Switching providers

Henry supports both OpenAI and Gemini. You can switch from one to the other at any time. This page covers the switch, what changes for the team, and how to set up a tested alternative for outage failover.

​

Why switch providers

Common reasons:

• Cost. One provider’s rates fit your team’s usage pattern better than the other’s.
• Existing relationships. Your organisation already uses one provider for other workloads; consolidating reduces administrative overhead.
• Compliance. Specific regulatory or contractual requirements favour one provider.
• Outage failover. The current provider is down; you want Henry working again immediately.

Both providers are competent for Henry’s reading-and-explaining role. The choice does not significantly affect the quality of Henry’s answers for most teams.

​

Switching, in detail

The switch takes effect immediately. There is no migration period or staged rollout; the next question uses the new provider.

​

What changes for the team

For the team, almost nothing changes:

• Henry is still in the same place (bottom-right chat icon).
• Henry answers the same kinds of questions.
• The conversation history from before the switch is preserved.
• Citation format is the same.

The slight differences a careful user might notice:

• Response style. Each provider’s models have subtle differences in phrasing. The overall voice is consistent because Brand Atlas’s system prompt is consistent; the micro-style varies.
• Response time. Different providers and models have different latencies. Differences are typically a few hundred milliseconds.
• Edge-case behaviour. When pressed on unusual questions, the providers behave slightly differently. Most teams do not notice.

​

Setting up tested failover

For teams whose tolerance for Henry downtime is low (organisations that rely on Henry for live work, support teams, multi-time-zone operations), a tested failover setup is useful. The setup:

1. Have accounts with both providers. Active accounts with valid API keys on both OpenAI and Gemini.
2. Document the failover procedure. A one-page runbook stored alongside Brand Atlas operational documentation. Names, the steps, the expected time to switch.
3. Test quarterly. Run a controlled failover once per quarter to ensure the runbook still works.

A tested failover means a 5-minute downtime during a provider outage, instead of a multi-hour wait for the provider to recover.

​

When the current provider is partially degraded

A provider can be slow without being down. Common signs:

• Henry responses are taking 15+ seconds.
• Rate limit messages are appearing despite normal usage.
• The provider’s status page shows incidents on the relevant model.

In partial degradation, the choice is to:

• Wait. Most degradations resolve within hours.
• Switch. Move to the other provider for the duration; switch back when the original is restored.

Both are legitimate. The right choice depends on the team’s tolerance for slow Henry vs. the cost of a temporary switch.

​

What happens to conversation history on switch

Conversation history is stored by Brand Atlas, not by the provider. Switching providers preserves history. The team member’s past conversations remain available; new conversations are processed by the new provider. If the customer ever wants to delete past conversations (a data-handling requirement, a clean-up), this is done from Settings → AI → Henry → Conversation history and is independent of the provider switch.

​

Cost reconciliation across providers

If you switch mid-month, you will receive a partial-month bill from the old provider and a partial-month bill from the new one. Both will be smaller than a full month at either; the total is roughly the same as a full month at one. Brand Atlas does not aggregate the bills; the customer reconciles with each provider directly.

​

Switching back

A switch back is identical to a switch forward. Paste the previous provider’s key (or a freshly created one); the new provider’s key is replaced. There is no limit on how often you can switch. Most customers switch zero or one times in the life of an atlas; some switch back and forth based on cost optimisation.

​

Related pages