Why AI Finance Tools Fail on Legacy Systems and No-API Bank Portals

Pull up your last close and list where the numbers came from. An ERP that your IT team has not upgraded in years. A bank portal that sends a statement only if someone logs in and clicks Download. A vendor site that exports one PDF at a time. Three spreadsheets that hold the real mapping.

Now read a demo of any AI finance tool. The demo connects to a modern cloud ERP through a clean API, pulls a tidy feed, and posts a journal entry. None of it resembles your close.

That gap is why AI finance tools fail on legacy systems. The tool works as shown, on the systems it was built for. Your controller's problem sits on the systems it was not.

The assumption underneath most AI finance tools

Most tools in this market start from one design decision: every system has an API, and the API is open. Build on that and you get fast demos, clean integrations and a short roadmap.

The finance stack you run breaks that decision in four places.

The bank portal with no API. Aggregators cover a large share of banks. Some banks, credit unions, insurers and payers fall outside them. For those accounts, a person logs in, downloads the statement and drops it in a folder. An AI tool that needs a feed has nothing to read.

The ERP behind a network boundary. On-prem and hosted legacy ERPs sit behind firewalls and tunnels. Some need an SDK installed on a machine inside the network. QuickBooks Desktop is a common example. A cloud-only tool cannot reach the ledger, so it cannot read the balance it is supposed to reconcile against.

The vendor system that sells you the API. Some systems hold your data and charge extra to let you query it. Others publish an API that omits the fields your close needs. The tool inherits the limit.

The spreadsheet in the middle. The mapping from vendor name to GL account lives in a workbook one senior person maintains. No integration reads it, so the tool guesses, or skips it.

What failure looks like in practice

Failure arrives as scope.

A vendor scopes the pilot to the accounts it can connect. Those are the easy ones. The hard accounts stay manual, so your team still downloads the same statements and still chases the same exceptions. The tool automates the part of the close that was already fast.

Three months later the controller can show a reconciliation for ten accounts and still has forty-odd that need hands. The vendor calls it phase two. Phase two needs a connector the vendor has not built.

A second pattern: the tool reads from an export instead of the source. Someone uploads a CSV each month. That works, and it also reintroduces the manual step you wanted gone, plus a copy of the data that can drift from the ledger.

Why an API-first design cannot stretch to cover this

You can add connectors one by one, and vendors do. Each connector takes engineering time, and each depends on the system's owner keeping the API stable. The long tail of bank portals, regional ERPs and industry-specific systems never reaches the top of a roadmap, because each one serves too few customers.

A tool built on the assumption of open APIs treats a missing API as an exception. A finance team with legacy systems has that exception in half its accounts. For that team, the exception is the workload.

What works on a stack that does not fit the demo

A tool that handles legacy systems needs three capabilities that an API-first tool does not.

It operates the interface a person uses. Where no API exists, the agent logs in to the portal with credentials held in a vault, the same login your staff use today, and downloads the statement the way a person would. Adopt agents work with portals and legacy interfaces this way, and they work with the interface, not around it.

It turns the interface into a repeatable tool. A one-off script breaks the first time a bank redesigns its login page. Adopt's Zapi records a browser session and converts the calls between the browser and the backend into tools an agent can call. Where a system offers OAuth or a public MCP server, Adopt uses that first. Zapi covers the systems that do not.

It sits on top of the ERP you have. Your ERP stays the system of record. The agent reads from it, prepares the work, and a person approves before anything posts. That requires no migration and no process redesign, which is the right answer for a team that cannot replace a working ERP to get better automation.

Honest limits belong in this list. QuickBooks Desktop needs a network tunnel and an SDK. Some systems need custom work before an agent can operate them. A vendor who says every connector takes a day is describing a demo.

Five questions to ask before you sign

Put these to any vendor, Adopt included.

1. Which of my accounts and systems can you connect today, and which need new work? Ask for the list by name. A vague answer means the hard accounts are not in scope.
2. What do you do when a system has no API? "We build a connector" is an answer with a timeline. Ask for the timeline.
3. Where does the data go? Ask whether agents run in the vendor's cloud, your cloud or on-prem, and whether a hosted model sees your data.
4. Who approves what posts? Look for a named reviewer on every entry and no write to a control account without sign-off.
5. Can I run it on one of my ugly accounts first? The proof that matters is on your worst reconciliation, not on a sample file from the vendor.

If a vendor can answer the first two with specifics, the product reaches your stack. If the answers come back as roadmap items, plan for a pilot that covers a fraction of your close.

What to do this month

Start with an inventory. List every system and account your close touches, mark which ones expose an API, and mark which ones a person reaches by logging in. That list is your evaluation scorecard.

Then take the hardest reconciliation on it and ask each vendor to run that one. Our guide to evaluating AI finance tools lays out the full set of criteria, and the AI for finance teams guide covers where agents fit across close, AP and AR. For the reconciliation work itself, see account reconciliation.

If your close runs through a bank portal, an on-prem ERP or a spreadsheet, send us the messiest reconciliation you have and we will show you what an agent does with it.