Field Notes · Integrations
By Jigesh Shah · 6 min read · Updated 2026-07-20

When Your Software Has No API, Put an AI Employee on the Screen

A lot of the software small businesses run on has no API to plug into. Computer-use AI can operate it by clicking and typing like a person. That’s powerful — and easy to misuse. Here’s the honest version.

The short answer

When a tool has no API, a computer-use agent can still operate it — reading the screen and driving the mouse and keyboard the way a person does. That makes it the right tool for the legacy, regional, and portal software SMBs are stuck on and can’t otherwise automate. But screen-driving is slower, more brittle, and higher-risk than a real integration: a redesigned button can break it, and it acts with your full logged-in access. So treat it as a bridge for the last mile — governed with human-in-the-loop approval — not as a default, and graduate to a real API once a task gets frequent or high-stakes.

A surprising amount of the software that runs a small business was never built to be automated. The regional accounting package, the bank and government portals, the fifteen-year-old ERP nobody wants to replace — they have a screen for a human and nothing else. No API, no webhook, no export worth the name. So a person sits there and clicks. That gap is where the most interesting — and most misunderstood — kind of AI automation lives.

The integration gap most SMBs actually live in

Most automation advice quietly assumes an API exists. Wire system A to system B, let the data flow. But the tools that eat the most human hours in a small business are often exactly the ones with no way in. The work gets done by a person reading a screen and typing into another one — which is slow, easy to get wrong at 5pm, and impossible to hand to a normal integration.

What “putting an AI on the screen” means

Computer-use models — the screen-operator approach — change the shape of that problem. Instead of calling an API, the agent looks at the actual screen, finds the field, and clicks and types like a person would. No cooperation from the software is required, because from the software’s point of view it is a person. For a whole class of no-API systems, that’s the difference between “can’t be automated” and “can.”

Where it genuinely works

  • No-API portals and legacy tools — the government filing site, the insurer’s claims portal, the accounting app with no integration. If a human can do it in a browser, a screen agent usually can too.
  • Repetitive, well-defined tasks on stable screens — the same five steps, the same layout, many times over.
  • Reaching a system into a larger workflow — using the screen as the last hop so an otherwise-blocked agent can finish a job end to end.

Where you should build the API instead

Screen-driving is a bridge, not a foundation. The honest comparison:

DimensionScreen operator (computer-use)Real API / integration
SetupFast — point it at the screen, nothing needed from the vendorSlower — an API has to exist and be wired up
SpeedHuman pace — one screen at a timeFast — direct data, batched
ReliabilityBrittle — a redesigned screen can break itStable — contracts change rarely
Best forNo-API portals, legacy tools, low-volume tasksHigh volume, high stakes, anything mission-critical

The rule of thumb: use the screen to reach what you otherwise couldn’t, but the moment a task becomes high-volume, time-sensitive, or involves moving money, a real integration pays for itself in speed and reliability. Screen-driving buys you access today; an API buys you dependability tomorrow.

The part people underestimate: oversight

A screen agent acts with your full logged-in privileges. It is signed in as you, clicking real buttons in real systems — which means a misread screen isn’t a bad API response you can catch and retry; it can be a payment sent or a record overwritten. This is precisely where a human-in-the-loop approval step stops being optional. Log every step (a screenshot per action helps), start on read-mostly and low-risk tasks before trusting it with anything that writes, and gate consequential actions behind a person. The capability is impressive; the discipline around it is what makes it safe.

How we think about it at Neural Infrastructure

We treat screen-driving as one tool in the kit, not a silver bullet — part of meeting a business where its software actually is. Use it to reach the systems that were never built to be automated, keep it governed with memory and human oversight, and graduate a task to a real integration when its volume or stakes justify the work. It fits the same principle as everything else we build: “isolated instances, coordinated agents” — capable where it needs to be, contained by design. Put honestly: an AI on the screen is how you cross the last mile to a system you can’t integrate. It is not an excuse to never build the road.

Frequently asked questions

Can AI use software that has no API?

Yes. Computer-use (screen-operator) agents read the screen and drive the mouse and keyboard the way a person does, so they can operate GUI-only tools — regional accounting software, bank and government portals, legacy ERPs — that offer no API to integrate with.

Is a screen-operator agent the same as RPA?

They rhyme, but classic RPA follows brittle recorded scripts tied to exact screen positions and breaks when anything moves. A computer-use agent interprets the screen more like a person, so it tolerates small changes better — but it is still slower and less reliable than a real integration, and both need supervision.

When should I build an API instead of using a screen agent?

When volume is high, the stakes are high (especially moving money), the task is time-sensitive, or the underlying screen changes often. Screen-driving is a bridge to systems you cannot otherwise reach; once a task is frequent or critical, a real integration pays for itself in speed and reliability.

Is it safe to let an AI drive my software?

Only with oversight. A screen agent acts with your full logged-in access, so a misread screen can do real damage. Keep a human-in-the-loop approval step before any consequential action, log every step, and start on low-risk, read-mostly tasks before trusting it with writes.

JS
Written by Jigesh Shah
Founder & CEO, Neural Infrastructure

Jigesh Shah is the founder and CEO of Neural Infrastructure, the operating layer for autonomous AI. He also runs RYVR, a marketing agency operated end-to-end by AI employees — the flagship proof that autonomous AI can run a real business, not just demo one. His work focuses on making AI agents production-grade for the companies that actually deploy them: governed, observable, and owned.

AI employees for the tools you actually run.

Neural Infrastructure builds autonomous AI around your real operations — governed, with memory and a human-in-the-loop approval path — so automation can reach even the systems that were never built for it.

See the platform