A production-grade AI integration on Exact Online is a layer that validates bookkeeping actions against your ledger context, logs every action at acceptance level, and maintains the integration when Exact changes its API. The difference with a chat-AI plus an open connector is not in what you see working on your screen, but in what keeps working in production. What a controller can do today on Exact Online with AI is more than last year and less than vendors promise. This page shows which three workflows actually save time or money in production, how Nance goes beyond a chat-AI plus connector, and when a lighter path is actually the right call for your situation.
What it is
An AI assistant on Exact Online means you ask in natural language and Exact does or returns something. Three sentences controllers literally say to Nance:
- "Book these twelve supplier invoices according to the existing dimensions and wait for my approval."
- "What is my cash position over 30 days based on actuals plus open invoices?"
- "Explain the variance between budget and actuals for Q1 per cost centre and point out where the deviation comes from."
The moment it clicks for a controller almost never happens on the first sentence. It clicks when they take one step further: "can we run this check every month on the third and email me when the month can be closed?". At that moment the tool turns from assistant into colleague. One command runs every month on its own, checks whether bank lines are processed, whether invoices are missing, whether VAT has been filed, whether payroll is paid, and sends an email with the status. The controller no longer has to think about it.
The three workflows you put into production today
1. Month-end close checklist with automatic ticks
In practice the first workflow controllers adopt. AI walks through the close checklist every month: bank lines processed, supplier invoices booked, revenue reconciled, VAT ready for filing, intercompany settled, cost centres assigned. Per item a tick or an open point with the specific line that is still missing. The controller gets a summary on the third of the month saying "you can close" or "these eight bookings are still missing, here are the links". One-time setup, every month on its own.
2. Invoice-to-booking with human-in-the-loop
AI reads the incoming invoice, validates against Exact's accounting rules (VAT code consistent with supplier and country, cost centre required for this ledger account, division restriction respected), proposes a booking, and waits for approval. Posting only happens after the controller's click, or after a four-eyes step if that is agreed. For controllers processing hundreds to thousands of invoices a month, this saves on average five to ten minutes per invoice. The win is not in the booking itself but in eliminating wait time for data: AI collects context that the human used to chase down.
3. Ask-your-data: natural language queries on ledger and receivables
"Which debtors have been paying later than their baseline for three months?" "What was my margin per service line in Q1 compared to Q4?" "Which cost centre deviated most from budget this month?". AI parses the question, queries Exact, and gives an answer with links to the source lines. No report, no export, no BI ticket. For those losing thirty minutes every Monday to questions where the answer was already in the data, this saves the thirty minutes.
What Nance does that a chat-AI plus connector does not
Technically it is not hard to connect a chat-AI to Exact Online. A developer with OAuth experience and two weeks of time gets a long way. That is just not where the cost sits. The cost sits in the steps you only notice when they are missing.
Actions are validated against the bookkeeping, not just against syntax. A chat-AI saying "book this invoice as business meals" can pick a ledger account that is syntactically correct but substantively wrong. With Nance, every action runs through a layer that checks whether the VAT code matches the supplier and country, whether the cost centre is required for this account, whether the user's division access actually allows that booking. An invalid action is refused, not executed and undone after the fact.
Every action produces a structured audit log. Not the raw model call log that auditors get no value from, but a recorded line with user, timestamp, intent, approver and the exact mutation in Exact. An auditor can reconstruct in five minutes what happened and who was responsible. That is what an audit trail means in finance.
Access management follows your organisational structure. Divisions, projects, roles, four-eyes rules per action type. One AI integration that opens everything for everyone is not an integration, it is a risk. Nance respects the user's Exact permissions and layers action-level permissions on top (who can post, who can approve, who can only ask questions).
The integration is maintained when Exact changes its API. A non-maintained integration can break silently: a field that no longer returns data, an endpoint that gets stricter, a refresh-token behaviour that deviates from the spec. Waking up to an integration that has broken silently is expensive. Nance versions the adapter, monitors for breakage points, and adjusts before your month-end close falls over.
It is the sum that counts. Sam summarises it like this: building an AI workflow is eighty percent easy. The last twenty percent are the production pieces, and that is where most of the time goes. Nance takes that twenty percent off your hands so the controller can focus on the creative eighty: which workflow do I actually want?
There is one more point that is often missing from comparisons: use AI to be creative and figure out the workflow, but make sure the workflow itself runs as deterministically as possible once it stands. If you let an LLM do every step of your process, you can never be certain it happens exactly the same this month as last month. What you want is a workflow that runs as code, breaks when something unexpected happens instead of silently continuing, and brings AI back in only when an actual decision is needed. For auditors this is the difference between a tool they accept and a tool they reject.
What this means for the controller in practice
The choice is no longer "AI or no AI"; the choice is which layer between you and Exact fits your scale and your liability. Four categories, honestly compared.
| Dimension | Chat-AI plus open connector | Generic workflow tool (iPaaS) | Build your own on Exact's API | Nance |
|---|---|---|---|---|
| Setup time | Minutes to hours | Days to a week | Two to eight weeks | Under 30 minutes single-division, under two hours multi-division |
| Time-to-first-value | Immediate, for one query | A week for one flow | Two weeks for one flow | Immediate after connection |
| Audit trail | Raw call logs | Step logs without finance context | What you build yourself | Structured actions with user, intent, approver |
| Multi-division support | None | Limited | What you build yourself | Out of the box |
| Maintenance when API changes | Your problem | Your problem or vendor | Your problem | Caught by Nance |
| Suitable for controller liability | No | Limited | Depends on the builder | Yes |
| Suitable for IT liability | No, breaks at laptop level | Limited, low-code lock-in | High initial and ongoing cost | Yes |
Reading this table as an IT manager, you immediately see where the risk sits. A chat-AI plus connector works beautifully on one laptop. The question is who maintains it when the management team has three of them in flight, or when the building controller leaves. That is exactly the challenge low-code and no-code tools historically had too: building was fast, scaling was not. With AI the pattern is the same, only with more risk because the output itself is semi-deterministic.
Safety and data boundaries
Four claims at the assurance level, no implementation detail.
- Enterprise-grade data protection: no training on customer data.
- EU hosting available for customers who require it.
- ISO 27001 certification in progress.
- Customer data from Exact does not leave your tenant without an explicit masking step. By default, personal data, banking details and specific transaction details are not included in queries to the language model.
Setup
Five steps, five minutes of reading, between fifteen and thirty minutes of elapsed time for a single-division Exact environment.
- Connect your Exact environment. OAuth flow, division selection, scope choice (read, write, or both with four-eyes).
- Set roles. Who can post without approval, who gets proposals to review, who has read-only access.
- Define your first workflow. Month-end close checklist is the most chosen start. Other options: invoice validation or a weekly cash update.
- Test on one month or one batch. Compare what Nance proposes with what you would do manually. Adjust thresholds and approval paths based on what you see.
- Activate automation. One-time say "do this every month on the third". Email notification on. Done.
Multi-division you set up here by running the OAuth flow per division and extending the access matrix. Plan two hours for the full set up to ten divisions, longer if you want to configure intercompany rules.
Frequently asked questions
Does this work for multi-division and multi-entity Exact?
Yes. The access layer respects division restrictions and intercompany boundaries configured in Exact. One Nance user can have rights in one or multiple divisions, with separate approval rules per division. Multi-entity with consolidated reporting works by connecting per entity and running aggregations at the Nance level.
What happens if I want to reverse an AI action?
Every action is reversible via Exact's standard journal-entry mechanism. The Nance audit log links the original action to the reversing entry, so the timeline in Exact and in Nance stays aligned. For actions still pending approval, cancelling is one click; no Exact mutation has occurred.
How is this different from Exact's own native AI?
Exact's own AI features work inside Exact and on Exact data. Nance works across your entire finance stack: between Exact and your email, your CRM, your HR system, your payment processor. An ERP is the connective tissue between finance systems, and an AI colleague that lives in only one system misses half of reality. For controllers who work only in Exact, native AI may suffice; for controllers who move between five systems all day, it does not.
What if Exact changes its API?
It happens every few months. Nance versions the adapter and publishes release notes per Exact update. In practice customers do not notice an API change; the adapter is updated by us before the change reaches customers.
How does this handle VAT, divisions and cost centres?
VAT codes are validated against supplier and country before a booking is proposed. Divisions and cost centres follow the user's Exact permissions structure. Required cost centres are refused if missing; defaults can be set at user or supplier level.
What does my auditor say about the audit trail?
We have had the log reviewed by external auditors. The feedback determined the contents of the log: user, intent, approver, timestamp, exact Exact mutation with line ID. What an auditor wants to see is which human took which decision, not which prompt a model received. The log meets that bar.
When Nance is not the right call
Honesty belongs here too. For those who just started, sell one product through a simple billing model, work in one country, and process no more than a few dozen invoices a month, Nance is overkill. A chat-AI plus open connector, or even Excel with a fixed monthly routine, gets the work done. For those past the complexity threshold (multi-division, multi-entity, intercompany, four-eyes rules, an auditor who comes by every quarter for controls, an IT team responsible for what runs in production), Nance becomes the cheaper choice from month two because maintenance and validation are baked in. For concrete costs and packages: see the pricing.
Closing thought
What controllers tell us after the first month is strikingly stable: Nance does not feel like software, it feels like a colleague you give a task to. Not a SaaS tool that works the way the developers decided, but a teammate where you decide the work split yourself. One command, from now on every month, email me when it is done.
Book a Nance demo on Exact Online or connect your Exact environment in a test setup to see the three workflows running on your own data.
Further reading: AI for your books: when chat AI is enough, and when you need an AI colleague (the strategic context behind this integration), how Nance keeps debtors in check without spamming customers (the AR side running on this integration), the broader CFO solutions page for the full Nance offering, or why the SMB finance stack is breaking for the architecture context.
