Dext AI Connector vs Custom MCP: What Firms Should Ask

By Jude Lee · · News

Two accountants reviewing software settings on a laptop in a small firm office

Last verified: November 2026. Product announcements in this space date quickly — check each linked article’s own date stamp and the vendor’s current documentation before acting on anything below.

What shipped, and why it matters

Two announcements landed within days of each other. CPA Practice Advisor reported on 22 October 2026 that Dext rolled out an AI Connector, and The Paypers described it as linking practice data to AI assistants. Separately, AiThority reported that Financial Cents launched an MCP server so firm data can be used inside everyday AI assistants. The Paypers and AiThority items carry the same late-October 2026 window; confirm the date shown on each page, because if you’re reading this well after publication the “just shipped” framing no longer applies and the feature sets will have moved.

Treat the headlines as the starting point, not the spec. Feature details, pricing, and permission models change fast — confirm everything against the vendor’s own documentation before you plan around it.

The pattern is what counts: the tools you already pay for are starting to ship the connective tissue that lets an AI assistant read your firm’s data directly. Not long ago, getting Claude or ChatGPT to see your practice-management queue meant exports, middleware, or a custom build. Increasingly, it means a toggle.

What an AI connector actually is, in plain terms

Most of these connectors are built on MCP — the Model Context Protocol, an open standard for giving an AI assistant governed access to a specific set of data and actions. An MCP server is a small adapter that sits in front of a system (Dext, a practice-management tool, your general ledger) and publishes a defined menu: “list open client queries,” “fetch this receipt’s extracted fields,” “show jobs due this week.” The assistant can only call what’s on that menu. It is not a firehose into your database, and it is not the same thing as uploading client files into a chat window.

That distinction matters because an MCP connection is governed by scopes you set, while a copy-paste into a consumer chatbot is governed by nothing. If you want the underlying mechanics, we walked through them in connecting an AI assistant to QuickBooks or Xero via MCP.

Where a vendor connector earns its place in a firm workflow

Here is a concrete use that does not require a single line of code. A bookkeeping supervisor opens an assistant on Monday morning and asks: “For my ten CAS clients, which receipts are still awaiting coding, which jobs are past due, and draft me a short client-by-client status note.” With a receipt-capture connector and a practice-management connector both attached, the assistant reads both, cross-references them, and drafts ten notes in one pass. The supervisor edits and sends.

Nothing was posted. Nothing was emailed automatically. The agent read, reasoned across two systems, and produced a draft — which is where the current generation of this technology is genuinely strong. Compare that to the same supervisor exporting two reports into Excel and writing ten notes by hand.

The weaknesses are equally concrete. The assistant will summarize whatever the data says, including when the data is wrong. If half your clients’ receipt queues are stale because nobody chased them, a confident AI status note makes the mess look organized. SmartCompany’s piece on AI finance mistakes put it well: AI takes a mess and amplifies it.

A connector doesn’t give you clean data. It gives your mess a confident narrator.

Vendor connector, custom MCP server, or neither

Vendor-supplied AI connector
Fastest path: turn it on, authenticate, done. Maintained by the vendor through their own API changes. Scopes and available actions are whatever the vendor decided to publish — if the action you need isn’t on the menu, you wait. You also inherit their logging granularity and data-retention defaults as-is, with no ability to tune either to your quality-management policy. And consider exit cost: if the connector is sunsetted, re-priced, or folded into a higher tier, any process you built on top of it reverts to manual with little notice. Best when your workflow lives mostly inside that one system.
Custom MCP server you commission
You define the menu: exactly which fields, which clients, which actions, which audit trail. Can span systems the vendors will never connect for you — a firm-specific job taxonomy, a legacy workpaper folder, a pricing table. Costs real money to build and keep alive as upstream APIs shift. Best when the valuable workflow crosses three systems and none of them own it.

And “neither” is a legitimate answer. If your bottleneck is that clients don’t send documents, a deterministic reminder sequence beats an agent.

There’s a second case worth naming, because clean data doesn’t rescue it. Where the assistant’s output has to be tied back to a source document anyway — a fixed-asset addition memo, a loan covenant calculation, an accrual supported by a specific invoice — the reviewer opens the underlying document regardless. The drafting time you saved gets spent re-reading the source to confirm the draft. If a task’s review step requires inspecting the evidence line by line, a connector mostly relocates the work rather than removing it. We ranked which firm tasks actually reward AI versus plain rules in accounting automation examples ranked by AI readiness, and the broader vendor-versus-custom tradeoff is covered in practice management MCP: vendor vs custom vs neither.

The questions to ask before you switch one on

  1. Read-only or write?

    Does the connector only read, or can the assistant change records? Start read-only. The failure modes of a write-capable agent are different in kind, not degree — see read-only vs write MCP for accounting agents.

  2. What are the permission scopes?

    Can you grant access to one client’s data, or is it all-or-nothing across the firm? Least privilege is the whole point.

  3. Per-user or per-firm identity?

    Does the assistant act as the logged-in staff member, inheriting their existing permissions, or as a single service account that sees everything?

  4. What gets audit-logged?

    You want a record of every call: who, which client, which records, when. If the vendor can’t show you the log, you can’t evidence it in a quality-management review.

  5. Where does the data go, and is it used for training?

    Get the retention and model-training terms in writing. This is the line item most likely to decide whether your confidentiality obligations allow it.

  6. Which assistants are supported?

    Claude, ChatGPT, Copilot, a desktop app, a browser extension? Support varies and changes.

  7. What happens at rate limits or outage?

    An assistant that silently returns partial data is worse than one that errors loudly.

  8. Can you disable it per client?

    Some clients will say no. You need a switch at client level, not just firm level.

What automation in accounting means now

Automation used to mean deterministic rules: bank feed rules, recurring journals, scheduled reports — same input, same output, every time. Agentic automation adds a second layer: software that reads unstructured context, decides what to do next, and takes multi-step action. The useful mental model is that rules handle the bulk of predictable, repeating work, and agents handle the judgment-flavored remainder in draft form, with a human approving. Neither replaces the other. We argued the split in rules, AI agents, or neither.

Whether this is a threat to the profession

The honest read: connectors like these compress preparation and status-reporting work, not judgment, relationship, or sign-off work. Inc.’s profile of accounting platform Pennylane frames automation as expanding what accountants can take on rather than eliminating them — a vendor-adjacent view, so weigh it accordingly. My own opinion: the practical risk isn’t that AI replaces CPAs; it’s that firms that never reallocate recovered hours just absorb the savings into lower realization.

As for what large firms run: based on what the Big 4 publicly describe in their own technology announcements, the pattern appears to be heavily customized internal platforms layered on enterprise audit and tax suites. Read that as inference from public statements, not measured fact. Either way, the advantage there is engineering capacity, not a secret tool — and the gap narrows when open protocols like MCP let a ten-person firm wire the same assistant to its own systems.

Sizing the win without making numbers up

There is no credible industry benchmark for what a connector saves, so model it yourself before you buy:

hours/week × weeks × loaded rate
Recovered prep time — use your own time-tracking data
recovered hours × realization × billing rate
Only counts if hours get reallocated to billable work
connector cost + setup + review time
The honest denominator

Pull two weeks of actual time entries for the task you’d automate. Plug in your own rates. If the result isn’t obviously positive before you’ve optimized anything, the connector is probably solving a problem you don’t have.

A sane 30-day test

Pick one workflow, one connector, read-only, three clients who have consented. Make consent concrete and recorded: a clause in the engagement letter, or a documented email confirmation that names the assistant, the system it will read, the categories of data it can access, and the fact that output is reviewed by a human before use. Run it in parallel with the manual process for two weeks and compare outputs line by line. Keep a log of every place the assistant was wrong — that log is your real evaluation, and it’s also the start of the review gate you’ll need if you ever let an agent write.

Not sure where to start?

Get a free automation audit: we map your bookkeeping, month-end close, client onboarding, document collection, and AP/AR — and show you what's worth automating before you spend a dollar.

Get a free automation audit

Related guides