Alice Bonneau.
Back to portfolioPortfolio

Case 06 · Inqom · Product design · UI/UX · Accounting

Accountants didn't need more alerts. They needed the reasoning behind each bank line.

Inqom already flags anomalies and accountants close them without reading. I redesigned bank reconciliation around three ideas. When a bank line won't match, its receipt, its history and the balances now appear right next to it. What was checked is kept. The AI suggests an account and explains why, but never decides.

Role
Sole designer · research → prototype
Product
Inqom Expert · accounting firms
Scope
Bank reconciliation, AI suggestion
Timeline
2 days · business case, 2026

My role

A design brief, answered end to end.

This is a business case for Inqom, done in two days and presented to the product team. I didn't design the accounting rules or the existing product. I designed how an accountant understands, decides and hands over a bank line that won't match.

What I was given

The brief, user verbatims and product data: alert usage, Excel exports, help center visits. Plus the existing reconciliation screen.

What I owned

The synthesis, the scope, the journey, every screen in Figma, the state matrix, the MVP split and the test protocol.

Who I worked with

I worked on this case alone. Every design decision on this page is mine, from the first sketch to the last screen.

Context

Two people, one bank line, the same lost time.

Inqom is accounting software used by firms to keep their clients' books. Reconciliation means matching every bank line to an accounting entry. Most lines go through on their own. The ones that don't are where the time goes.

When a line won't match, the screen only says "not reconciled". The document is in one place, the history in another, the balances in a third. Doubts get settled by phone, email, or Excel. None of it is kept.

The current Inqom reconciliation screen: a global banner of 14 anomaly alerts and bare 'not reconciled' statuses.

Before: 14 alerts on the file, one "close all" button, and nothing that says what is blocking each line.

Two profiles share that pain, and the design had to serve both: the one who produces the entry, and the one who reviews it.

Portrait of Julien, the accountant persona

Julien · accountant, 6 years, ~50 files

"Show me why, then let me decide."

Loses time finding the document and the history. Exports to Excel to see everything flat. Starts from zero when the client answers. His name is on the entry, so he only accepts a suggestion that explains itself.

Portrait of Nadia, the team lead persona

Nadia · team lead, 4 accountants

"Let me understand without calling back."

Her review load is the heaviest. She doesn't know why an entry was made this way, and the accountant has forgotten too. She wants to save time without the AI doing the job: juniors need to keep learning.

The problem

What people said didn't match what they did.

The interviews pointed one way. The data pointed another. I designed for what actually happens.

01

Alerts nobody reads

Julien said he reads every alert. In reality, 72% are closed without action, in bulk. Anomalies are already detected. Adding detection would only add noise.

02

Help lives outside the product

Julien said he doesn't need help. He calls Nadia two or three times a week and exports to Excel. Fewer than 1 in 20 users opens the help center.

03

Coming back to a line means starting over

Julien puts a line on hold and emails the client for a receipt. When the answer arrives days later, he has forgotten what he already checked.

04

The reviewer can't see the reasoning

Nadia doesn't know why an entry was made this way. She calls the accountant back, who has forgotten too. Juniors don't dare to call.

So the problem wasn't detection. It was that the reasoning behind each line got lost between tabs, phone calls and Excel.
Synthesized from the verbatims and product data in the brief.

Where I invested

Three moves. Two of them need no AI.

Julien forgets what he checked. Nadia can't see what was checked. I focused on the bank line that won't match.

01 · Gather

No AI

The document, the history and the balances in one place, on the line.

02 · Suggest with reasons

✦ AI

The AI suggests an account, explains why, and says what is still missing before it can be sure.

03 · Keep the trace

No AI

What was checked and decided. Julien finds it again, Nadia reads it.

I challenged the brief. The panel and the trace don't need AI, and they carry most of the value. They ship first.

Left asideAlert triage (V2) · automatic validation · a dedicated review view for Nadia

What the brief demanded

Every decision answers a constraint. Each one has a price.

The accountant's name goes on every entry, so his reputation is on the line, not the software's. The design could not make the AI look smarter than it is. Every decision here is a direct answer to a constraint.

Constraint 01

The accountant signs, not the AI

Julien's name is on the entry. He won't validate what he doesn't understand.

Decision

The AI suggests. It never decides.

Every suggestion needs a human click. Everything can be undone.

Trade-off: less time saved on obvious cases.

Constraint 02

One mistake breaks trust for good

"If the tool is wrong once, I stop trusting it."

Decision

Verifiable reasons, no confidence score

The suggestion lists reasons taken from the data, and what's missing to be sure. When the AI isn't sure, it says so clearly, with the data to back it up.

Trade-off: a less spectacular display, less flexible wording.

Constraint 03

Too many signals, none are read

72% of alerts are closed without action.

Decision

One label per line, the most urgent one

A "To do" column says what blocks each line. Secondary information moves into the panel.

Trade-off: less information at a glance in the list.

Constraint 04

Nadia hasn't decided on automation

"Why not", then "if it's the AI, no".

Decision

No automatic validation at all

Repetitive cases stay manual. It also answers Nadia's wish: juniors keep learning by doing the work themselves.

Trade-off: repetitive work remains.

The flow, step by step

One line, from blocked to reviewed.

The main journey follows a single card payment: CB STATION, €86.70. It is blocked on day 1, waits for the client, comes back on day 4, and ends in Nadia's review.

UNDERSTAND Julien opens a line he can't match. DECIDE Julien validates, right away or once the document is back. REVIEW Nadia understands the entry without calling Julien. 01 Open reconciliation Each line says what it's missing. 02 Click the line The panel opens beside the list. 03 Read the line panel Document, history, suggestion, reasons. AI 04 Ask the client for the document Pre-filled message. Line on hold. 05 Save the trace What was checked is kept on the line. Auto 06 Receive the document The line rises: "Document received". 07 Reopen the line The trace shows first. 08 Accept the suggestion Entry booked. Can be undone. 09 Review the entry Who, what, why, in three lines. Account sure? Yes No 3 days later
Something Julien doesThe AI suggestsSaved automaticallyJulien decides
01

Understand without leaving the page

Day 1

Each line now says what it's missing. One click opens a panel beside the list: the document, similar past payments, and a suggestion with its reasons. Here, three payments at the same merchant in 2024 were booked to 6251 Travel. What's missing: no receipt confirms it's a business expense.

Bank reconciliation screen: transactions table with the detail panel open on the right
Hover or touch the panel to scroll it yourself.
Full bank reconciliation screen including the whole detail panel
02

Put it on hold, with a reason

Day 1

Since the document is missing, the main button is "Put on hold", not "Validate". Julien picks the reason, and a pre-filled message goes to the client. The line moves to the suspense account and will come back by itself.

Bank reconciliation screen: transactions table with the request panel open on the right
Hover or touch the panel to scroll it yourself.
Full bank reconciliation screen including the whole request panel
03

Pick up where you left off

Day 4

The line rises to the top: "Document received". The trace shows first: what was checked, what was suggested, what was asked. The receipt confirms the date and amount, and the client confirms it's business fuel. What was missing is now resolved. Julien accepts, and can still undo.

No starting from zero.

Bank reconciliation screen: transactions table with the detail panel open on the right, receipt received
Hover or touch the panel to scroll it yourself.
Full bank reconciliation screen including the whole detail panel
04

Review without calling back

Nadia

The trace is saved automatically and can't be edited. Nadia sees who decided, what was booked, and why, in three lines. She can tell the AI's steps from Julien's.

Decision trace: who, what, why, then every recorded step with its author: Julien, the client or the AI.

AI steps and human steps, clearly separated.

When it doesn't go to plan

The edge cases decide whether people trust it.

The happy path proves nothing. Trust is won when the history disagrees, when there's nothing to go on, or when Julien changes his mind.

Contradictory history: two accounts shown side by side, neither preselected.

History disagrees: both accounts shown, nothing preselected.

Not enough elements: the AI proposes nothing and lists why.

Not enough to go on: the AI suggests nothing rather than guess.

SituationWhat the screen does
Two accounts in the historyBoth shown, neither preselected
No historyThe AI suggests nothing and says why
Julien corrects the suggestionHis choice applies, the trace keeps both the AI suggestion and his choice
Julien undoes after validatingThe line goes back to not reconciled
The client doesn't answerThe line shows its age, follow-up in one click

The contract with the devs

One matrix: which state can carry which label.

Rows are the line's state. Columns are what the "To do" column can show. Each cell says whether the combination exists, what it means, and where it appears in the journey.

Document missingBlocked, Julien must act
✓ Document receivedUnblocked, ready to process
✦ Account suggestedThe AI suggests, nothing missing
noneEmpty column
Not reconciled
✓ YesNo document in the file: it has to be requested or added.e.g. CB STATION, day 1
✓ YesBack from waiting: the client answered, the line rises to the top.e.g. CB STATION, day 4
✓ YesNothing's missing: the AI suggests an account or an invoice to match.e.g. VIR SEPA CLIENT
✓ YesThe AI has no sure answer: Julien chooses, the panel explains why.e.g. PRLV ASSUR (shared history)
Suspense
× NoJulien has already acted. Red here would be noise.
× NoAs soon as the client answers, the line goes back to Not reconciled.
× NoNothing is actionable while waiting. The suggestion stays in the panel.
✓ Yes, alwaysThe ball is with the client or Nadia. The reason shows in grey under the label.e.g. CB STATION, day 1 to day 4
Reconciled
× NoA reconciled line has its document, or its absence is accepted in the trace.
× NoThe document was used; the information lives in the trace.
× NoOnce validated, it becomes "AI suggestion validated by Julien" in the trace.
✓ Yes, alwaysIt's done. Nadia reads who decided, what and why.e.g. URSSAF, TELECOM, BANK FEES

Rule 1

One label per line

If two apply, the most urgent wins: Document missing › Document received › Account suggested › none.

Rule 2

Labels only on lines to process

On hold or reconciled, there's nothing for Julien to do: the column stays empty.

Rule 3

One color, one meaning

Red = blocked. Green = good news or done. Purple ✦ = AI. Orange = waiting. Everything else is grey.

Scope

An MVP in three batches, deliverable with 3 devs.

The two batches without AI come first and solve most of the problem on their own. Batch 3 then adds the AI suggestion, on top of a context the accountant can already check.

Batch 1

No AI

See what's blocking

  • The "To do" column, one label per line
  • The panel: document and similar operations

Batch 2

No AI

Stop starting from zero

  • The automatic, read-only trace
  • Put on hold with a reason and a pre-filled message
  • The line comes back when a document arrives

Batch 3

✦ AI

The suggestion

  • One account only if the history agrees
  • Two accounts if it disagrees
  • "Not enough elements" otherwise
  • Undo after validation
V1 doesn't
  • Validate automatically
  • Suggest an account without history
  • Triage alerts
V2
  • Client request and reply inside the product
  • A review queue for Nadia
  • Alert triage, filters
Line panel annotated for developers.

Annotated for handoff

  1. Header

    Label, amount, date: from the bank statement. A single label, the most urgent.

  2. Document

    Searched in the client file. States: found, missing or received.

  3. Similar operations

    Same merchant, 3 at most, link to the general ledger.

  4. Suggestion ✦ AI

    Only if the history agrees. Otherwise two accounts, or nothing. Reasons drawn from the data.

  5. Buttons

    The main button follows the state: document missing → Put on hold.

  6. Trace

    Saved automatically, read-only: what was shown and what was decided.

How I worked with AI

I worked with AI the way I want Julien to.

It suggests with its reasons. I check. I decide. The same rule I designed for the product applied to how I made it.

Domain vocabulary
Claude AI

AI did

Translated the brief and the accounting vocabulary into plain terms, so I could understand the domain better and faster.

I decided

Built my own understanding of reconciliation from it, then framed the problem and made every design call myself.

Diagrams and visuals
Claude AI

AI did

Laid out the journey diagram, the annotated screen and the matrix visuals, and suggested some layouts.

I decided

Built every screen in Figma myself, the matrix and its three rules, and all the edge cases.

Consistency check
Claude AI

AI did

Re-read my mockups for inconsistencies.

I decided

Fixed my own data errors (balances, dates, debit and credit) and the ambiguous language, such as replacing "Suggestion seen" with "Account suggested".

AI widened what I could check. It never signed anything.

Validation · test plan

A protocol built to catch blind acceptance.

7 participants: 5 accountants, 2 of them juniors, and 2 team leads. 45 minutes on the prototype, thinking aloud, with an observer taking notes.

"It's the end of March. You're working on client file SAS A••••. Four lines are left unreconciled. You want to finish before your team lead's review."

Task 1

Process the fuel station line

Does he read the reasons? Open the history? Leave the screen?

Task 2

Process the insurance line

Does he pick the most frequent account by reflex?

Task 3

Resume the fuel line to simulate a real scenario (3 days later)

Does he find what he did without searching? Re-check it anyway?

Task 4

Review the entry (team leads)

Can she tell the AI from Julien? Validate without calling?

Success

They decide without leaving the screen, explain their decision in their own words, and don't re-check what's in the trace.

Red flag

Someone accepts without reading the reasons.

How I'll measure it

Time to decide, not acceptance rate.

Main metric

Time between opening an unreconciled line and the decision.

Stay in the product

Weekly Excel exports (6 in 10 users today). Lines reopened before deciding.

Stop letting lines linger

Days spent in the suspense account.

Trust, without excess

Suggestions accepted, corrected, refused. Undos after validation.

Baseline measured over 4 weeks before setting targets with the PM.

The takeaway

It suggests. Julien decides. Nadia understands.

The strongest AI decision in this project was where not to use it. Gathering the context and keeping the trace solved most of the problem on their own.

The AI came last, and only where it could explain itself.

Enlarged image