rodion prokhortsev
ru
All work

Role
Product designer
Team
PM, designer, 2 developers, LLM developer, analyst
Platform
Web
Year
2025

AI assistants that take over a sales manager's first consultation when a B2B client chooses hardware. The client describes the task in their own words and gets verified configurations with prices, stock and a ready request.

Context

A distributor of server, network and office hardware for B2B clients.

The client's path before launch: a request on the website → a manager clarifies the task → a presales engineer puts the configuration together → a commercial quote. The “request → quote” cycle took up to 72 hours.

Problem

Managers and engineers spent their time on typical requests: “a server for 1C with 60 users”, “an office for 20 seats”, “licences for a department”.

Without expertise, a client can't put a configuration together from the catalogue. So they leave to make a call, or go to a competitor.

Context was lost when a manager handed the task to an engineer: the client had to tell it all over again.

More requests meant more staff, and selection didn't scale.

Goals

Reduce the load on managers and engineers from typical selections.

Shorten the “request → quote” cycle (up to 72 hours before).

Raise the website's conversion to requests (24–26% before).

Increase LTV through repeat selections and renewals.

Hypotheses

Hypothesis 1. Typical requests can be closed by a dialogue without a person if the bot asks in the language of the task (users, 1C, budget) rather than the language of hardware.

Metric: the share of requests the bots close on their own.

Hypothesis 2. The client shouldn't have to choose which of the four bots to go to: there should be one entry point that routes them to the right one.

Metric: the share of requests that go through the assistants.

Hypothesis 3. A B2B client will trust AI with the selection only if they see the logic: what the bot understood, why it recommends an option and what an engineer will check.

Metric: the share of selections that reach a request or a quote.

Hypothesis 4. If the path from selection to request needs no re-entering and no phone call, conversion to requests will grow.

Metric: the website's conversion to requests.

Hypothesis 5. The bot should hand complex tasks to an engineer honestly and with the context, or we lose large deals.

Metric: the share of handed-over dialogues that reach a request.

What we turned down

One universal assistant. The categories have different results (a configuration, a set of workstations, a network diagram, a licence matrix) and different clarifying questions. We kept four assistants and one entry point with routing.

A step-by-step wizard. It doesn't work with a client's spec or free text. From the wizard we took clickable answer options and the parameters panel.

An instant answer without questions. Without the budget and the load, the bot errs both ways: overpaying or falling short. We limited it to 2–3 questions that can be skipped.

Open access for every visitor. Prices and payment deferral are contractual, company details come from the B2B profile. Access is closed; new clients request it by their tax ID.

Automatic ordering without a manager. B2B purchasing needs an engineer's check, so the outcome is a request, not a payment.

Process

Analysis of the customer journey and of requests to sales.

Scenarios for the four categories.

Dialogue logic: questions, answers, options, handing over to an engineer.

First mockups and testing the hypotheses with the sales team.

Final mockups and development.

  • 3.1entry

    Shop catalogue

    “Servers” section, banner and the Orbit button

  • 1.2entry

    Single interface

    assistant selection page

  • 5.1entry

    Personal account

    history, requests, drafts

  • 1.1screen

    B2B sign-in

    access for client companies only

  • —branch

    Access request by tax ID

    a manager activates the company

  • 3.2screen

    Orbit in the side panel

    respects the catalogue filters

  • 1.2screen

    Assistants hub

    shared input or a choice of bot

  • 2.1screen

    Dialogue start

    Orbit · Desk · Grid · Stack

  • 2.2screen

    Clarifications

    ≤ 3 questions, parameters in the panel

  • 2.3screen

    Generation

    transparent log of steps

  • 2.4 · 4.1–4.3screen

    Solution options

    2–3 configurations / bundle / specification

  • 2.5branch

    Details and comparison

    table of differences

  • —branch

    Shop cart

    save and come back later

  • 2.6screen

    Request checkout

    company details from the B2B profile

  • 2.7goal

    Request sent

    a manager confirms within 30 min

  • 5.2branch

    Handoff to an engineer

    summary + dialogue history

  • 5.1screen

    History and requests

    continue a draft

  • main path
  • alternative / branch
  • entry point
  • screen
  • target action

01

Assistant hub

The user types the task into one shared field and the system routes it to the right assistant — or picks an assistant by its card

User task

A single entry point for those who don't know which of the four bots they need. Choosing a bot shouldn't be the user's job

Design decisions

Checks hypothesis 2

  • Shared input with auto-routing removes the mental load of choosing between Orbit, Desk, Grid and Stack
  • Example chips come from real sales requests and show that free text is fine
  • A bot card speaks about the result, not the technology: tags “Spec in PDF”, “As a bundle”, “Specification”, “Licences”

02

ORBIT. Starting the dialogue

The user describes the task in their own words, picks a typical one or attaches the customer's spec

User task

Remove the fear of a blank page and set the right expectations: what the bot can do, what the result will look like and where its limits are

Design decisions

Checks hypothesis 3

  • Three capability cards frame expectations: I understand text, I read specs, I give 2–3 configurations with price and stock
  • Frequent tasks are ready-made wordings for those who'd rather pick than write
  • “Attach spec” right in the input: a B2B client often already has a document
  • The note that an engineer confirms the result builds trust and honestly marks the limits of AI

03

ORBIT — server selection

The bot has understood the message; the user answers by clicking an option or in their own words, and can skip a question

User task

Collect the missing parameters without turning the dialogue into a questionnaire. Every extra question is a risk that the user goes off to call a manager.

Design decisions

Checks hypothesis 1

  • First, confirm what was understood: tags “60 users”, “1C”, “≤ ₽900,000” show the user they were heard
  • 2–3 questions at most. Options come with hints (“recommended from 20 users”), so a non-expert can answer too
  • “Skip question” with a clear consequence: the bot picks an option with headroom
  • The “Selection parameters” side panel shows progress 5/7 and lets the user edit any parameter without typing

04

ORBIT. Generating the solution

The user watches the selection progress and can add a clarification or stop the generation

User task

Make the wait transparent. The reasoning log shows the same expertise as a live manager and explains where the numbers come from

Design decisions

Checks hypothesis 3

  • Skeletons of the three cards show the result format in advance, so the wait feels shorter
  • The composer stays active during generation: a clarification isn't lost and is taken into account
  • The estimated load in the panel bridges the client's task and the hardware

05

ORBIT. Configuration options

The user compares three options, reads why one of them is recommended, adds extras, places a request or downloads a quote

User task

The key value screen: instead of a consultation the client immediately gets verified configurations with price, stock and the option to order

What's on screen

  • Three tiers — “Basic / Recommended / With headroom”. An anchoring effect and a clear choice without studying specs
  • The “Why CFG-02” block explains the recommendation in terms of the client's task, not as a black box
  • “Download quote” — a document for approval inside the client's company

Design decisions

Checks hypotheses 1, 3 and 4

  • A specification with prices per item, VAT and a price fixed for 3 days — the transparency B2B procurement needs
  • Iteration chips (“Cheaper”, “+ UPS and rack”) — adjust in one click instead of a new request

06

ORBIT. Details and comparison

The user opens the comparison from the chat, looks only at the differences, picks a configuration or exports the table

User task

In B2B the buyer has to justify the choice to management. The comparison helps defend the decision and cuts calls to the manager.

Design decisions

Checks hypotheses 3 and 4

  • The “Differences only” switch removes the noise of identical specs
  • Rows “Load at 60 users” and “1C ceiling” translate hardware into the client's task
  • XLSX export — an artefact for internal approval

07

ORBIT. Placing the request

The user checks the pre-filled company details, picks delivery and payment terms, leaves a comment and sends the request

User task

Turn the selection result into a request with minimal input. What the system already knows about the client, we don't ask again

Design decisions

Checks hypothesis 4

  • Company details come from the B2B profile — no need to enter the tax ID and KPP
  • Deferred payment under the contract with a visible limit — the payment model B2B is used to
  • The dialogue history is attached to the request: the manager doesn't re-ask what the client already told the bot
  • “This is not a payment” and the name of the personal manager ease the worry before sending

08

ORBIT. Request sent

The user sees the request number and status, downloads the quote, writes to the manager or moves on to another assistant

User task

End the scenario with confidence: what happens next, who is responsible and when. And show the next step in the assistant ecosystem

Design decisions

Checks hypothesis 4

  • A status timeline with an SLA “by 14:37” makes the wait measurable
  • The request number copies in one click — it gets forwarded to colleagues and accounting
  • A cross-scenario via Desk ties the four mini-products into one ecosystem

09

Desk — workstation

The user describes the department and requirements, gets workstation profiles, swaps items, changes quantities and places a request

User task

Show that the dialogue pattern scales to another kind of result — not one item but a set for a group of employees

Design decisions

Checks hypothesis 1

  • “Replace” on every item — a precise edit without a new selection
  • Price per workstation and total for the set — two scales a buyer understands
  • A cross-link to Stack: software is offered right with the hardware

10

Grid — network specification

The user describes the office, floors and devices, checks the diagram and specification, and exports an XLSX for their engineer

User task

A network project is hard to grasp from a list of equipment. A diagram in the bot's answer makes the logic of the solution visible and verifiable

Design decisions

Checks hypotheses 1 and 3

  • Topology right in the answer: perimeter → core → access → end devices
  • The table by network level matches the diagram — easy to compare
  • Key parameters (ports, PoE budget, uplinks) — the client's engineer checks the solution in seconds
  • Specification export — the familiar format for tenders and approvals

11

Stack — software by role

The user lists roles and headcount, sees a “product × role” matrix, swaps products and calculates the cost of ownership

User task

Choose software from employees' roles rather than from products — that's how a department head puts the task

Design decisions

Checks hypothesis 1

  • A “product × role” matrix: you see at once who has which software
  • One-off and annual payments are split, showing the real cost of the first year and the following ones
  • A mark of the Russian software registry and a ready alternative — the answer to import-substitution requirements
  • “Remind me about renewal” — the basis for repeat subscription sales

12

Selection history

The user finds a past selection, continues a draft and tracks request statuses

User task

A B2B selection rarely ends in one session: it needs approval, a budget, clarifications. The account is the point of return

Design decisions

Checks hypothesis 4

  • The “Continue where you left off” card — the most frequent scenario is at the top
  • One status scale: draft → cart → request → payment
  • Filters by assistant and the original request in each row — a selection is easy to recognise weeks later

13

Handoff to an engineer

The user sets a non-standard task, gets an explanation of why an engineer is needed and continues the conversation with a person in the same chat

User task

An honest limit of what AI can do. Complex tasks go to an expert without losing context — the client doesn't have to tell everything again

Design decisions

Checks hypothesis 5

  • The bot recognises a non-standard task itself and explains why it hands it over
  • An automatic summary for the engineer, with data gaps highlighted
  • The engineer joins the same chat, and Orbit stays on call (@orbit)
  • Escalation rules are visible in the panel — the system behaves predictably

Impact

Measured a month after launch, outside the peak quarter. Hypotheses 1, 2 and 4 were confirmed. Hypotheses 3 and 5 were not measured separately: their contribution shows through conversion to requests.

Both shares are of all requests: of the requests that reach the bots, they close about 9 in 10 on their own.

The client gets a quote right after the selection instead of in 72 hours: the “Download quote” button on the options screen.

  • ~40%Of requests are handled by the bots. Hypothesis 2
  • 37%Of requests the bots close on their own. Hypothesis 1
  • +7 ppWebsite-to-request conversion: from 24–26% to 31–33%. Hypothesis 4

What's next

LTV can't be judged in a month. We watch repeat selections in “History” and the renewal reminders for subscriptions in Stack.

We go through the dialogues handed over to engineers and turn recurring tasks into new typical scenarios. That is how the share of requests closed on their own grows.

This site uses cookies for analytics