From 7,490 manual tasks
to a reply in one minute

"We didn't expect it to start with an audit. Mary just looked at our CRM and showed us numbers we hadn't seen ourselves: how many deals were sitting still and how many times managers had manually given themselves a 'contact the client' task. After that it was clear where to start."

About the company

FF makes printed products: photo books, portfolios and custom printing. It sells through fabrika-fotoknigi.com, and clients write in via chats, e-mail and web forms. Client records live in Bitrix24, while order fulfilment runs in an ERP system and on the production floor.

Inside the company there are two teams: MIU handles sales and client development, and TsPK deals with questions about the website, layouts, defects and delivery. For the client it's all one "Fotoprint", and they don't pick a department.

The challenge

The company did have processes, and fairly detailed ones. Things got complicated not where a process was missing but at the hand-offs: one client is in sales, support and orders at the same time, a request changes owners along the way, and order data is split between Bitrix24 and the ERP.

Before automating anything, we needed to see the picture in numbers. Mary connected to the working Bitrix24 and traced the client's path from the first message to a repeat order.

What the audit showed

• A single entry point doesn't mean single record-keeping. The same message in different open channels is handled differently: in one it creates a lead, in another it opens a deal straight away, in a third it stays in the chat only.
• The CRM piles up dead deals. 87.7% are marked "in progress" but haven't moved for months. Nobody closes them, because closing is something nobody does by hand.
• The routine shows in the numbers. The "contact the client" task was created manually 7,490 times. That's exactly the kind of work an agent does.
• Half of the deals have no source. 53% have no record of where the client came from, so there's no way to tell which ads bring in money.

A separate finding of the audit was what we couldn't see: tasks and projects, automation robots, production operations and the exact org structure required additional access. We recorded the data gap as a gap instead of filling it with guesses.

The solution

Mary isn't built as yet another board. It's a single layer on top of the existing systems, right where context gets lost today when work is handed on.

One queue for all incoming requests

Chats, e-mail and forms land in a shared inbox, and every message becomes one task regardless of the channel. If a channel or the model is temporarily unavailable, the message waits in the queue and doesn't get lost.

Mary recognizes the client before replying

Mary looks for a match in the database by phone, e-mail or username and opens the client card with their history, orders and past questions. If there are several matches, it asks for a check and doesn't merge people at random. If the client isn't there, it creates a card once, checking for duplicates.

Topic, risk and priority

Before the first reply, Mary works out the topic of the request, the possible damage and the next step. Defects, lost shipments, missed deadlines and sharply negative messages go to the top of the queue, so an important request no longer waits just because it came in later.

Auto-replies only where there's a confirmed source

At first Mary suggests a reply to the manager. Sending on its own is switched on only for five to ten agreed scenarios and only if the answer is based on the approved knowledge base. No source means no confident auto-reply: the question goes to a person, and the gap in the knowledge base gets flagged.

Difficult cases go straight to a person

Low confidence, critical risk or a non-standard decision is always handed to a specialist, together with the history, the order and a suggested next step. The person starts with the decision instead of digging through the conversation.

Personal data is scrubbed before it reaches the model

Before a message goes to an external model, it passes through local rules: phone numbers, documents and bank details are masked inside the company's infrastructure. Masking quality is checked on a test set: how much was found, how much was hidden unnecessarily and how long processing took.

Plan and results

The work is split into four checkpoints, each with a result you can open and show to the team.

1. by August 1 One queue for all incoming requests

   Client lookup in the database, new cards created with a duplicate check.

2. by August 15 Request topic and urgency

   Critical requests go to the top and are handed to a person along with the context.

3. by August 31 Working pilot with the staff

   5–10 auto-replies based on the confirmed knowledge base, a personal data filter on Fotoprint's server, speed and stability metrics.

4. by September 15 Client owner and segments

   A main manager and a backup, three client portfolios, a supervisor queue.

By the end of the cycle a message goes all the way through: it lands in the inbox, the client is found or created, the topic and priority are set, and either a reply is sent or a person steps in. This route is measured and repeated. The next scope is chosen based on the pilot's metrics, not on a wish to expand.

Results

Half of all requests, 52%, are closed without a manager: Mary replies on its own where the knowledge base has a confirmed source. Everything else goes to a person along with the history and a suggested next step.

The average response time to a client is about one minute. Before, it depended on which channel the client wrote in and when a manager got to them.

About 600 requests a week go through a single process. For comparison: the "contact the client" task was created manually 7,490 times, and that volume of routine work was the first candidate for automation.

What's next

After the first cycle, Mary gradually moves from assistant to main operator: first it handles safe topics, then a selected client portfolio, then the whole client journey. People stay on critical decisions, quality control and changing the rules.