Case study · Business banking · AI agents

Allowance: an agent that pays the routine bills and stops on the risky ones

Most of a company's vendor bills are routine, and the one that isn't can cost tens of thousands. Allowance puts an agent in the business bank account. It pays the routine bills, stops and asks about the risky ones, explains every payment, and waits a day before any money leaves.

Context

Concept exploration, set in a business banking app built with Kinetic

My role

Product designer and design technologist. I did the research, designed both apps and built them

Built with

Kinetic, my banking design system. A desktop app and a phone app that run the same scenario

Status

Working prototype. Not tested with users and not released

Allowance · live app

Both are the working app, and every link in them goes to a real screen. Try stopping a payment, or open the Harlow bill.

01 · The problem

Routine work, with a rare and expensive failure

Most of a company's bills in a month are the same vendors for about the same amounts. Someone approves them one by one anyway.

They do it because the one that is wrong can be very wrong. A criminal takes over a vendor's email and sends an invoice with new bank details. The same bill arrives twice under two numbers. An amount is far above the usual. Approving twenty routine bills to catch one of those is slow, and a person who approves twenty in a row stops reading them.

Today's bill pay tools flag the risky bill and still leave every approval with a person. That catches fraud, but it gives no time back on the routine ones.

$3.05bn

lost in 2025 to scams that trick a business into sending a payment to the wrong account, across 24,768 complaints.

FBI Internet Crime Report 2025, business email compromise

74%

of US organizations were affected by that kind of scam in 2025.

AFP 2026 Payments Fraud and Control Survey, 465 treasury staff

0.8% to 2%

of a company's yearly payments are duplicates or errors, from the best run companies to the worst.

APQC benchmark, reported by CFO.com, 2020

What this evidence does not show. The fraud survey covers organizations of every size and leans large, and the error benchmark is from 2020. I also looked for how long a company of this size spends paying bills and found no figure with a named source, so this page does not quote one.

02 · Constraints

What the design had to work within

Most of these are about money that has already left, or about who is allowed to say yes. A screen can slow a mistake down. It cannot always undo one.

A transfer is hard to take back

A bank can be asked to reverse a transfer only within 5 banking days, and only for a duplicate, a wrong recipient, a wrong amount or a wrong date. A payment someone confirmed is not on that list. Modern Treasury, on the US bank transfer rules

Owners do not trust an agent alone yet

78% of small business owners do not fully trust AI with low level tasks unless someone is watching. The agent has to show its work before anyone lets it act. Bluevine survey of 942 US owners, April 2026, reported by Stacker

Some answers belong to one person

New bank details and a new vendor are the owner's call. The person at the desk sees the same question and has to be able to pass it on, not answer it.

The agent has to read the invoice right

Every check depends on what the agent took from the bill. If it misreads an amount or an account, its explanation is wrong with confidence. So the invoice as it was sent is always one tap away.

Who pays for a mistake is not a design question

If the agent pays a bill it should have stopped, who covers the loss is set by the bank's terms. The design can make that rare. It cannot decide it.

03 · Decisions

Three decisions that shaped it

Each one gives something up so a wrong payment is caught before the money moves.

What every option was measured against

It had to give the routine bills back, put every risky bill in front of a person, and leave time to stop a payment before the money moves. And it had to feel calm: nothing leaves the account that would surprise the owner, and every payment says why.

How four ways of paying bills compare against the three goals
Approve every bill Flag the risky ones One monthly cap ChosenAllowance
Routine bills paid without a person NoA person approves each one NoEvery approval stays with a person Yes YesWhen every check passes
Risky bills reach a person In theoryAfter twenty in a row, people stop reading Yes NoOne wrong payment fits under a month’s cap YesNew bank details, a new vendor and a bill far above the usual always stop
Time to stop a payment NoNo waiting day NoNo waiting day NoNo waiting day YesPayments leave at 9:00 AM the next business day
Decision 1
Pay the routine bills, not just flag them

What I chose

The agent pays a bill on its own when every check passes, and says why.

What else I considered

Approving every bill, as companies do now. Or an agent that flags problems but never pays, as today's tools do.

Why

Both keep a person on every routine bill, so no time comes back.

What I gave up

A yes on every payment. The company has to trust the checks, so every payment has to explain itself.

What I kept

The flag. A risky bill still stops and waits for a person, as the Harlow bill does.

Decision 2
Three things always stop, whatever the limits

What I chose

New bank details, a new vendor and a bill far above the usual always stop and ask, even under every limit.

What else I considered

One monthly spending cap. It is the easiest rule to explain and to build.

Why

A cap big enough for a month of bills is big enough for one payment to the wrong account. The risk is in the bill, not the total.

What I gave up

Some honest bills wait for a person every time, like a first order from a new supplier.

What I kept

The cap, as one of two limits set at setup: the most it pays on one bill and the most in a month.

Decision 3
Every payment waits a day

What I chose

Scheduled payments leave at 9:00 AM the next business day, and anyone at the company can stop one with a tap until then.

What else I considered

Paying straight away, which vendors like. Or putting the agent in the accounting tool, where the invoices already are.

Why

The waiting day is the undo. After the money leaves, the bank's options are narrow. And only the bank can hold a payment, and knows which account the last 14 payments went to.

What I gave up

Speed. Vendors are paid a day later, and nothing can go out the same day.

What I kept

No one on the routine bills. They are scheduled on their own, and the waiting day is the only delay.

04 · The scenario

One company, two people, one month

Everything on this page follows the same story, and both apps run it.

Rivera Landscape is a commercial landscaping company with 60 staff. It banks with Kinetic. By the middle of October the agent has read eight vendor bills. Five passed every check. Three did not.

Dana Okafor, accounts payable. Works through the bills at a desk. She can stop, hold and approve, but not everything.

Alex Rivera, owner. Mostly on site, with a phone. Two answers are Alex's alone: new bank details and a new vendor.

The eight vendor bills in October that this case study follows
Vendor Bill What the agent found What happens
Fleetwise, fuel $19,246.80 Nothing. In its usual range PaidThe next morning, without asking.
Metro Waste, disposal $4,180.50 Nothing PaidThe next morning, without asking.
Greenline Wholesale Nursery, plants $42,360.00 Nothing. Paid every month PaidThe next morning, without asking.
Coastal Mutual, insurance $9,860.00 Nothing ScheduledIt leaves tomorrow at 9:00 AM.
Greenline Wholesale Nursery, again $42,360.00 The same bill, sent again under a new invoice number. Every check on the bill itself passes ScheduledWith a heads up that it is the second Greenline bill this month. Someone stops it before it leaves.
Ridgeline Stone and Gravel $14,720.00 Far above the $6,000 to $9,000 Ridgeline usually bills Stops and asksDana can answer.
Sunbelt Irrigation Supply $6,275.40 A new vendor. Nobody has paid them before Stops and asksOnly Alex can add a vendor.
Harlow Equipment Rental $28,400.00 The bank details on the invoice are new Stops and asksOnly Alex can answer, after calling Harlow.

The job

"Let the routine bills go without me. Stop on the ones that could hurt us, while there is still time to do something about it."
05 · The design

Four parts, in the bank account the company already has

The agent has three ways to act, and the whole design is about which one a bill gets. It pays quietly when every check passes. It pays and tells you when the bill is fine but something around it is odd. It stops and asks when the bill could be a fraud or a mistake. Each phone below is the working app, built with Kinetic.

Set up: three choices on one screen

The most it can pay on one bill, the most in a month, and which vendors it can pay.

The vendors offered are the ones the company has already paid, so the list starts from its own history. A sentence under the choices says the rules back in plain words and changes as they change.

  • Three presets for each limit, not a number field. A preset is one tap, and nobody has to invent a figure.
  • Three things are not settings at all: new bank details, a new vendor, and a bill far above the usual always stop, whatever the limits say.

Stop and ask, and ask the right person

Harlow's invoice is under every limit. The agent stops anyway, because the account it asks for is not the one the last 14 payments went to.

The screen puts the two accounts side by side and says what to do: call Harlow on the number already on file, not one from the invoice. Paying the new account needs a tick to say that call was made. Holding the bill is the first button, because it is the safe one.

  • Dana sees the same stop at her desk and cannot answer it. Her button is "Send it to Alex", and it reaches his phone with everything on the page.
  • "Pay the account on file" is offered too. It is often the right answer, and it is safe.

Every payment explains itself

A payment the agent made without asking still has to answer one question later: why did this go out?

Each bill lists the checks the agent ran, in words: the vendor is on the list, the bank details match the last 23 payments, it is under the limit, it is in the usual range, it fits in the month. Below that sit the bill, the invoice as it was sent, the history with that vendor, and what happened and when.

  • Each check has a tick, a warning or a cross as well as its color, and a hidden word for screen readers.
  • When a person says yes to something the agent stopped on, the check stays on the record with their name beside it.

Stop it in time

Every payment the agent schedules waits until 9:00 AM the next business day. Until then, anyone at the company can stop it with one tap.

Greenline's second bill passes every check, so the agent schedules it and says it is the second this month. Stopping it takes one tap and can be undone. Once a transfer has left, the app is honest about what remains: the bank can be asked to reverse it for a few days and a short list of reasons, and after that the only route is to ask the vendor.

  • On a card you can dispute afterwards. On a bank transfer the cheap moment to fix a mistake is before the money moves.
  • Stopping asks for no confirmation, because it is the safe direction and it can be undone.
06 · Two designs

The desk works the queue. The phone gives the answers

The two apps are one design. The same screens, the same rules and the same data, laid out for where each person is.

Dana has eight bills to get through, so the desktop keeps the list on the left and the bill beside it, and nothing is opened and closed. Alex has one question to answer between site visits, so the phone opens straight on that bill with the safe button first.

Desktop · Dana's view of the Harlow bill Open standalone

List and detail together

On the desktop, bills, vendors and activity keep the list in view with the chosen item beside it. On the phone the same two screens follow one another.

Who is signed in matters

The desktop opens as Dana and the phone as Alex. The two answers that are Alex's alone show Dana a "Send it to Alex" button. The desktop can switch person, to show both.

Nothing is a dead end

Home, Bills, Vendors, Agent, Activity and Alerts are all real screens. Every row opens, every vendor has a page, and every action writes a line in the record of what was done and by whom.

07 · Built with Kinetic

What the design system already had, and what it gained

Most of Allowance is Kinetic as it stands: the phone, the activity list, the limit meter, the buttons, the status pills, the switch and the timeline.

Kinetic was built for a consumer banking app on a phone. This is its first business account and its first desktop app, and four larger pieces were missing. I built them from Kinetic's own tokens so they can go back into the system.

Checks list

One line per check, each with an icon and words, for passed, heads up and stopped.

Account comparison

The account on file beside the account on the invoice, with how many times each has been paid.

Sheet and dialog

One component that rises from the bottom on the phone and sits in the middle on the desktop. It holds keyboard focus and closes with Escape.

Desktop shell

A rail of sections and a list that keeps its detail beside it, drawn from the same screens as the phone.

08 · Outcome

What success would mean: no payment ever reaches a criminal's account

There are no results yet. Allowance has not been released or tested with users, so these are the targets I would hold it to.

What does exist is the working app: both designs run the same month of bills end to end, so every stop, hold and payment in the scenario can be tried. The first two targets are about safety. The rest are about whether the time really comes back.

The seven measures for Allowance, each with its target, the reason for the number and what to change if it is missed
What is measured Target Why this number If it is missed
Safety
Payments sent to new bank details without the owner confirming Zero This is the failure the feature exists to stop. It is not traded against anything. Stop the release until it is fixed.
Wrong or duplicate payments stopped before the money leaves 9 in 10 The waiting day is the one cheap moment to fix a mistake. After it the bank has 5 banking days and a short list of reasons. The heads up is not being seen. Move it from the bill to an alert, or hold a flagged payment for a second day.
Time back
Routine bills paid without anyone touching them 4 in 5 In the scenario five of eight pass every check. A real month should do better as each vendor's history builds up. Find which check stops the most bills. The usual range is probably too tight for vendors whose bills swing.
Questions answered with a hold, a change or a no At least 1 in 4 If nearly every question gets a plain yes, the agent is asking about things that did not need asking. Loosen the check that produces the most plain yeses.
Vendor bills paid late Fewer than in the three months before Each company is its own baseline. In one US survey 42% of small businesses said they had paid their own vendors late. Look at how long questions wait for an answer. A bill held up by an unanswered question is late for a new reason.
Hours spent on bills in a month Measured on day one and at 90 days I found no public figure with a named source, so this is a method and not a number. If it has not dropped, the time has moved from approving to answering. Go back to the row above this one.
Companies that still have it on after three billing months 60% One month is a single run. Three is a habit. Ask the ones who turned it off what happened just before.

The 9 in 10, 4 in 5, 1 in 4 and 60% figures are my own starting points, and I would expect to move them after the first months of real numbers. The late payment figure is from the Intuit QuickBooks 2026 Late Payments Report.

09 · Reflection

What I would test first

The second Greenline bill. It passes every check, so the only thing between it and the bank is a line someone has to notice before 9:00 AM. If people miss that line, it belongs in an alert, not on the bill, and that is the first thing I would put in front of real users.

More case studies

Three more about money and oversight: the real cost of a payment abroad, a home offer signed at the property, and a supervisor stepping into a live call.

1QR, a travel payment app in an outdoor theme
A payment app for scanning QR codes abroad. It runs in two outdoor modes from the Wolf-Rayet design system: a high contrast day mode that stays readable in direct sun, and a dim night mode in one warm color that protects night vision.
KeySign, signing a home offer at the property
Buyers decide right after a showing, but offer tools need a desk and every state has its own forms. KeySign lets the agent write, sign and send a compliant offer from their phone at the property, then open escrow.
Watchtower, supervising AI voice agents on calls
One person watches every AI voice agent on the line from one dashboard. When an agent tries something outside its limits, Watchtower pauses it and asks the supervisor to approve, edit, hold or take over.