AI implementation in fence manufacturing: 636 prices out of the red and 84% of leads that never reached a site survey

An AI implementation case in fence manufacturing: a full-cycle fence and gate maker with its own workshop and seven installation crews, about 30 people, roughly ₽160M a year in turnover and about 4,200 leads a season. In five weeks the corporate Claude became the owner’s working tool: it is connected to seven company systems, rebuilt a year of financials from source documents, took the website price calculator out of the red, found where leads are lost, reviews the sales managers’ work every morning and, since October 20, answers customers on Avito on the night shift. Part of the effect is already working; part is still a calculation that can be checked against facts in spring 2027.

Published October 21, 2026 · Breakdown by Anton Ro Baten

Point A: where they started

• ·The data sat in five disconnected systems: the CRM, Avito, the 1C accounting of two sole-proprietor entities, an installation spreadsheet and the website. A simple question meant merging spreadsheets by hand.
• ·Nobody saw the real profit: more than half of the expense invoices in 1C were stuck as drafts, and the two entities’ costs were never consolidated.
• ·Website prices had grown historically: nobody knew which fence made money and which sold at a loss.
• ·Nobody listened to calls. The owner personally pulled postponed leads, wrote to customers and assigned follow-ups.
• ·The goal of “15 crews and a second site” existed without any calculation of what it runs into.
• ·The keys to the infrastructure were not with the owner: a contractor controlled the website and the domain, there was no corporate email, and work lived in personal mailboxes. A year earlier a manager left with the access, and passwords were changed by hand.

What we did

1. 1

   An environment, not a chatbot

   The first three sessions went not into AI but into the foundation. Corporate Google Workspace on the company domain with the owner as the super administrator; email moved without changing the address; single sign-on (SSO); a Claude Team plan; one working environment for everyone. Access is granted and revoked with one button, and the data stays with the company, not with whoever left.

2. 2

   Our own connectors to the CRM, Avito and 1C

   There was no ready connector to amoCRM for the task, so we wrote our own integration: a separate $6-a-month server, Docker, a secure connection on a company subdomain. The same for Avito (about 55 functions) and 1C (a separate read-only “AI analyst” user). Claude works with the data directly, without exports.

3. 3

   Safeguards from day one

   Any change in the CRM, on Avito or in prices is shown as a preview and applied only after the owner confirms. Messaging customers is allowed only for deals with a dedicated tag, with quiet hours, a limit on unanswered messages and no duplicates. Customer phone numbers are hidden and every action is logged. Before touching prices, a full copy of the calculator settings was taken — more than 1,100 fields, rollback in minutes.

4. 4

   A year rebuilt from source documents

   More than 40 exports from 1C and the CRM for the year: profit and loss reports, invoices, cash and bank of both entities, piece-work orders, payroll, the sales funnel. Four independent revenue sources — sales, reports, receipts and purchases — matched each other.

5. 5

   Prices from cost

   A cost model: purchasing, crew rates, delivery, site survey, the manager’s commission, taxes, rent, the winter payroll and advertising at ₽9,652 per deal. The bar is a margin of at least ₽20,000 per crew-day. Installation rates and minimums were changed, material prices were not; every change was checked with a live calculation on the website.

6. 6

   A daily review of the sales managers

   Every night all calls of the day are transcribed and the chats analysed. By 9:00 the head of sales has a document per manager in the CRM: metrics, what worked, what to fix, a breakdown of every call and a ready message to the customer to book a site survey. Done by hand from October 9 to 19, then on a schedule.

7. 7

   An AI manager on the night shift

   Trained on 35 calls of the best manager and the department’s chats: strong techniques kept as they are, mistakes fixed — a conversation always ends with a survey date or a callback. Tested in a dialogue where the head of sales played the customer, then the owner corrected every step. Since October 20, from 21:00 to 9:00 it answers Avito chats, prices the job with the calculator, books surveys no earlier than the next day and sends a report in the morning. It works under the head of sales’ supervision, and the company does not hide it.

8. 8

   Purchasing, finance, strategy

   The supplier order is assembled from the installation schedule, with unclear items — colour, thickness, quantity — flagged separately. An expense report for both entities: 1,954 payments in 22 categories; a passport of the 1C database. The 2027 strategy starts from a Point A confirmed by the owner: 10 crews by June, four priorities, roles and a 90-day plan.

What AI found in the data

• ·Only 16% of leads reached a site survey, while 91% of surveys became contracts. The bottleneck is the first contact, not the contract price or the surveyor. Avito chats are the worst at 8%, and the main reason is silence (70 of 104 such losses company-wide).
• ·Of 35 “they will get back to us” losses, 28 were phone calls: the price was named, the next contact was not set.
• ·Response time is 4–14 minutes during working hours, but one in ten evening Avito leads waited about 11 hours for an answer until morning.
• ·Long fences on the website sold at a loss of up to −5%, chain-link at up to −41%. Work on 90 m cost ₽159,176 and on 91 m ₽124,522: 100 metres came out cheaper than 90. Channel steel sold at ₽1,500 per metre — below the purchase price.
• ·The installation minimum was calculated as ₽42,000 instead of the intended ₽15,000; a sliding-gate frame cost differently in two calculators — ₽22,000 and ₽24,000.
• ·The workshop is loaded at 68% and the surveyor at 46%: the “welders are the bottleneck” hypothesis was disproved, and a second surveyor is not needed yet. Crews are loaded at 87% over the season — 110–119% in summer and 35–80% in spring and September: 189 free crew-days.
• ·Overhead is ₽41,904 for every crew working day: an order can now be priced by crew-days.
• ·Advertising is ₽7.2M a year at ₽1,720 per lead; an Avito deal costs half as much as one from Yandex Direct.
• ·Sliding gates with automation are ₽20–40K more expensive than competitors’: this explained the managers’ complaint that “we lose on automation”.
• ·On Avito, fences were listed under “Goods”, so for “turnkey fence” searches the company ranked 32nd–64th or not at all. Some listings spent ₽1,800–2,640 per contact versus ₽45–315 for the working ones; about ₽12K went to job-vacancy listings.
• ·The growth ceiling: tax regime limits, the long GAZelle truck and survey capacity converge at ~10 crews. The goal of “15 crews” moved to 2028.
• ·A ₽176,300 estimate in the CRM differed from the calculator by ₽6,300, and inside the difference was a ₽1,000 shortfall: 100×100 posts priced as 80×80. Managers sometimes edit PDF estimates by hand, so the CRM amount differs from what the customer received.

Point B: results

Why +2.4% on price gives +24% on profit: profit is about a tenth of turnover, and the mark-up is shared with nobody and goes into it in full. Hence the tenfold leverage — and the answer to why a ₽160M business should go through 1,100 calculator fields.

How we measured

• ·On September 22 the first estimate on incomplete data showed a loss of about ₽1.5M: costs ran ahead of revenue, and prepayments for unfinished work created the feeling of money. That figure was not accepted as a result, and the year was rebuilt from source documents.
• ·The verified picture: ₽13.7M stays in the company for January–September, about ₽15.5M for 12 months — after the owner’s salary.
• ·Two cross-checks matched on their own: ₽7.2M of advertising ÷ ₽1,720 per lead = 4,186 leads against the company’s “about 4,200”; ₽459K of Avito spend for July–September is 3.3% of the channel’s contract total against an independent estimate of “about 3.8% of revenue”.
• ·+₽3.3M is a calculation: the new prices applied to actual January–September orders, assuming the number of customers does not change. It will be checked against 1C and the CRM in spring 2027.
• ·The funnel is September: 582 leads in the main funnel after the CRM accounting was fixed. These leads are still being worked, so the shares compare across channels but run higher than the annual 16%.
• ·Every figure was cross-checked against several sources and confirmed by the owner.

What AI does not get credit for

• ·We don’t credit AI with the drop in Avito cost per contact from ₽1,095 to ₽382 or with the growth in September contracts: the new listing bids are only prepared so far, and October contracts have not matured.
• ·Decisions on prices, bids and messages to customers are made by the owner: AI shows every change as a preview and applies it only after confirmation.

What didn’t work

• ·Two figures turned out wrong: the first financial estimate showed a ≈₽1.5M loss instead of ₽13.7M staying in the company, and the first CRM report showed “99% of leads unanswered” because calls and chats were stored on contacts, not on deals.
• ·Three calculations diverged from reality: gate bundling distorted the mark-ups in 1C, a 5% target slipped into the model instead of 2%, and for fences with sliding gates the calculator applies a lower minimum. All three were caught by checking against the website before any decision.
• ·Two incidents in the website admin: 3D panel prices were zeroed and the thickness list collapsed. Both were fixed quickly; now there is a reference copy before changes, test edits only with consent, and a reload and check after every save.
• ·Mass data collection from Avito through the browser triggered an IP block. Since then — the official API, and the browser only in small portions.
• ·Of three stalled deals Claude handles in messengers, one is lost: the customer had already installed a cheaper gate from a competitor. Two are being negotiated.

What’s next

• ·Automate the first contact: response time, a task for the next contact, follow-ups without the owner. The goal is lead-to-survey conversion from 16% to 22%.
• ·A night bot in messengers for tagged deals, a voice bot for missed calls (the prototype is ready) and a customer bot on the knowledge base.
• ·Early booking for the 2027 season from the base of lost leads and past customers — for the 189 free crew-days; a recalculation of fences with sliding gates.
• ·Management reports from 1C: an audit of counterparties and balances, then profit and loss, cash flow and monthly balances.
• ·A measurement in three to four weeks against the September baseline: “Avito chat → survey” conversion (8%), “they will get back to us” losses (28 of 35), night leads handled by the AI manager, cost per contact, contracts and their total at the same budget (39 contracts, ₽7.5M).
• ·The 2027 goal is 10 crews by June and about ₽30M staying in the company versus ₽15.5M now.

Takeaways for your business

1. 1.AI in a mid-sized business starts not with a chatbot but with the question of who holds the keys to the infrastructure, where the data lives and whether it can be trusted.
2. 2.On dirty data AI gives a convincing and wrong answer: the first estimate showed a loss where the company makes ₽13.7M. Cross-check the sources first, then draw conclusions.
3. 3.Price is the shortest lever on profit if it is calculated from cost: with profit at a tenth of turnover, +2.4% on price gives +24% on profit.
4. 4.Look for the bottleneck in the funnel, not in feelings: “the welders can’t keep up” and “we need a second surveyor” were not confirmed, while 84% of leads were lost before the survey.
5. 5.Changes in live systems only with a copy and a check after every save.