Yes, AI agents can work with old software that has no API, but how they get in matters more than the AI itself. The options, roughly from most to least reliable, are: reading and writing the system's database or file exports directly, an RPA bot that follows fixed screen steps with AI handling the judgment calls, and a computer-use or browser agent that looks at the screen and clicks like a person. AI automation for legacy systems works well when you pick the most stable entry point available and keep the AI's job narrow; it disappoints when an agent is asked to improvise its way through fragile screens on every run.
Below: why so many businesses are stuck with API-less software, the four integration routes with their trade-offs, a few realistic workflows, where each approach breaks, and when it's smarter to modernize the system instead of automating around it.
Why is so much business software still missing an API?
Plenty of the software that actually runs businesses was built before APIs were a selling point. Think of an on-premise inventory package installed in the early 2000s, a hospital or lab information system with a thin web front end, a government or insurer portal that you log into to file claims, a desktop accounting tool customized years ago by a vendor who no longer exists, or an industry-specific ERP where the "integration" option is a nightly CSV.
Replacing these systems is expensive and risky because they hold years of data and encode business rules nobody wrote down. So teams work around them by hand: someone copies order details from email into the green-screen terminal, someone downloads a report from the portal every morning and pastes it into a spreadsheet. That copying is exactly the kind of work businesses want agentic AI to take on.
What are the ways an AI agent can connect to a legacy system?
There are four main routes. A good integration often uses two of them together.
| Route | How it works | Reliability | Best for |
|---|---|---|---|
| Database or file integration | Read from (and sometimes write to) the system's database, or use its import/export files | High | Reporting, syncing records, bulk updates the vendor's import supports |
| RPA with AI judgment | A scripted bot clicks through fixed screens; AI handles reading documents and deciding values | Medium-high | Repetitive data entry on stable screens |
| Browser automation scripts | Code drives a web app's pages using page structure (selectors), not pixels | Medium | Web-based portals with consistent layouts |
| Computer-use agents | An AI model sees screenshots and decides where to click and type | Lower, improving | Low-volume tasks, exploratory work, screens that change often |
Database or file-level integration
If you can reach the database underneath, you often don't need to touch the screen at all. An agent can query open orders, outstanding invoices or stock levels directly, which is fast and dependable. Many old systems also have an import format (a fixed-width text file, a CSV template, an XML drop folder) that the vendor supports. Writing to that is safer than writing to the database directly.
Be careful with direct database writes. Legacy applications frequently enforce rules in the application layer, not the database, so inserting a row behind the app's back can skip validations, triggers or audit logs. Reads are usually fine; writes deserve a conversation with whoever knows the system best.
RPA plus AI
Robotic process automation (RPA) tools record or script a sequence of clicks and keystrokes. On their own they are brittle about anything unstructured: they can't read a messy supplier email and decide which product code it refers to. Pairing them with an AI model fixes that. The AI reads the email or PDF and produces clean, validated fields; the RPA bot types those fields into the same screens it always uses. We compare these approaches in more detail in AI agent, chatbot or RPA: which one do you need?
Browser automation scripts
For web-based legacy apps, tools like Playwright or Selenium let code drive the browser by finding elements on the page (a field labeled "Policy Number," a button with a known ID). This is more stable than pixel-based clicking and far cheaper to run than a vision model on every step. It breaks when the page structure changes, but those breaks are loud and easy to diagnose.
Computer-use agents
The newest option is an AI model that looks at screenshots and operates the interface the way a person would. It can cope with screens it has never seen and adapt when a button moves. The trade-off is speed, cost per run, and occasional wrong clicks. Our explainer on computer-use AI agents covers how they work and where they're ready for real use.
What does AI automation for legacy systems look like day to day?
Three illustrative workflows, each using a different route:
Order entry into a desktop ERP. A distributor receives purchase orders as PDFs and WhatsApp messages. An AI step extracts customer, items, quantities and requested delivery date, matches item descriptions to the product master, and flags anything uncertain for a person. A scripted RPA bot then enters the confirmed order into the ERP's order screen. Humans review exceptions instead of typing every line.
Daily claims status from an insurer portal. A clinic's billing team logs into a payer portal every morning to check claim statuses. A browser automation script logs in with a dedicated account, downloads the status report, and an AI agent reconciles it against the clinic's billing records, drafting follow-ups for denied or stuck claims. A person approves the follow-ups.
Reporting from an old inventory system. A manufacturer's inventory package has no API, but its database is a standard SQL database on a local server. A read-only connection lets an agent answer questions like "which raw materials will run out before the next scheduled delivery?" without anyone exporting spreadsheets.
In all three, the AI does the reading and judging, and the most stable mechanism available does the typing. That division is the main design principle. If your systems include Tally, SAP or Zoho, our post on adding AI on top of your existing ERP covers the options specific to those.
Where does screen-based automation break?
Any automation that works through a user interface inherits the interface's fragility. Plan for these:
- Layout and version changes. A vendor update moves a field or renames a button. Scripted bots fail outright; computer-use agents may adapt or may click the wrong thing confidently.
- Pop-ups and session quirks. Session timeouts, "your password expires in 3 days" banners, maintenance notices and unexpected dialogs derail scripted flows.
- Login security. Multi-factor authentication, CAPTCHAs and IP restrictions exist to stop exactly this kind of automated access. Work with the system owner on a sanctioned service account rather than trying to defeat them.
- Speed. Screen-driven steps are slow compared with an API call. Hundreds of records a day is usually fine; tens of thousands may not be.
- Silent wrong entries. The worst failure isn't a crash, it's a value typed into the wrong field that nobody notices. Read-back verification (after saving, read the record and compare it to what was intended) catches most of these.
- Terms of service. Some portals prohibit automated access. Check before you build, especially for government, banking and insurer portals.
How do you make legacy integrations reliable enough to trust?
A few practices separate the integrations that run for years from the ones that get switched off after a month:
- Pick the most stable layer available. Database read beats file import beats page selectors beats screenshots. Use the screen only for the steps that truly require it.
- Keep the AI out of the clicking where possible. Let AI decide what to enter; let deterministic code decide how to enter it.
- Validate before writing. Check extracted values against master data (valid customer codes, item codes, tax rates) before anything touches the legacy system.
- Verify after writing. Read the record back and compare.
- Queue, don't block. When the legacy system is down or a step fails, park the item in a queue with a clear reason, alert a person, and retry later.
- Log screenshots on failure. A screenshot of the moment something broke saves hours of guesswork.
- Give the bot its own credentials with only the permissions it needs, so its activity is auditable and separable from staff.
When should you modernize instead of automating around old software?
Automating around a legacy system is a bridge, not always a destination. It makes sense when the system is stable, the vendor isn't going away, and the volume of manual work is the real pain. It makes less sense when:
- the screens change often enough that you're constantly repairing bots;
- the system can't keep up with your volume even when automated;
- the system is a security liability (unsupported operating system, no vendor patches);
- you're automating a workaround for a process that should be redesigned anyway.
A middle path many businesses take: wrap the legacy system with a small, modern service that exposes the handful of operations you need (create order, check stock, fetch invoice) and hide the messy integration behind it. Your AI agents and other tools then talk to a clean API, and when you eventually replace the legacy system, only the wrapper changes. It's the same pattern we describe in how AI agents connect to your CRM, ERP and inbox.
Frequently Asked Questions
Can AI work with a desktop application, not just websites?
Yes. RPA tools and computer-use agents can operate desktop applications on Windows, including older ones. The same reliability rules apply: prefer the most stable method, validate inputs, and verify results. Desktop automation usually needs a dedicated machine or virtual desktop where the bot runs.
Is it legal to automate logins to a third-party portal?
It depends on the portal's terms of use and the laws where you operate. Many business portals allow it or offer an official bulk option; some prohibit automated access. Check the terms, ask the provider if unclear, and never bypass security controls like CAPTCHAs.
How much does legacy system automation cost compared with API integration?
Usually more to build and more to maintain, because screens change and edge cases surface over time. Database or file-based integrations are closer in cost to an API integration. Budget for ongoing maintenance whenever the automation touches a user interface.
Will the automation break when the software updates?
Screen-based automation can, so plan for it. Ask the vendor for advance notice of updates, test after each one, and design failures to land in a queue rather than silently dropping work. Database and file integrations are less affected by interface changes.
Should we use RPA or an AI computer-use agent?
For high-volume, repetitive tasks on stable screens, scripted RPA with AI for the judgment steps is usually cheaper and more predictable. Computer-use agents fit low-volume tasks, screens that vary, or work where writing a script for every path isn't worth it.
Stuck with a system that won't talk to anything?
Most businesses have at least one system everyone copies data into by hand. Tell us which one it is and what the copying looks like, and we'll tell you the most stable way in, or whether it's time to replace it. Our agentic AI development work includes exactly these integrations.