How to choose an AI assistant for a founder or small team

Name the failure you want to prevent

'I need an AI chief of staff' is a broad product category, not a testable requirement. 'I lose client promises between meetings and messages' gives you a concrete workflow to evaluate. So does 'I spend an hour rebuilding context before replying' or 'I need someone to handle repetitive scheduling.'

A founder with five teammates may need personal attention management, while a service business with five agents may need a shared inbox. Similar headcounts do not mean the same product need. Start with the failure, then choose the category.

• Forgotten promise: evaluate commitment capture and source context.
• Lost background: evaluate recall, attribution and correction.
• Repeated cross-tool work: evaluate execution, permissions and recovery.
• Shared customer operations: evaluate a CRM or shared inbox with team controls.

Run a small acceptance test

Use sample data first. Put one request in a message and one promise in a meeting note. Include a sentence that sounds like a task but is only an idea. Then change a date. A useful assistant should help you distinguish what is actionable from what is uncertain, and show enough evidence to correct it.

For an action-capable assistant, inspect the exact proposed text, recipient and destination. Check whether 'done' means a draft was created, a message was sent, or the underlying outcome was achieved. These are different states. Test what happens when a connector is unavailable, rather than judging only the happy-path demonstration.

Check the privacy and cost boundaries

Ask where source data is read, where AI processing occurs, which metadata is retained, whether memory is optional, and what deletion actually removes. 'Runs on your Mac' does not necessarily mean 'processed only on your Mac.' Ask about provider accounts and usage charges as well as the product subscription.

For Berry, the ledger is local, AI processing runs through your Anthropic or OpenAI account, and Berry's servers retain item metadata. Optional hosted memory and screen capture add separate data flows. Read the privacy page before connecting confidential work; Berry does not claim SOC 2 certification or cloud-free processing.

Where Berry is a fit—and where it is not

Berry's current pilot is designed around commitments from supported conversations and meeting notes, organised into Today, Waiting, People and Projects. It is a candidate for Mac-using founders, consultants and solopreneurs who want to know what needs them without rebuilding every thread.

It requires macOS 14+ and your own Claude Code or Codex account. Slack replies require exact-text approval, Gmail is drafts only, and WhatsApp opens the chat for you to send. End-to-end execution is future work, not a current feature. If you need mobile-only access, automatic calendar editing, or a shared customer pipeline, choose around those needs instead.

Measure the outcome after a working week

Count useful commitments recovered, incorrect items you had to correct, and how often you needed to reopen the original app to understand an item. Notice whether the assistant creates another inbox to maintain. A small number of accurate, actionable items can be more valuable than a long impressive-looking summary.

Keep the comparison grounded in your own workflow. The right product is the one you can rely on for the job you are buying today—not the one with the most ambitious roadmap.