AI assistance: suggestions for the tedious parts, decisions stay with you

Where exactly does Santix use AI?

There is no chat screen. AI prepares a draft at specific steps of existing screens, returns structured fields and shows the result labelled "Yapay zekâ önerisi" (AI suggestion). The app interface is Turkish; screen names are given in Turkish with a translation.

How are drawing layers classified?

When you upload a DWG or DXF, Santix's CAD server first matches layer names against a rule dictionary: Turkish office names (DUVAR, KAPI, MAHAL, KOT), English and AIA names (A-WALL, A-DOOR), ISO 13567 patterns and the layer list of Türkiye's August 2026 regulation on digital project drawings. Only layers the rules cannot resolve go to the AI.

The drawing itself is not sent. For each layer the request carries the layer name, entity counts (lines, arcs, blocks, text, hatches, dimensions), block names and a few sample texts. The answer gives each layer one of 14 meanings (wall, door, window, room tag, room boundary, fixture, level text, finish text, dimension, grid, floor hatch, furniture, services, or ignore) with a confidence.

You see the result in the layers step of Rooms › Review. Each row is marked as rule, AI suggestion or confirmed, with uncertain ones at the top. Fix individual rows or confirm the rest in one tap. Corrections apply to later uploads in the same project. PDF sheets have no layers, so the step is skipped.

How is the architect's room schedule matched to rooms?

Load the room schedule as Excel or CSV, or paste the table text from a PDF, in the finishes step. Santix first reads it on the device: if the header has columns such as room number, floor, wall and ceiling, matching happens without AI at all.

If headers are not recognised, "Yapay zekâ ile oku" sends the room codes and names from the plan together with the table rows (up to 5,000 rows or 200,000 characters of text). The suggestion holds, per room, floor finish and thickness, skirting, wall finish (per direction if the source splits it), ceiling and ceiling type, waterproofing, thermal and acoustic insulation, paint code and notes.

Matching tolerates spelling differences: "Z-01", "Z01" and "z.01" are the same room. Rows covering several rooms, such as "all wet rooms", become one suggestion per room with lower confidence. Codes found in the table but not on the plan are listed separately rather than dropped. Before applying, you choose the rooms and either "fill blanks" or "overwrite"; applied values carry the architect source badge. See room data sheets and the room schedule guide.

How is an Excel estimate imported?

1. Choose import from Excel in the budget, or import items in a purchase request, and pick the file.
2. Santix sends the header row and at most 20 sample rows, never the whole file, and finds the real header if title rows sit above it.
3. For the budget it proposes item code, category, description, unit, quantity, unit price and total; for purchasing, description, unit, quantity and unit price.
4. Warnings appear on screen, for example when it is unclear whether prices include VAT or when material and labour prices are in separate columns.
5. Correct the mapping, check the preview and import. No row is added without your confirmation.

Offline, or with AI unavailable, columns are guessed from header names and marked as such. More on progress payments and budget.

What is read from a scanned document, and what is not?

In Documents, scan with the camera or pick a photo or file. The image is read and the form is filled with type, title, document number, issuer, issue date and expiry date. Fields that were hard to read are underlined in orange, and a banner says how many to check.

Santix checks the values too: impossible dates, an expiry before the issue date and tax numbers that are not 10 or 11 digits are flagged or rejected. If the document has no expiry date, one is suggested from the statutory validity of its type (24 months for an OHS risk assessment, 6 months for a scaffold inspection report) and flagged for review.

Invoices and delivery notes are stored as documents with type, number, issuer and date. The amount is not carried into the budget or a progress payment; you enter payments in the budget yourself. If reading fails or there is no connection, you fill in the form by hand and the file is kept. See document tracking.

What does AI do when drafting a contract?

Instead of answering questions one by one, you can describe the deal in a sentence, such as "2.4 million TL, 10 % advance, starting 1 December and finishing end of May". The helper extracts the template variables from it, lists them as proposed values and you tick which ones go into the draft. After each answer it asks at most three more questions, required fields first, then payment, retention, time and delay penalties, health and safety, insurance and tax.

It also flags risks: no retention or bond, a delay penalty without an overall cap, multi-year work that triggers the 5 % withholding, unassigned stamp duty (9.48 per mille) or VAT withholding (4/10), pay-when-paid terms or an unsecured advance, and warranty periods that are short for the trade. For custom clauses you describe what you want and get one to five formal clause drafts; unknown amounts are left as placeholders, never invented. Nothing enters the contract until you add it. See contract templates.

How is the bid recommendation prepared?

Scores are not computed by AI: Santix calculates the weighted score itself (price 60 % by default, lowest price scores 100). "Öneri iste" sends the computed comparison, meaning totals excluding VAT, duration, payment term, score, inclusions, exclusions and notes, and returns a 3–5 sentence recommendation with reasons, up to six risks and up to six negotiation points. It is labelled as an AI suggestion; selecting a bid stays with you. Comparison and scoring work offline; only the recommendation needs a connection. See bid comparison.

What data is sent, and where?

Requests go to OpenAI through Santix's server on Cloudflare, not directly from the app. The server checks your session, sends only the content the task needs, asks OpenAI not to store it (store=false) and sends no user id. Santix keeps no content, only a usage counter of organisation, feature and units processed.

Sub-processors and retention are listed under security and privacy.

Limits: what AI does not do

• It never creates, changes or deletes a record on its own; every suggestion passes an approval screen.
• It does not pick bids, approve payments or answer revision requests.
• It does not compute room outlines or areas; the CAD server derives those from geometry.
• Web price research is switched off for now.
• It needs a connection; without one, the same work is done by hand.
• There is a monthly usage limit per organisation and a per-minute request limit per person, with a clear message when either is reached.
• Suggestions can be wrong, which is why low-confidence fields are flagged and nothing is applied silently.