An AI assistant for accounting is a layer between your questions and your ERP that turns natural language into actions on your ledger, receivables or payables. The 2026 question for a controller or CFO is no longer whether that layer exists; it is which form fits your scale: a chat-AI plus an MCP server you install yourself, a script you build in an evening with a coding agent, or a platform that takes the maintenance off your plate. This piece explains what each path actually does for you, when it breaks, and why the decision in practice is less about models than about who keeps it running.
What MCP actually is
MCP stands for Model Context Protocol. It is an agreement that lets an AI model talk directly to other systems. Practically, it means a chat-AI is no longer dependent on what you type into a prompt and no longer dependent on individual plugins per tool. An MCP server between the chat-AI and your ERP lets the AI read your ledger, answer questions about open invoices, or propose an action. What it solves is the copy-pasting. What it does not solve is everything that comes after: validation, maintenance, audit, production reliability. By 2026 nearly every major AI vendor publishes MCP-compatible approaches, so it is no longer a niche technique.
The problem is not building it, it is keeping it running
In customer conversations we see the same pattern. Someone with technical affinity installs an MCP server on their ERP within a few weeks, connects it to a chat-AI, and automates a first workflow. A weekly export, a quick ledger query, a first attempt at an automated month-end close check. It works on their laptop. Everyone is excited.
Two months later comes the first reminder of how fragile the setup is. The script stops working, the builder is on holiday, nobody remembers how it was put together. The documentation is a few loose Notion pages and some half-modernised Python in a Cursor tab. The question that remains once the builder is away is not about the AI; it is about the set of scripts, validation rules and integrations that ran on one laptop.
This is not a critique of DIY. It is an observation. What sits between "I built something" and "it runs every day for the team" is the twenty percent that is not about model choice or prompt. It is maintenance, validation, audit, scale, and colleagues who must be able to take it over.
What a chat-AI plus connector actually does for you today
Honestly and concretely: a chat-AI plus an MCP server is an excellent copilot for one user who knows what they want. Time savings within weeks on specific ad-hoc tasks. Ask a question, pull raw data, propose simple write actions with user-confirm, run a calculation that used to take thirty minutes in a couple of minutes.
What it does not do is autonomously take over your daily work. A typical finance workflow is not a linear script. It is a loop: look at a new supplier invoice, check whether the supplier is known, read the email context, remember that this supplier made a VAT mistake last month, decide whether this is a routine booking or whether it goes to the controller. The manual workflow is full of knowledge that sits in someone's head. AI in chat form speeds up the steps; it does not take over the loop.
There is also a boundary people underestimate. General language models are trained on broad material, not on current local tax reality. Tax law changes more often than a model is updated. A chat-AI that explains, convincingly enough, how a specific tax credit works, can just as easily give outdated information. A production system for finance must validate rules against local, current sources, not against a model's generic training.
The difference between a copilot and an AI colleague
A six-dimension table makes the difference visible.
| Dimension | Chat-AI plus connector (copilot) | AI colleague |
|---|---|---|
| Who initiates the work? | You, every time | The system, on trigger or schedule |
| Who knows the accounting rules? | General training | Domain-specific validation |
| Who reviews the output? | You, after the fact | Validated before action |
| Who maintains the integration when the API changes? | You | The platform |
| Who handles multi-division or multi-entity access? | Nobody | Built in |
| Who is liable on a mistake? | Unclear | Captured in audit trail |
On all six, a chat-AI plus connector scores at copilot level. An AI colleague scores senior on all six. Not because of better models, but because of specialisation, production maintenance, and organisation-aware access management.
Three scenarios where the difference shows
Scenario 1: Exact changes a field in its API. An unmaintained MCP server breaks or returns wrong data without anyone noticing immediately. The builder may be on holiday or have moved on to another project. In a production-grade integration the change is caught through versioned adapter management, with release notes and breakage monitoring. You do not notice anything.
Scenario 2: An AI suggestion leads to a double booking or a wrong VAT code. A chat-AI without a validation layer misses the semantic check (this invoice was already booked last week, or this VAT code does not match this type of supplier). A finance-specialised platform stops the action because domain rules fire before the booking lands.
Scenario 3: The auditor asks for an audit trail at year-end. Raw model call logs are not an audit trail. What the auditor needs is a structured action log with user, intent, approver, timestamp, and the exact ERP mutation, per action. A chat-AI plus connector does not deliver this out of the box.
"Can you just connect everything for me, like MCP does?"
A common question from customer conversations. The short answer: integrations are maintenance work. For the ten systems on which European mid-market finance runs you can justify the investment. For the two hundred niche tools beyond that, you are better off with a lighter path: connect MCP yourself, or let the AI colleague log in via a password manager for occasional use. Honesty here builds trust.
Six paths from light to heavy
Anyone searching for "AI assistant for accounting" or "MCP for accounting" runs into different solutions without an honest comparison. The six paths that matter in practice.
| Path | Setup time | Maintenance cost | Production fit | Audit trail | Who keeps it running |
|---|---|---|---|---|---|
| Chat-AI plus open MCP wrapper | Minutes to hours | High over time | Limited | Raw call logs | You |
| Chat-AI alone, with copy-paste | Immediate | None | Personal productivity | None | You |
| Build your own on the ERP API with a script or CLI | Two to four weeks | High and ongoing | Possible with discipline | What you build | You or your dev team |
| Have a coding agent (Cursor or similar) generate an integration | An evening to a week | High and fragmented | Limited without review | Unpredictable | You, and whoever comes after |
| Have a freelance dev or IT partner build a custom integration | Four to twelve weeks | Fixed maintenance contract | Good | What the builder delivers | Them |
| Agent platform with finance specialisation | Under 30 minutes single-division | Included | Good | Structured | The platform |
Two things stand out when you read this table with a finance lens. The first five paths share that you or someone in your organisation becomes the owner of a piece of production software. For the first two, that is you. For paths three and four, you sit on code that dev tooling knows but the finance team does not, and the ERP API has its own quirks (token refresh, division restrictions at endpoint level, batch limits, webhook stability) that you only learn when they break. For path five you buy a maintenance commitment from someone who is not in the office. The problem finance teams are solving here is rarely "I cannot build it". The problem is "I do not want to own something I do not understand".
When a lighter path is the right call
Honestly. For those not yet at platform scale.
Tech-savvy SaaS startups with a small finance team. One founder or operator who codes, one ERP, a simple bookkeeping model, low invoice volume per month. The DIY route gets you a result in an evening and you can run it for a few months. One-off analyses or pilots. If you want to know whether a type of analysis adds value before you invest in a process, build it ad hoc. Engineering teams that want to test. A dev team validating an idea before finance is brought in can use a coding agent to make a first integration and test assumptions.
In all these cases: use it, learn from it, and be honest about the moment you cross a threshold.
What a production-grade AI colleague categorically does differently
One paragraph, no pitch. A production-grade AI colleague is a platform category that validates finance output against accounting rules, maintains integrations through versioning, delivers audit trails at acceptance level, and handles access management aligned with how your organisation actually runs (divisions, projects, users; not one shared password). This category recently came of age for European mid-market finance. Ask every vendor how they fill in those four points.
What you can do this month
Three actions, no purchase required.
Test an MCP integration or a wrapper against your ERP. An evening's work, nothing you cannot uninstall. The goal is not to automate, the goal is to feel where the sharp edges of a DIY approach sit. What works? Where do you have to think? What would happen if your laptop disappeared tomorrow?
Make a list of finance tasks where copilot level is enough and where it is not. On the copilot side: ad-hoc questions, quick lookups, a calculation, a raw export. On the not-copilot side: anything on a schedule, anything that needs an audit trail, anything that needs validation by a second person, anything that crosses your liability threshold.
If that list has more than two not-copilot tasks, book a conversation or demo. A platform decision is not needed tomorrow, but the problem you are trying to solve is worth a look. The cheapest mistake is to keep building on your laptop for a few more months before admitting that keeping five solo flows running costs more than handing the maintenance to something that does that for you.
Frequently asked questions
What is MCP and do I need to understand it?
MCP (Model Context Protocol) is an agreement that lets an AI model talk directly to other systems instead of going through copy-paste or individual plugins. As a buyer you do not need to understand the protocol. You do need to know that whoever uses it themselves also maintains it themselves. Anyone buying a platform that uses MCP under the hood sees none of it.
Can I connect my accounting system to ChatGPT or Claude?
Yes, technically. With an MCP server or another connector you can wire a chat-AI to your ERP. It works for ad-hoc questions and simple actions with user-confirm. It works less well as soon as you need validation, audit trails, multi-user access, or automated schedules.
How safe is an MCP integration with my accounting system?
It depends on who maintains it. An open MCP server without a validation layer exposes your ledger to a chat-AI with general training. A production-grade platform keeps customer data inside your tenant, masks personal data before it reaches the language model, and logs every action at acceptance level.
What if my ERP changes its API?
In a DIY setup that is your problem. In a production-grade platform the adapter is versioned and updated before the change reaches you. In a DIY approach plan on half a day to a full day per quarter for maintenance on every integration.
Is MCP a temporary trend or the new standard?
By 2026 nearly every major AI vendor publishes MCP-compatible approaches. Standardisation is happening. But standardising the protocol does not change the fundamental question of who maintains it and who is liable for the outcome.
Closing
AI is neither a protocol nor a colleague. It is a lever that can multiply your productivity three or four times if you deploy it at the right scale, and that on average makes you more fragile if you let it run on a laptop for a team of five. The difference between those two outcomes is not which model you pick. It is who becomes the owner of keeping it running.
Start with chat-AI plus MCP if you are building solo. Switch to a platform when you notice the questions have outgrown your laptop. Between those two sits a learning trajectory you cannot buy and cannot skip.
See how Nance connects to your ERP, production-grade, or book a demo to see it run on your own stack. Further reading: AI in Exact Online: what you can actually automate today, AI for accountants: where the admin stops, your work starts, why the SMB finance stack is breaking, or the CFO solutions page.
