Case study 04 / 06 · B2B retail CRM · In-store sales app · Agentic UX
From ten decisionsto one confirmation
In a B2B retail CRM used by sales teams in store, the most critical step is the transaction recap: a sale worth several thousand euros has just been paid, and the client is waiting at the counter. Sales associates had to close it by hand across three tabs. I redesigned this last step around an agent that prepares the closing plan, does what is safe on its own, and stops only where a human has to decide. The sales associate no longer does the closing. They supervise it.
In 30 seconds
- 3 → 1tabs to visit after payment
- ~10 → 1decisions, plus one compliance check that stays human
- 3levels of autonomy: done by the agent, one tap, you decide
- 3states designed: closing plan, awaiting approval, closed with audit trail
- The problem
- The system already knew the amount, the client, the product and the tax-refund eligibility, yet the screen asked the sales associate about all of it, one tab at a time.
- What I did
- I analyzed the screen with four Agentic UX lenses, sorted every post-sale action by reversibility and stakes, and designed a closing plan the agent prepares and the sales associate validates.
- The outcome
- One screen, one confirmation. The only friction left is the friction that protects the client and the business: the compliance check and the manager approval.
- Role
- Product Designer, owner of the Retail Transaction flow
- Client
- B2B retail CRM for in-store sales teams, under NDA
- Focus
- Agentic UX · Human in the loop
- Status
- Concept presented to the client
Le cas est présenté comme un CRM retail B2B (client sous NDA), sans nom de marque ni secteur luxe ; visuels anonymisés, données fictives. Confirme le statut : présenté au client en juillet, et ensuite ? (retenu, en dev, en pause).
The heaviest screen, at the worst moment
Step 4 of the in-store checkout, the transaction recap, comes right after payment, at the end of three already long steps. The client is still at the counter, the basket is worth thousands of euros, and the sales associate wants to end on a perfect note. Instead, the screen spread about ten small tasks across three tabs: launch the tax refund, set up the gift receipt, choose documents and a sending channel, register the product service, run the compliance check, request an approval. And “Close transaction” appeared on every tab, so nobody was ever sure the job was done.

The striking part: the system already knew most of the answers. It displayed “Client is eligible” for the tax refund and still asked “Launch now?”. It knew which channel the client always used and still showed three radio buttons.
The question I set out to answerHow might we let the sales associate close a sale in one confident gesture, without hiding anything that needs a human?
What the business needs
- A flawless end to a sale worth thousands of euros
- No missed tax refunds or services
- Compliance that can be proven
What sales associates need
- Know at a glance what is left to do
- Stop re-entering what the system knows
- Get back to the client fast
Constraints
- The compliance check is a legal attestation
- Some sales need a manager's approval
- Stay inside the existing checkout flow
Thinking first, then using AI as a mirror
Before generating anything, I wrote my own answers to a set of Agentic UX questions, grouped in four lenses. Only then did I compare them with an AI's analysis. That order matters: it keeps the designer steering the agent, instead of reacting to whatever it produces.
One sentence instead of three tabs
The screen served ten tasks, but the sales associate had one intent: “close this sale properly”. The agent can assemble everything around that intent.
Reversible or not? Costly or not?
Reversible and low stakes: the agent acts. Reaches the client or money: one tap. Legal or approval: always human, whatever the confidence.
Show the “why”, on demand
Each pre-filled choice explains itself in a few words (“client's channel for the last 3 sends”) and stays editable in one tap.
Plan before action
On an €8,600 sale, the agent never acts and reports afterwards. It shows its plan, and everything can be changed before confirming.
An autonomy map, not a flowchart
The key design decision was not a screen. It was deciding, for each action, how much freedom the agent gets. I placed every post-sale action on two axes: can it be undone, and what does it cost the client or the brand if it is wrong?

- 01
Assemble the plan when the screen opens
Amount, client, product, eligibility and the client's usual channel are already in the system. The agent gathers them before the sales associate even looks.
- 02
Sort by autonomy
Safe actions are already done, client-facing ones are pre-filled and wait for one tap, legal ones are isolated and can't be skipped.
- 03
Explain, on request
A short reason under each line, and a “why?” link when the agent made a judgment, such as tax-refund eligibility.
- 04
Execute, then log
After confirmation the agent runs everything and writes an audit trail: who did what, the agent or the sales associate, and when.
What I designed
The agent lives inside step 4 of the same checkout. There is no chat bubble and no separate “AI space”: the sales associate stays in the flow they already know. I designed the three states of the step.

1. The closing plan
Three groups, read top to bottom: “Already done”, “Ready, one tap” and “You decide”. The button tells the truth: “1 item still needs your check”. Service registration, once hidden in its own tab, is now just one more line in the plan.
2. Awaiting approval
When a sale needs a manager, the agent follows the request and tells the sales associate they can go back to the client. Everything else stays prepared but on hold, and the confirm button unlocks only after approval.
3. Transaction closed
Instead of a simple “Done”, a short audit trail. Each line is tagged “Agent” or “You”. On a sale of this size, being able to show who attested the compliance check matters as much as the sale itself.
Autonomous, never invisible
Some friction is the feature
The compliance check is never automated. It is a legal attestation, so a person has to carry it. Knowing where the agent must stop is the design.
Nothing hidden
Everything the agent did on its own is listed, not buried. The sales associate sees the whole plan before anything reaches the client.
Every choice can be overridden
Channel, gift receipt, refund, service: each line has “change”. The agent suggests, the sales associate owns the decision.
A trace and a way back
An audit trail separates agent actions from human ones, and agent actions can be undone or resent.
What changes
tabs to visit after payment
decisions, plus one human compliance check
of agent actions visible in the audit trail
Ces chiffres décrivent le design, pas une mesure. Si tu as un retour du client ou un test avec des vendeurs (temps de clôture, oublis de tax refund), ajoute-le ici.
What I learned
- The real lesson: after three long checkout steps and a payment worth thousands of euros, the last step asked for even more technical and sensitive input. That is exactly where people are most tired and most at risk of a mistake. The final step had to become fluid, with the system doing the work instead of asking the user again and again.
- Automating was a way to relieve sales associates and lower their cognitive load, not to remove them. Everything the system already knew is now done or pre-filled, and only two decisions stay human: the compliance check and the manager approval.
- I used this screen to test the Agentic UX exercise I then ran with my design team: pick a heavy screen, find where an agent removes clicks, bring a before / after.
At the end of a long, high-stakes flow, the best thing a product can do is stop asking. Automate what the system already knows, so people only spend their attention where it truly matters.
Van Cleef & Arpels: High Jewelry app
Private access to the High Jewelry catalog, local pricing and a direct line to a personal sales associate.
Read the case study