SynthesisLogic

AI agent that works in your websites and portals

For websites and portals with no API: an agent logs in with a limited account you create, clicks through the screens like a person, downloads or enters the data, stops on anything unexpected, and logs each step with a screenshot.

Book an intro call How it works
A morning with this agent
  1. Logged in to 3 supplier portalsDone
  2. Downloaded 9 new order confirmationsDone
  3. Updated delivery dates in the ERPDone
  4. One portal asks for a new security codeAsks a person

An example. Your agent is built around your own work.

How it works

  1. Log in with its own account

    The agent signs in with a dedicated account that has only the rights the job needs. The password sits in a secrets vault, not in a prompt or a file your staff can read.

  2. Work through the screens

    It opens the pages, fills forms, downloads reports, checks statuses and uploads documents. It reads the page the way a person looks at it, and each step is recorded with the page address and a screenshot.

  3. Check what it found

    Extracted values are compared with the expected format and with your own records. A missing or out-of-range value does not go any further.

  4. Bring the data back, or stop

    Approved values go to your ERP, spreadsheet or report. If a page looks different from what the agent expects, or a payment or legal step appears, it stops and asks a person.

New to AI agents? How they work, in plain words

Real examples, setup steps and tool connections for this agent

An example conversation

An example of how a conversation runs. The grey lines show what the agent does in your systems meanwhile.

  1. CustomerPlease pull yesterday's delivery statuses from the Northline carrier portal and update our order sheet. The portal has no export.
  2. Checked for an API or export first: Northline offers neither for status. Logged in to the portal with the dedicated read-only account. Opened the shipment list for yesterday.
  3. Read 46 shipments: 41 delivered, 3 in transit, 2 marked "delivery attempted". Matched all 46 tracking numbers to open orders in the order sheet.
  4. AgentI have updated 44 orders: 41 delivered and 3 in transit. Two shipments show "delivery attempted" and I have left those for you.
  5. CustomerWhich ones?
  6. AgentOrders 8814 and 8827, both to addresses in the same district. The portal gives the reason as "recipient not available". I did not contact the recipients.
  7. CustomerOk. Notify the customers using the usual template. Did the portal show anything new today?
  8. While on the shipment list the portal displayed a new "Accept updated terms" notice. Did not click it. Stopped at that page and saved a screenshot.
  9. AgentYes. After I finished reading the list, the portal asked for acceptance of updated terms. I did not accept them, because that is a legal step. The screenshot is in the run log for you to read. I will send the customer messages for the two delayed orders now.

How it is built

First look for an API or an export
The browser is the last choice. If a portal offers an API, a report download or a CSV, we use it, because that is steadier than reading screens. The agent works in the browser only where nothing better exists.
A dedicated, limited account
Every portal gets its own account for the agent, with only the permissions the job needs, never a personal login. The credentials are kept in a secrets vault, and any new second factor makes the agent stop and ask a person.
A model that reads pages, code that sets the boundaries
The agent drives a real browser (such as Playwright) and reads each page through its structure or a screenshot, so it can cope with layout changes. The language model chooses the next click within a script that fixes the allowed addresses, the steps for each job, a step limit and a time limit.
The portal's rules come first
Before automating, we check the portal's terms of use. Some forbid automated access. In that case we ask the owner of the portal for permission or an export, instead of going around the rule.
Every step is recorded
Each action is logged with the page address, the time and a screenshot, with passwords and sensitive fields masked. If something goes wrong, you can see exactly what the agent saw.
Stopping is part of the design
An unfamiliar page, an error message, a login challenge or a payment step ends the run and alerts a person. The agent does not retry endlessly on a slow page. Repeated retries trigger an alert.
Values are checked before they leave the browser
What the agent reads is compared with the expected format and your own records, then written to your systems through their APIs. After a portal update, we re-test the key steps. The agent runs in the cloud or on your server.

What you can connect to it

Supplier and carrier portals
Checks order and delivery statuses, downloads invoices, statements and price lists that the portal offers only on screen.
Government and customs websites
Looks up the status of filings and licences, and downloads the documents your team needs. Submissions that carry legal weight wait for a person.
Banking and payment portals (read access)
Downloads statements and checks that payments have arrived. It does not make payments.
Older web applications
Enters or reads records in systems with no API, using a limited account.
ERP (Odoo, SAP, Microsoft Dynamics and others)
Receives the checked data through its API or import, and supplies the records the agent needs to compare.
Accounting software (QuickBooks, Xero)
Takes bills and receipts that were fetched from portals.
CRM (Salesforce, HubSpot and others)
Receives supplier or customer details that were collected from websites.
Excel and Google Sheets
Holds the extracted rows for review, or the data the agent types into a website.
Outlook and Gmail
Sends the run summary, and reads emails with codes or links only where you have allowed it.
Book an intro call

Connects to your tools. You stay in control.

Works with what you have
Supplier, carrier and government portals, older web applications, your ERP, accounting software, CRM, Excel and Google Sheets, Outlook and Gmail.
A person approves what matters
Payments, orders, discounts or anything you choose wait for approval. Every action is logged.
Can run on your own servers
The agent, its data, its logs and even the language model (open models such as Llama, Qwen or Mistral) can run on your servers, so documents and customer data stay inside your network. How that works

What it does not do

  • It does not make payments or accept contracts and terms. Those steps stop the run and wait for a person.
  • It does not use a personal login. It uses a dedicated account with limited rights.
  • It does not get around captchas or ignore a portal's rules against automated access.
  • It does not keep going on a page it does not recognise. It stops and says what it saw.
  • It does not guess a missing or unreadable value. That value goes to a person.

Questions

Is it safe to give an agent a login?

Give it its own account with limited rights, never a personal one. The password stays in a secrets vault, every action is logged with a screenshot and the agent stops for any new security prompt.

What if the website changes?

The agent notices when a page no longer looks as expected, stops and tells us, and we adjust it. After a portal update we re-test the important steps rather than waiting for a failure.

Can an AI agent log in to a website and use it for me?

Yes, for jobs we have defined: logging in, finding the right page, reading or entering data and logging out. It works on a dedicated account and within the addresses and steps you approve.

Why not just use an API or a CSV export?

If one exists, we use it, because it is more reliable than reading screens. Browser work is for portals that offer neither.

Is it allowed to automate a website?

It depends on that site's terms. Some portals forbid automated access. We check the terms of each one and, when needed, ask the owner for written permission or an export before building.

Can it handle two-factor login?

Only in ways you have agreed. If a portal asks for a new second factor, the agent stops and alerts a person. We do not work around security checks.

How long does a job take, and what if a page is slow?

It depends on the portal and the number of pages. Each job has a step limit and a time limit, so a slow or looping page ends the run and alerts a person instead of running on.

Can it also type spreadsheet data into a website?

Yes. It reads the rows from Excel or Google Sheets and enters them into the site's form, then logs which rows went in. If the site has an import or API, we use that instead.

How much does an agent like this cost?

It depends on the task and the systems it connects to. After a first call we give you a clear estimate of cost and time.

Want this agent built around your work?

Tell us in a few lines what the job looks like today and we will set up an online intro call.

Type of organisation