Four recurring tasks of a real estate agency
An agent's work consists of many repetitive small jobs. Each has a different cost of error.
- A listing description. A form (area, floor, condition, surroundings) becomes the text of an advert. The error here is a fact that is not in the form.
- A first reply to a lead. Someone writes 'I am interested in the flat on Lisova'. You need to answer quickly, ask about budget and timing and pass it to an agent.
- Price comparison. Pick comparables and justify the price for the owner. The error here is the client's money.
- Photos. Editing, captions, 'improvements'. The error here is no longer lost time but a risk of misleading the buyer.
A model does the first two well when it has the form and the rules. It does the third only on the data you gave it. The fourth has a legal line, covered below. General rules for phrasing requests are in the article on prompts for a business task.
What it costs per month
A hypothetical agency: 40 new listings and 600 incoming leads a month. A description takes 1,500 input tokens (the form and style rules) and 700 output. A first reply to a lead takes 1,500 input tokens (the object card, the client's message, the rules) and 300 output. The volumes are hypothetical; Ukrainian tokenises less economically than English, so recalculate on your own texts.
Prices of three current models from the OpenAI API pricing page on 21 October 2026 (standard processing, short context, per million tokens): gpt-6-luna is $0.10 input and $0.50 output, gpt-6.1-sol $2 and $10, gpt-6-astra $10 and $50. The National Bank rate on 21 October 2026 is UAH 45.0791 per dollar.
| Model | One description | One lead reply | 40 descriptions and 600 replies a month |
|---|---|---|---|
| gpt-6-luna | $0.0005 (UAH 0.02) | $0.0003 (UAH 0.014) | $0.20 (UAH 9) |
| gpt-6.1-sol | $0.01 (UAH 0.45) | $0.006 (UAH 0.27) | $4.00 (UAH 180) |
| gpt-6-astra | $0.05 (UAH 2.25) | $0.03 (UAH 1.35) | $20.00 (UAH 902) |
Calculation: (40 × (1,500 × input price + 700 × output price) + 600 × (1,500 × input price + 300 × output price)) ÷ 1,000,000 × 45.0791. Volumes are hypothetical; OpenAI API prices and NBU rate on 21.10.2026.
Even the highest sum is small next to the commission on one deal. So do not choose a model by price: compare them on twenty of your past descriptions and leads, and pick the one that invents fewer details. The sum excludes the cost of an intermediary app (a website chat, a CRM integration): take that from the tariffs of the platform you choose.
Price comparison: median, not average
An owner asks: 'why do you set exactly this price?'. The agent picks comparables, similar flats, and divides the price by the area. A common slip here is to take the average. One renovated unit among six ordinary ones pulls the average up.
A hypothetical example. A 54 m² flat. Seven comparables in the same district, price per square metre in dollars:
| Metric | Calculation | Result |
|---|---|---|
| Prices per m² (ascending) | 1,080, 1,150, 1,190, 1,230, 1,260, 1,310, 1,720 | 7 comparables |
| Average | sum ÷ 7 | $1,277/m² |
| Median | the fourth value | $1,230/m² |
| Flat price by average | 1,277 × 54 | $68,966 |
| Flat price by median | 1,230 × 54 | $66,420 |
| Difference | 68,966 − 66,420 | $2,546 |
| Spread of comparables | (1,720 − 1,080) ÷ 1,230 | 52% of the median |
There is one object at $1,720/m² here, and the average absorbed its influence: the owner would get a guide $2,546 higher than a typical comparable shows. Without that object the median would be 1,210 and the average 1,203: an outlier lifts the average and barely touches the median.
How to split the work between model and spreadsheet:
- The spreadsheet calculates. The median, minimum and maximum come from a spreadsheet formula, not from the model. A formula's result can be checked, and an error in price costs the client money.
- The model reads. Give it the table of comparables with a note on condition (renovated, not) and ask it to flag objects that clearly differ and say briefly why.
- The agent decides. Whether to keep the renovated unit in the sample is for the person who has seen the flat.
- A minimum number of comparables. With fewer than five, the median is unreliable; say so to the owner. The threshold of five is our rule, not a legal norm.
Do not ask a model to name the market price without data: it will answer confidently, but without a source, and you cannot verify it. Supply your own price data (for example from adverts).
Photos and descriptions: the line of truth
AI can 'furnish' an empty flat, swap a grey window view for a sunny one, remove cracks. The temptation is understandable and the risk is direct. The Law of Ukraine 'On Advertising', in Article 1, defines advertising as information about a person or a product that is distributed, among other purposes, as self-promotion to form interest in the product. An advert for selling an object fits that definition. The same article defines unfair advertising as advertising that misleads or may mislead its audience through inaccuracy, untruthfulness, ambiguity, exaggeration or concealment. Article 7 lists legality, accuracy and reliability among the principles of advertising.
From this comes our practical line (our reading of the norms, not legal advice):
- Allowed: correcting light and colours, cropping, removing items from the frame that are not part of the object (someone's car, rubbish).
- Not allowed: changing what the buyer is buying: the state of walls, floors, windows, utilities, the view from the window, room size, furniture that will not stay.
- Grey zone: a visualisation with renovation or furniture. A 'visualisation' label is more honest, but whether it is enough for a buyer who looks only at the first picture is a question for a lawyer; we have no conclusion.
In descriptions the same rule applies: the model works only from the form. Add to the request: 'Do not invent facts. If data is missing, write "not specified"'. The agent checks the surroundings (school, metro) personally: a model may invent what is not there. Before publishing, the agent checks every fact of the description against the form.
First-line lead handling
The model's job in the first reply is simple: confirm the object is still available, answer from the object card and ask two or three questions (budget, timing, payment method), then hand over to an agent with a short summary. Every reply must contain only what is on the card. The model does not promise a price, a discount, the legal status or when you can view: those are for the owner and the agent.
Client data — phone, name, budget, sometimes family composition — is personal data. Before you connect a model to correspondence, decide where it is processed, following the article on data for an AI assistant. Do not give the model copies of passports, technical passports with addresses or deal details. On choosing between a chatbot builder, a ready assistant and custom development, see the article on a chatbot for business.
What to give the model and what not
- Is every fact in the text in the object's form?The model must write only from the data you gave. No → complete the form, do not ask it to 'add something nice'
- Does the job change a photo so that the object being bought changes?Condition, look, size, furniture, view. Yes → do not do it, it risks unfair advertising
- Must it name a price, a discount or the legal status?That is a decision for the owner and the agent. Yes → a person only; the model drafts from the table of comparables
- Will passports, deal details or owners' addresses end up in the request?A first reply does not need them. Yes → remove them from the request
A first step for a week
- Take 10 past objects: the form and the description already published.
- Give the forms to two or three models with the same style rules and a 'do not invent' requirement.
- Compare with your descriptions: count facts that are not in the form. That is your measure of invention.
- For leads: take 20 past messages and check whether the model's reply contains only data from the card.
- Start keeping the table of comparables now: a median from your own data will be needed whatever the model.
Our position
An agency should give descriptions and first replies to leads to a model provided it works only from the form and the object card and an agent checks the facts before publishing. Compare prices in a spreadsheet by the median; the model flags outliers and writes explanations, but does not calculate or guess the market. Change photos only so that the object itself does not change.
The position does not hold for commercial property and complex deals, where a form is not enough and an error in the text has legal consequences: there a lawyer checks the descriptions. It also does not hold if you have fewer than five comparables: the median of such a sample is unreliable. The cost calculation assumes short texts and the volumes named above; recalculate for long forms.
Prices and the exchange rate were checked on 21 October 2026 on the OpenAI API pricing page and at the National Bank of Ukraine; the norms were checked against the text of the Law 'On Advertising' on zakon.rada.gov.ua. The arithmetic was recalculated in code. The numbers of listings, leads, tokens and the comparables' prices are hypothetical. This article is not legal advice.