01 · The problem
A chatbot beside the accounting system is not enough.
Value emerges only when AI can understand the question, retrieve the right evidence, propose a concrete action and leave the decision to an accountable person.
Finance work contains many tasks that are simple individually but time-consuming together: finding a customer, checking a supplier, comparing vouchers, preparing an accounting entry or drafting a response. Generative AI can shorten the path, but the flexibility that makes it useful also makes it unsuitable for uncontrolled writes.
The solution must therefore treat AI as a qualified proposal engine, not an administrator with unrestricted access. It should show the evidence, describe what will change and wait for explicit approval.
The right level of ambition
Automate the path to the decision first. Automate the decision itself only when the rules are clear, risk is low and the outcome can be verified.
02 · Workflow
From an open question to a controlled change.
Steg 01Ask
The user describes the job in natural language and specifies a period, company or counterparty when needed.
Steg 02Retrieve
Rekly reads the relevant Fortnox records and can enrich them with current company data.
Steg 03Propose
The AI formulates answers or proposed actions with visible evidence and uncertainties.
Steg 04Review
The user sees exactly which fields are affected and can edit, reject or approve.
Steg 05Execute
Only after approval is the bounded change made in Fortnox.
Step 06Log
The question, evidence, proposal, decision and result are linked for follow-up.
Show the difference before writing
A good approval step shows more than a button. It shows a diff: which records will be created or changed, old and new values, the rule or evidence behind the proposal and what happens if the user continues.
↳
Continue reading
Unlock the rest of the guide.
Leave your email to unlock the complete guide immediately and receive the PDF in your inbox. If you choose, we can also follow up with concrete examples of AI workflows for finance and administration.
Your interest is tagged asAI for financeRekly · Hamn AI · BolagsAPI
03 · Building blocks
Combine system access, AI and Swedish context.
WorkflowRekly
Connects the user's AI to Fortnox, presents proposals and keeps writes behind an approval step.
rekly.se ↗
AI infrastructureHamn AI
OpenAI-compatible inference with local models for organisations that need control over data and operations.
hamnai.se ↗
Company contextBolagsAPI
Enriches customer and supplier records in the accounting system with current Swedish company information.
bolagsapi.se ↗
The products can be used separately. Together they illustrate an important architecture: the AI model should not carry permissions or business rules itself. A separate workflow layer retrieves data, limits allowed actions and requires approval where risk warrants it.
04 · Control model
Give every action a risk level.
The same control does not fit every task. A read-only question can be answered directly, while changing payment details should always require stronger verification and perhaps dual approval.
| Level | Example | Control | Result |
|---|
| 1 · Read | “Which invoices are due this week?” | Permission check and source reference | Answer without a write |
| 2 · Preparation | Create a draft customer or accounting entry | Visible evidence and manual review | Draft, not a posted change |
| 3 · Bounded write | Update an approved field | Diff, explicit approval and log | One defined API operation |
| 4 · Sensitive action | Payment, bank details or a large bulk change | Strong verification, approval and amount limit | Separate security flow |
Limit the tools, not just the prompt
Instructions in a prompt are not an authorisation model. Limit in code which operations the AI flow may call, which fields may change, how many records may be affected and which amounts require additional approval. Control then sits outside the model.
05 · Implementation
Choose a job that recurs every week.
- Map the task. Document the starting point, sources, decisions, writes and definition of done.
- Start read-only. Let the AI gather evidence and answer, but not change anything in the first phase.
- Add drafts. Measure how often proposals are accepted, corrected or rejected and why.
- Enable one bounded write. Choose a reversible operation with a clear diff and mandatory approval.
- Scale based on evidence. Add new operations only when quality, time saved and control can be measured.
06 · Measurement
Measure more than how often the AI answers.
- Acceptance rate: share of proposals approved without changes.
- Correction rate: which fields or reasoning users most often adjust.
- Time saved: minutes from question to completed task compared with the previous process.
- Prevented writes: when the control step catches an unsuitable or incorrect proposal.
- Source coverage: whether the answer used the correct period, company and evidence.
A rejected proposal can be a successful control step.
Do not judge the system only by its level of automation. A clear stop before an incorrect write is also delivered value.
07 · Checklist
A safe first AI feature in finance.
The first use case has a clear start and goal.
The AI may call only explicitly allowed tools.
Sources and period are shown with the proposal.
All writes show a clear diff.
Approval is required before Fortnox is changed.
Sensitive actions have additional approvals and limits.
Question, proposal, decision and result can be traced.
Quality and time saved are measured per task.
By Hugo!