What can AI business agents actually do today, and where do they stop?
They draft. On ESR.co an AI team researches your industry, sets and tracks goals, breaks them into tasks, and writes the strategy, the marketing and the content, using Claude. What they do not do is send, post, publish or book. Email and CRM work comes back as a draft. Social publishing is built into the product but stays switched off until it is turned on, and even then it hands you something to approve rather than posting for you.
What ESR agents do
- Research your industry and specialise to it. You describe the business in plain English and the team is assembled around it, with named roles: a Strategist, a Marketer, a Builder, plus a Board of Advisors.
- Hold goals over time. A goal is not a prompt. It is tracked, broken into tasks, and worked through in phases you can watch in the Command Center.
- Write the work. Strategy, positioning, campaign plans, marketing copy, content drafts. This runs on Claude.
- Keep going overnight on the higher plans. On Growth and Empire the team works while you sleep and queues decisions for you.
Where they stop
Every one of these has an effect outside the product, and an ESR agent never self-approves an action with an external side effect. A person acts:
- Email. Draft only. The agent writes it. You send it.
- CRM. Draft only.
- Social publishing. The integration is built and shipped in the codebase, and it is dormant. Until the publishing credentials are configured there is no Social agent in the picker, no Social card and no “coming soon” tease. It is not displayed at all.
- Booking and scheduling on your behalf. Not something the agents do.
- Ads. Google Ads and Meta Ads are not connected. Ask for a campaign and you get the plan plus step-by-step instructions to run it yourself, not a live campaign.
Why the dormant feature is hidden rather than teased
There is a rule in the codebase behind this, and it came from getting it wrong once. ESR had shipped setup instructions for a Telegram bot that did not exist. One cause: a capability was shown before it was wired. So the social capability now sits behind a single switch that checks two things together, whether the publishing service can be called and whether a tenant’s credentials can be stored encrypted. A key without encryption is not a capability, it is a way to strand someone’s credentials. If either is missing, the feature does not appear anywhere in the interface.
The practical version of that rule for anyone evaluating agent products: a feature you can see in the interface but cannot complete is worse than a feature that is absent, because it costs you the time to find out.
What to do with this if you are comparing tools
Separate the two questions. “Can it produce the work” and “can it deliver the work” are different, and most of the current generation of agent products is much better at the first than the second. Ask for a list of actions the product performs end to end without a human click. A short list is not a weakness. A vague answer is.
Last reviewed July 26, 2026 by ESR AI CORP. Everything on this page describes what ESR.co ships today. Where a number is given it is read from the product’s own pricing module, not from marketing copy.