLM Configuration are recorded in the tenant Audit Log with the admin who made them. The key itself is never written to the log.

Reading a Failed Connection Test

The pattern of failures across the three levels tells you where to look before the message does.

Tenants With Mixed Subscriptions

A tenant can hold more than one Autopilot subscription, and they do not all have to be the same type. The LLM Configuration applies only to the Autopilot Direct ones; ACCELQ-managed subscriptions keep running on ACCELQ's own providers regardless of what is configured here.

Keep in mind: Which subscription a user falls under is set in Manage Users. If a user reports that Autopilot is unavailable while a colleague has it, check the subscription assigned to each user before checking the provider.

Practical Tips

• Set a spend limit on the key before you connect it. Autopilot usage is driven by how many people use it and how often. A monthly cap on the provider side turns an unexpected bill into a visible failure you can act on.
• Keep the defaults unless policy says otherwise. The preselected models per level are what ACCELQ validates its prompts against. Changing them is supported, but the defaults are the tested path.
• Test after any provider-side change. Key rotation, model access changes, region changes and billing changes all happen outside ACCELQ. Running Test Connection afterwards catches a break before users do.
• Name the key for ACCELQ on the provider console. It makes usage reporting and rotation unambiguous when the same account serves other tools.
• Rotate, then revoke. Save the new key in ACCELQ before revoking the old one on the provider side; the reverse order causes an outage.