How to automate legacy software that has no API

Most of the software a business actually runs on was never built to be automated: an accounting program from 2009, a supplier portal with no export button, a government cabinet, an ERP whose vendor charges for every integration. You can still automate legacy software without an API in four ways: through its database, through the files it already imports and exports, with a recorded RPA bot, or with a computer-use AI agent that operates the screens like a person. Try them in that order, and give anything that clicks through a screen a human approval step before it changes money or records. With an AI-first team the first automated process takes days, not months, and costs 4 times less than traditional hand-written development, 4 times faster.

What changed in 2026 is the fourth route. Computer-use agents from Anthropic, OpenAI and Google can now read a screen and act on it well enough for real work, but not well enough to be left alone. Anthropic's own system card for its Fable 5.1 model reports 77.9% on the OSWorld 2.0 desktop benchmark with partial credit, and 41.7% when only fully completed tasks count. That gap is the whole story: the agent gets most of the way through most tasks, and finishes fewer than half without help.

Four ways to automate legacy software without an API

1. The database. Most old programs keep their data in an ordinary SQL database on a server you own. Reading from it is safe and fast; writing to it needs care, because the program may keep business rules in its own code. When the vendor allows database access, this route is often 30 to 50% cheaper than a bot over three years, according to a 2026 cost comparison by Kolonell.
2. Files and email. If the program imports a CSV, exports a report, or sends a registry by email, automation can live entirely outside it. In Sewing Lab the postal cash-on-delivery registries arrive by email; the system pulls them from the mailbox and reconciles every shipment, with nobody opening a file.
3. RPA. Tools like UiPath or Power Automate Desktop record a sequence of clicks and replay it. Fast to set up for one stable screen flow, but brittle: Kolonell puts RPA failure rates at 5 to 15% of runs and expects several days of fixes a year on an application that changes.
4. Computer-use AI agents. The agent sees the screen, decides what to click, and checks the result. A moved button or a new pop-up does not break it the way it breaks a recorded bot. It is slower than code, and every step is a paid model call, so it fits low and medium volume.

The honest default is a hybrid: give the agent direct access everywhere a database or file route exists, and let it use the screen only for the steps that have no other way in. Once a legacy system has a clean route, wrapping it as an MCP server lets any AI assistant use it as a tool.

When a computer-use agent is the right choice

Use the screen route when the answer to all three questions is yes:

• No other route exists. The vendor will not open the database, there is no import or export, and the program is a desktop app or a web portal you do not control.
• Volume is modest. Tens or hundreds of runs a day, not tens of thousands. At high volume a proper integration pays for itself.
• A person can approve the step that matters. Submitting a payment, changing a price or sending a document waits for a click from someone accountable.

If the system will be replaced within a year or so, screen automation is also the cheaper bridge: there is no point building a deep integration into software that is about to be switched off.

How to keep an agent in old software safe

An agent with a mouse has the same permissions as the account it runs under, so the safety rules are about infrastructure, not prompts:

• A separate account with minimal rights, in an isolated virtual machine. Telling the agent "never delete anything" is not the same as an account that cannot.
• Credentials kept out of the model. The agent works in a session that is already signed in; the password never passes through the prompt.
• Screen content treated as untrusted. Text inside a portal can contain instructions; the agent follows your task, not what a page tells it.
• Step, time and cost limits per run, with a clear failure message to a person instead of endless retries.
• Check the outcome, not the agent's report. After the run, read the record back from the system and confirm it changed.

This is how Zapyt works in production: its collectors operate Threads and X through the client's own account in a headless browser, clients connect their accounts through a streamed browser view so the password never reaches the operator, and no message goes out until a person approves it.

What it costs to automate legacy software in 2026

The market reference is Kolonell's 2026 analysis of per-process automation in Europe: EUR 4,000 to 12,000 and 2 to 6 weeks for an RPA process, EUR 8,000 to 30,000 and 6 to 12 weeks for a custom integration. We use the same figures in dollars, which rounds them down.

We are 4 times cheaper and 4 times faster because AI writes the code and senior engineers direct it: fewer hours for the same working system, not thinner work. Running costs are separate. A screen agent's model usage typically runs $30 to $800 a month depending on volume, against RPA licences that often cost $5,000 to $15,000 per bot per year. If you are still choosing between a ready platform and your own build, see automation platform vs custom build.

How we approach legacy automation at DForce

We start by looking for the boring route. In most audits there is a database, an export or a mailbox that removes half the screen work before an agent is needed. Then we automate one process end to end, measure how often it finishes without help, and only after that widen the scope. Of the processes that still need the screen, the ones that repeat most get promoted to direct integrations over time, which lowers both cost and failure rate.

If there is a program in your company that everyone complains about and nobody can connect to anything, book a discovery call and we will tell you which of the four routes fits, and what it would take.

Frequently asked questions

Can you automate software that has no API?

Yes. There are four routes: read and write the program's database directly, exchange files it already imports and exports, record the clicks with an RPA tool, or let a computer-use AI agent operate the screens the way a person does. Check them in that order: the first two are the cheapest to run and the most reliable, and the screen-driven routes are for what is left.

What is a computer-use AI agent?

A computer-use agent is an AI model that looks at screenshots of an application and returns mouse clicks and keystrokes, so it can work in desktop programs and web portals that offer no API. Unlike classic RPA it does not depend on fixed screen coordinates or element selectors, so a moved button does not break it. It is slower and less predictable than code, so it needs step limits, logging and a person approving anything that moves money or changes records.

Is RPA still worth it in 2026?

For a stable, high-volume screen flow that will not change, a recorded RPA bot is still cheap to run. Its weak point is maintenance: bots break when the interface changes, and licences often cost $5,000 to $15,000 per bot per year. For lower-volume work, or interfaces that change, a computer-use agent or a direct database route is usually the better buy.

How much does it cost to automate a legacy system without an API?

With an AI-first team, automating one process through the screen with a computer-use agent and an approval step costs $1,000 to $3,000 and takes 3.5 days to 1.5 weeks, against $4,000 to $12,000 and 2 to 6 weeks with traditional development. A proper database or file integration for one process is $2,000 to $7,500 over 1.5 to 3 weeks, against $8,000 to $30,000 over 6 to 12 weeks. That is 4 times cheaper and 4 times faster, because AI writes the code and senior engineers direct it.