AI & Automation
AI agents vs workflow automation: when browser use makes sense
A browser agent can work in software without an integration. Here is how I would decide which steps need AI, which need fixed rules, and where a person should approve the result.
· 5 min read
If a customer request arrives with a name, an email address, and a selected service, I would start with ordinary workflow automation. Save the record, check for a duplicate, and assign the follow-up. There is no reason to ask a model to work out those rules every time.
The decision changes when the request is a rambling email, a vendor portal returns unfamiliar information that someone must interpret, or someone has to interpret conflicting information. That is where I would consider AI. I would still decide exactly where its judgment belongs.
On August 26, 2026, Anthropic made Claude in Chrome generally available on paid Claude plans. It can read pages, navigate, and fill forms using existing browser logins. That opens another route into software without a convenient integration. Read the August 26 announcement (checked September 6).
An agent decides some of the steps
Workflow automation follows a path you define. A form submission creates a lead. A due date creates a reminder. A rule sends an incomplete record back for review. That workflow can include an AI step, such as summarizing a message, while the surrounding process stays fixed.
An AI agent has more room to choose its next action. It might read the request, look up a record, notice missing information, and decide what to check next. Anthropic describes the distinction as predefined code paths versus a model directing its own process and tool use. Its explanation of workflows and agents is useful here.
My preference is to give that flexibility a specific job. Interpreting an unclear message may deserve it. Deciding whether a customer has approved a price should have an explicit rule and a record behind it.
Browser use and API access are a separate choice
An API is a defined way for software to exchange information or request an action. Browser automation works through the interface a person uses. An AI agent can use either. A fixed workflow can use either too.
Check the existing product's supported integrations first. For a known action like creating a draft record, I want defined fields and a response I can verify. A suitable API often gives me a clearer starting point. It still needs permission checks, error handling, and maintenance.
Browser use becomes interesting when the useful information is trapped in a portal with no suitable API. I would consider it for a bounded task, such as checking a submission status and preparing an internal update. I would also check whether the vendor allows the proposed access and what happens when the login expires or the page changes.
Anthropic's August 20 platform release added a browser-use tool that combines screenshots with page structure. That developer tool is separate from the Chrome extension. Both are relevant to the integration question, but choosing one still depends on the actual application. Read the platform release.
| What the task needs | Starting approach | What I would verify |
|---|---|---|
| The same known action for each complete input | A workflow with fixed rules | Required fields, duplicate handling, and a recorded result |
| Interpretation of an unclear message | An AI step inside the workflow | The original source, missing details, and review of the suggestion |
| Choosing the next check requires interpreting unfamiliar information | A bounded agent using appropriate tools | Allowed actions, stopping conditions, and the person receiving exceptions |
| Work in a portal without a suitable API | Browser use within the chosen workflow or agent | Permitted access, login recovery, and confirmation before repeating a write |
A customer inquiry, worked through
Here is a hypothetical setup for a service business. A customer emails about a damaged tree, includes an address, and asks whether someone can visit this week. The request needs a record and a response. It does not yet justify a booking or a quote.
Have the workflow retain the original message and check whether that message identifier has already been processed. An AI step could suggest the service category and summarize the request. Missing details would stay visibly missing. The system should not invent an address or turn a customer's preferred day into a confirmed appointment.
If the job requires checking a vendor portal, a browser agent could retrieve the relevant status within an agreed scope. The workflow would save the result and its source. A person would then review the record and a drafted follow-up before anything consequential went to the customer.
That review needs to show the recipient, message, and proposed commitment. A vague 'approve' button is not enough. If the draft says Wednesday at ten, the reviewer needs to see that Wednesday at ten is actually available.
You can see a related boundary in the Carle Books draft-invoice walkthrough. It illustrates an internal tool with fictional records. Preparing a draft gives a person something concrete to review before sending.
Decide what happens when the task stops halfway
Before using a browser agent in a recurring process, I would run a small trial with representative examples. Include an incomplete request, a repeated message, an expired login, and a page that does not show the expected result. A successful run is useful evidence. So is knowing where the system stops and who gets the unfinished work.
For any write, I want a way to verify whether it happened before trying again. If the browser loses its connection after clicking save, repeating the whole task could create a second record. The next step should check the destination, then either continue or ask for help.
Count the time spent reviewing drafts and repairing failed runs. Saving five minutes of typing is less useful if someone spends ten checking the output. A trial should make that tradeoff visible.
Give the agent a smaller job first
Claude in Chrome includes safeguards and an option for manual action approval. Anthropic also says prompt injection remains an evolving problem. I would keep review in place for consequential sends and commitments while assessing a workflow. The announcement explains those controls.
Bring one recurring task to a business process automation discussion. I would map the rules, identify the part that needs interpretation, and check how the existing software can support it. That gives us a reason for each piece we build.