What Can an AI Agent Do on Your Behalf? Practical Examples

“Take care of it” works when the recipient understands the business and their authority. An AI agent needs that context, plus technical access to the tools. It can collect information, prepare files, fill forms or carry out an authorised operation. A prompt alone does not give it access to an inbox, shop or CRM.

The examples below separate vendor reports from our proposed pilot assignments. Product availability was checked on 1 October 2026. Muse has not launched in Poland or the EU. Dots are excluded from Pro in Poland, while business plans have different rules. See the dots and Muse guides for those distinctions.

What the published examples show

OpenAI describes a tester whose dot prepared a previously unsent invoice and sent it after approval. Its announcement also describes preparing code fixes for review. Meta presents Muse handling shopping tasks and helping prepare campaigns using connected business information. OpenAI dots, Meta Muse, Muse for Small Business

The supplier materials omit failure rates, total review time and retry costs. They illustrate a workflow worth testing: the agent prepares the material needed for a decision and can carry out the next step within its authorisation.

Five assignments worth testing

These are our proposed pilot scenarios. We are not claiming that all four products support every application involved without configuration.

For purchases, inspect currency, tax, delivery and quote validity. For travel, check airports, time zones and refund conditions. These details determine whether a promising search result is useful.

API access and browser control

An API accepts a defined operation, such as reading an order or creating a draft. The integration determines the data format and permissions. MCP can expose such tools through a common protocol; the protocol itself does not provide an account, a right to use the data or a finished business integration.

A browser exposes a website's interface. The agent reads elements, selects a field, enters text and checks the result. The browser documentation for Hermes and OpenClaw describes these capabilities, with local and external options depending on configuration.

Where a suitable API exists, it is usually a good starting point: the outcome and handling of repeated writes are easier to define. Browser control is useful when an application lacks an API for the required operation. Form changes, expired sign-ins or security checks can interrupt it.

One awkward failure deserves explicit handling: the agent clicks Submit but never receives confirmation. Clicking again may create a duplicate. A sound implementation checks the operation's state first, using an identifier or transaction history where available. If the outcome remains uncertain, it hands the case to a person.

A reusable assignment brief

Goal: compare three quotes for a laptop purchase.
Sources: only the Q4 Quotes folder and the specified manufacturer pages.
Scope: match model, quantity, warranty and delivery.
Output: a table citing every amount, plus missing information.
Access: read sources; create a new file in the For Review folder.
Approval: show recipient and message before sending anything.
Limits: no purchase, payment or change to source files.
Stop when: prices conflict, access fails or a write has uncertain status.
Trial budget: up to €1 for model and tools, then ask for a decision.

The budget is a limit you choose. Actual costs need measurement, and enforcement depends on the product. For a custom deployment, enforce it in the execution layer or provider controls and check the result. Folder access and sending permission also need configuration beyond the written brief.

A document or website can contain text that tries to redirect the agent. Reading it should not grant new authority. An instruction hidden in a quote might ask for a customer database to be sent elsewhere. Restricted accounts and tools limit the possible consequences even if the model misreads that content.

Measure the benefit without inventing savings

Measure manual completion time before the pilot. Afterwards, count review, corrections and failures. The difference shows recovered capacity.

For illustration, 40 monthly comparisons at 15 minutes each take ten hours. If checking and correcting agent output takes five minutes per comparison, the difference is about six hours and 40 minutes. Exception handling and maintenance still need to be deducted. These hypothetical figures need to be replaced with measurements from your pilot. Recovered time does not automatically become cash savings.

Track accepted outputs, model and tool spend, and any operations outside the agreed scope. If correction takes longer than doing the work manually, narrow the task or reconsider the approach.

Deploying a private agent with Syntalith

We can deploy Hermes or OpenClaw on your infrastructure and configure it for your work. The scope includes source access, tools, output locations, approval points and handling interrupted operations. Privacy requirements also cover the model, browser, communication channels and backups.

A personal AI operator starts at €600 excluding VAT. Hosting, model usage, external tools and maintenance are separate costs. This starting point covers a bounded personal setup; workflows spanning several employees and transaction systems require their own scope and quote.

Bring one task and an example of a correct output to a free process scan. Our dots, Muse, Hermes and OpenClaw comparison explains the technology choices.