AI helps businesses find wasted time, handle inquiries, prepare materials, and use accumulated knowledge. But a subscription to a neural network does not automatically increase profit. To evaluate the impact, you need to compare the full cost of the process and the quality of the result before and after implementation, then check whether the time saved turned into money or additional productivity.
This article includes seven scenarios from the provided transcript of the “Founders” episode and five additional public case studies. They cover construction, sales, marketing, the floral business, manufacturing, and services. At the end, you’ll find a pilot plan, a payback calculation, and a ready-made prompt for choosing the first task.
How to read the numbers. The transcript data are statements made by the speakers, not financial statements verified by the editors. The five additional case studies are based on publications from technology providers Microsoft and AWS: these are company and partner statements, not independent controlled studies. Project conditions differ, so repeating the results is not guaranteed. The episode featuring the entrepreneurs is listed in the Founders channel catalog; the transcript was provided to the editors.
*Material prepared on October 31, 2026.*
Where AI Brings Value to Business
| Function | What Can Be Handed Off to the System | What Remains with the Human | How to Measure It |
|---|---|---|---|
| Sales | Lead qualification, conversation analysis, reminders | Negotiations, exceptions, final agreements | Response time, conversion by stage, errors |
| Marketing | Drafts, adaptation, creative variations | Fact-checking, hypothesis selection, brand | Customer acquisition cost, share of verified materials, test results |
| Operations | Document matching and application preparation | Review, approval, handling exceptions | Time per operation, errors, rework volume |
| Knowledge | Search across instructions and documents | Updating and source verification | Search time, share of answers with the correct source |
| Production | Defect detection and guidance for specialists | Deviation confirmation and resolution | Missed defects, false alarms, losses |
The role of AI is to speed up a specific part of the process. If the margin is negative, the product is not wanted by customers, or the team is not delivering on its promises, speeding up those actions may increase losses.
1. Good Wood: Visualizations and Control of Customer Promises
Alexander Dubovenko, founder of the construction company Good Wood, said in the provided transcript that the company used to spend about 500,000 to 1 million rubles per month on 3D visualizations and produced about ten sets. According to him, after adopting generative tools, company employees began creating hundreds of images per month, and the total spending on subscriptions to various neural networks was about $1,000 per month.
This is not a calculation of net savings: the volume and mix of work changed, and employee labor, review, and integration were not disclosed separately. A generated image is also not the same as a technical design: it still has to be checked against dimensions, materials, and agreements.
In sales, according to the entrepreneur, the system analyzes communications, helps spot violations, and reminds the team about promises made to the customer. A practical idea for a small company is to start with one control point: after a call, automatically generate a list of promises, dates, and responsible owners, then have the manager confirm accuracy before it is entered into the CRM.
2. Floral Business: Orders, Data, and Plant Quality Control
Ilya Balakhnin of Paper Planes described in the transcript a project for an unnamed company that handles wholesale and retail flower sales as well as landscaping. According to him, the solution package processed standard requests, helped normalize data, selected plants from a knowledge base, and identified possible signs of disease in images. The consulting representative estimated staff time savings at roughly one-fifth, but did not specify the calculation base or time period.
For a flower shop, it is better to start simple: gather inventory, purchase prices, and sales in a single spreadsheet; then ask the system to prepare a draft purchase order and compare it with the florist’s decision. A diagnosis based on a photo should not be treated as definitive: a questionable plant should be inspected by a specialist. Before automating ordering, you need data on write-offs, seasonality, and shelf life.
3. Sales: Lead Qualification and Training on Calls
The host of the episode said that in one of his companies, two managers handle more than 1,000 qualified leads per month with a sales cycle of about a month and a half; previously, he said, a team of ten people handled that work. He also reported about a 30% increase in conversion after implementation, but the starting point, observation window, and comparability of the pipelines were not disclosed. It is not possible to conclude that AI caused all of the growth, or to convert that figure into percentage points.
Based on the description, the front line clarifies the need and passes the data to the manager, while reviewing transcripts helps create playbooks and train sales. A pilot can be limited to one lead source and should specify in advance when the agent must hand the conversation over to a person: a complaint, a discount request, uncertainty about an answer, or a request to speak with a manager.
Compare not only the number of leads processed, but also the share of target leads, conversion to payment, errors, and employee time spent on corrections. If there are more leads but their quality has changed, a before-and-after comparison without adjustments can be misleading.
4. Content and Advertising: More Options Without an Automatic Sales Guarantee
Roman Kumar Vyas said in the transcript that content production and creative analysis were automated. According to him, the share of winning ad variations increased from about 20% to 40%, and creating a banner through a neural network cost about 10 cents. These figures refer to a specific testing system and do not disclose the full budget, the definition of a winner, or the cost of the team’s work. The cost of generating an image cannot be equated with the full cost of the campaign.
A verifiable setup for an entrepreneur: take one advertising hypothesis, prepare several image and copy variations, check them for accuracy and rights, then run a comparable test with a preselected metric. Higher click-through rates are useful only if they do not worsen the cost of a qualified lead or the economics of the deal.
The host also described an editorial bot that was given past edits and examples of successful publications. He said the bot cost about five dollars, comparing it with an editor’s salary of more than 100,000 rubles. This is a comparison of individual costs, not the full cost of replacing the work: material preparation, editing, technical support, and responsibility for publication remain with the team.
5. Development Planning: Which Ideas Should Go Through Teams First
Balahnin gave the example of iRidium, a company developing smart home systems. According to his description, the planner helps assess the economic value of ideas and choose the order of work for multiple teams. He claimed a 11–12% increase in gross profit margin per quarter. The transcript does not clarify whether this refers to a relative change or percentage points, and the calculation method was not disclosed.
How to use this in a small business: list initiatives, estimate expected demand, gross profit, labor costs, dependencies, and uncertainty. The system can prepare sequencing options, but the owner must verify the original assumptions and account for customer commitments. A forecast without quality source data does not become a fact.
6. Food Manufacturing: Technologist Knowledge and Searchable Records
In another example from the transcript, Balahnin spoke about an anonymized company in food ingredients. According to him, the team digitized the technologist’s lab records from 18 years and connected them to a knowledge base search. Combined with analysis of new developments, this reportedly reduced the path from identifying an ingredient to a validated sample from 6–8 months to 6–7 weeks. The sequence of steps and independent verification of the result were not disclosed.
This is where RAG is useful: the system first searches for relevant passages in documents and then generates an answer based on what it finds. For a pilot, one folder of approved instructions or an archive of experiments is enough. Check that records are recognized correctly and require links to the specific document. The model’s output is a hypothesis for the technologist, not a replacement for lab testing or product safety requirements.
7. Founder Content: One Source for Multiple Channels
Dmitry Beskromny said in the transcript that he uses editorial materials written for his Telegram channel as the basis for adapting content into other formats. According to him, after resuming work on the account, he generated more than 2 million impressions in about a month. There is no data on paid promotion, unique reach, the composition of impressions, or sales, so the figure does not prove ROI.
To repeat the approach, choose one high-quality source piece and turn it into several drafts: a short post, customer Q&A responses, and a video script. The editor checks the meaning, timeliness, and platform requirements; separately account for legal status and channel availability for your audience. Track clicks and leads, not just views.
Five More AI Implementation Cases from Public Sources
8. Rau Consultants: Candidate Profile Drafts
The German staffing company Rau Consultants used Microsoft 365 Copilot to prepare candidate profiles. According to Microsoft's description, consultants created a prompt template, receive a draft, and then finish it themselves. The source reports time savings, but the amount cannot be reliably compared with other projects here.
What to take away: choose a document with a repeating structure, prepare an anonymized example, and ask AI to assemble a draft using approved fields. Verify facts against the source documents, and process personal data only when there is an appropriate legal basis and access controls are in place.
9. Acentra Health: Drafting Letters Based on Inquiries
Acentra Health created MedScribe to draft responses to inquiries related to healthcare services. In Microsoft’s December 2024 case study , it says the average time to prepare a letter fell from about six minutes to three minutes, and by mid-2024 nurses had saved 11,000 hours of work. Staff reviewed the letters; the source material and data protection requirements mattered.
What to take away: separate document preparation from approval. Measure not only the time for the first draft, but also review time, the correction rate, and the impact of errors. Medical decisions cannot be delegated to a generative model without professional oversight.
10. Softchoice: Customer Materials and Meeting Summaries
According to a 2024 Microsoft publication, employees at technology company Softchoice used Copilot for meetings and content work. For sales and marketing teams, the company reported a 62–67% reduction in the time needed to create customer materials. This is the result of company-described testing, not a forecast for every organization.
What to take away: record the time it takes to prepare ten comparable materials the usual way and with an assistant. Check quote accuracy and the quality of meeting summaries; agree on recording and processing conversations with participants under the applicable rules.
11. Presidio: Employee Training and Cross-Functional Rollout
According to the December 2024 Microsoft case study, Presidio first rolled out Copilot to 300 employees across different departments, developed training materials, and created a space to share insights. The company estimated total user savings at about 1,200 hours per month. The source does not provide enough detail to treat this as guaranteed payroll savings.
What to take away: assign a pilot owner, give employees examples of safe tasks, and collect not only successful prompts but also mistakes. Training and support are part of implementation costs too.
12. Oriserve: Voice and Text Support Agents
Indian developer Oriserve describes in an AWS case study the use of generative agents in customer conversations. According to the publication, the company cut its own AI application support costs by 50% after moving to the described infrastructure. This is a vendor-reported metric, not the cost of any customer’s contact center; AWS and Oriserve have an interest in promoting their experience.
What to take away: first assign the agent a limited set of questions allowed by the knowledge base and provide handoff to a human. For voice channels, separately measure speech recognition, the share of incorrect answers, customer dissatisfaction, and cost per minute including repeat contacts.
How a Small Business Can Start with One Process
Imagine a flower shop with one florist. The owner orders flowers personally, and inventory and sales are tracked in a spreadsheet. The goal is to reduce spoilage and avoid losing sales because popular items are out of stock. This is a training example, not a description of a real project.
- Take the baseline. Collect sales, purchases, inventory, and spoilage for comparable periods. Mark holidays and price changes.
- Define the solution. For example: once a week, suggest purchase volume by flower type based on inventory, shelf life, and deliveries.
- Start with a draft. AI receives the prepared table and explains the calculation. The florist checks every item; the order does not go out automatically.
- Check the exceptions. New products, weddings, holidays, and incomplete inventory can make past data a poor guide.
- Compare the result. Look at spoilage in rubles and as a percentage of purchases, stockouts, gross profit, and time spent on purchasing. Account for seasonality and promotions separately.
- Decide whether to expand the pilot. If data quality is low or corrections erase the time savings, fix the recordkeeping first.
This approach works for both a service company and a small wholesale firm: the documents and metrics change, but the principle remains the same — data → draft → review → measurement.
How to Calculate AI ROI
Compare the same amount and quality of work over the same period. Include subscriptions and model calls, integration, storage, support, result review, and training in the costs of the new setup. Count one-time implementation costs separately.
Monthly impact = verified cost reduction + additional gross profit − monthly solution costs.
Simple payback period = one-time implementation costs / monthly impactif the impact is positive and relatively stable. This is a simplified calculation without the cost of capital or changes in volume.
For example, a company used to pay a contractor 80,000 ₽ per month. After implementation, the same amount of work costs 20,000 ₽ for the service and 25,000 ₽ for review and support. With one-time costs of 140,000 ₽, monthly savings would be 35,000 ₽, and the simple payback period would be four months: 140,000 / 35,000. This is a hypothetical example, not a promise of results.
Important distinction: 100 freed-up hours do not automatically equal cash savings. Savings happen if the company actually reduced paid expenses or used that time for additional work that generated measurable gross profit. Additional revenue by itself is also not the same as profit.
Data, errors, and control: what to plan before launch
- Access to data. Before uploading contracts, call recordings, and customer databases, determine which data is actually needed, who gets access, and how long it will be retained. Check the applicable personal data requirements and the vendor’s contract terms; anonymization requires review and is not always as simple as removing names.
- Processing environment. On-premises deployment helps control infrastructure, but it does not by itself guarantee security: access control, updates, backups, and integration testing are still needed. Cloud products also differ in configuration and terms.
- Model errors. Ask for document-based answers, test the responses on a validation sample, and measure errors. The model can confidently produce an incorrect answer even when a knowledge base is available.
- Permission to act. In the pilot, restrict automatic sending of emails, payments, price changes, and orders. For external actions, assign a responsible person and an approval process.
- Business continuity. Plan a manual fallback in case of service failure, loss of access, or rising costs. Record versions of instructions and decisions so you can trace the error later.
For more on searching internal documents, see the breakdown of a corporate knowledge base with AI.
Ready-to-use prompt for choosing your first AI task
Copy the template and fill it in with safe information that does not include unnecessary personal data:
You are a small business process consultant. We work in [industry], and our team has [number] people. Our recurring tasks are: [list]. Available data: [types of documents and spreadsheets]. Constraints: [budget, security requirements, allowed tools]. Suggest five AI pilot candidates. For each one, specify the current process, expected outcome, required data, risks, human role, and before/after metrics. Do not invent cost or savings: ask questions if the data is insufficient. Then help choose one process with a clear outcome and a simple rollback path.
The neural network’s answer is a list of hypotheses. Review it with the employees who do the work and choose the task based on the real cost of mistakes and the readiness of the data.
Common questions about AI for business
How much does it cost to implement artificial intelligence?
There is no single price. The final cost depends on the complexity of the process, licenses, integrations, data requirements, and the amount of review needed. Calculate one-time setup and ongoing expenses separately. The example above shows the method, not a market rate.
Which process should you choose first?
A repetitive one with limited volume, available data, and a clear quality criterion. A good first candidate is drafting a response or searching an approved procedure, if an employee can review the result before sending it.
Do you need to lay off employees after implementing AI?
There is no mandatory rule. Some of the freed-up time can shift to customer work, quality control, or product development. Decisions about team structure are made after measuring the results and taking labor obligations into account.
How is an AI agent different from a regular chatbot?
In this article, an agent means a software workflow in which the model uses data and permitted tools to complete several steps. Its autonomy is limited by settings: human review, access rights, and restrictions on risky actions are defined separately.
Can AI guarantee sales growth?
No. Results depend on demand, the product, lead quality, pricing, team performance, and the measurement method. A comparable pilot with preselected metrics and change control helps assess AI’s contribution.
How Vladimir Lomtev’s AI Studio helps implement AI into business processes
On the Vladimir Lomtev AI Studio website are solutions for the kinds of tasks covered in this article: process audits and setup, AI agents integrated with business systems, corporate knowledge bases, and document processing automation. For a pilot, you can choose one process, define baseline metrics, data sources, constraints, and acceptance criteria. Then the team checks the solution’s quality and decides whether to expand its use.
Bottom line: start with a measurable task
The case studies point to several areas worth testing: document preparation, lead qualification, call analysis, marketing drafts, and knowledge search. Company claims are useful as a source of ideas, but they do not replace your own experiment. Choose one area, keep human oversight, and evaluate the full cost together with the quality of the result.