# searchit AI Agent — the memory of your AI

searchit turns your organisation's knowledge into the memory of your AI: files, mailboxes and internal systems, opened up through a knowledge graph and AI enrichment, with every user's permissions. The searchit AI Agent works with it under your sign-in.

<https://searchit-agent.ai/>

## Your AI gets a memory.

searchit turns your organisation's knowledge into the memory of your AI — opened up through a knowledge graph and a search that reads every document with AI. Everyone sees only what they are permitted to see, and your data lives in an installation for your organisation alone — on your premises or in the cloud.

### From memory to co-worker.

This is how the searchit AI Agent works with it: request, tools, result. One click on the run shows every step and further departments.

## Documents become knowledge.

Your memory is already there: searchit reads the documents your organisation already has and recognises who, what and where — the same person, the same company across all sources, with their relationships. The result is a knowledge graph that you and your AI can query. searchit works from your sources: what is in them counts from day one, and every entity carries the permission of its document.

### Try a question

Five records from a demo dataset, on the right the graph searchit built from them. Tap a node or a record.

> “How is the company whose credit of €2,645.47 arrived on 20 Dec 2023 connected to our framework agreements and open invoices?”

Left: five records from the data. Right: the graph. The credit in the bank statement, attached to an email, and the invoice spell the company differently; searchit merges both into one entry with 16 sources. Linked to it: Nordwind as contract partner with a framework agreement, the dunned invoice, the credit as an amount, the customer number and the location. The revenue report is released to Controlling only and does not appear in the answer.

Demo dataset, indexed by searchit: nodes and edges as the system built them. Permissions assumed for this example — dashed grey marks what only Controlling sees. Anyone outside it does not get the revenue report, not even as a gap in the graph.

The graph is open to you: you see every connection, name it in your own words, merge spellings and add what your department knows.

## Let your knowledge work.

The searchit AI Agent works on the same memory as your search and acts with your permissions. It reads and compares documents, writes the result as Word, Excel or PDF and sends it when you say so.

### Continue with the question above

AutoTeile Müller, customer K-10005: two framework contracts, two invoices with dunning letters. Give the agent its next task.

### Compare framework contracts

> “Compare the two framework contracts with AutoTeile Müller and save the differences as a table for me.”

**In the graph:** contracts linked to AutoTeile Müller queried: LV-2023-011 and LV-2025-003

**Read:** both contracts in full text

**Created:** table as an Excel file, saved in your folder

| | LV-2023-011 | LV-2025-003 |
|---|---|---|
| Start | 27 Oct 2023 | 25 Oct 2025 |
| Term | 12 months, extendable by 12 months | 24 months |
| Services | dangerous goods (ADR), import customs clearance, express air freight | food customs clearance, express air freight, toll |
| Payment term | 14 days | 14 days |
| Volume discount | 3 % from €15,000 per month | 5 % from €25,000 per month |

Values from the contracts in the demo dataset, as the agent read them.

### Draft a payment reminder

> “Draft a payment reminder to AutoTeile Müller for the two open invoices, referring to the dunning letters.”

**Read:** open items list: two open invoices

**Read:** both dunning letters with fees and interest

**Created:** draft as a Word document, saved in your folder

**RE-2023-1001:** €29,712.65, dunned on 7 Feb 2023

**RE-2024-1012:** €135,010.26, dunned on 4 Jun 2024

**Total claim per dunning letters:** €168,954.69

A draft in your name. It is sent only when you instruct it.

### Customer sheet to sales

> “Summarise AutoTeile Müller on one page for me and send the PDF to sales.”

**In the graph:** customer, contracts, open invoices and dunning letters queried

**Read:** revenue report by customer

**Created and sent:** customer sheet as PDF, email with attachment to sales

**Role:** customer K-10005, client for transport and logistics

**Contracts:** LV-2023-011 and LV-2025-003

**Open:** €164,722.91 in two invoices, both dunned

**Revenue 2023:** €27,467.00 net, rank 7

The email goes out because the task explicitly says so. Without that instruction it stays a draft.

### His tool board

Sixty tasks in seven areas. Pick an area — the work lies underneath, and next to it you see what he does there in detail.

### Find and prove

He searches everything your organisation has connected and names the document behind every statement.

- **Hybrid search** — Searches by wording and by meaning in one pass — across 100+ connectable sources: file shares, mailboxes, Teams, SharePoint, line-of-business databases, web sources.
- **Matching passages** — Only the relevant sections with their highlighting, when the whole document is not needed.
- **Breakdowns** — Counts and groups result sets — by year, folder, supplier, tag.
- **Filter by amount and period** — Narrows down to values pulled out of the documents — invoice amount, term, quantity.
- **Browse folders** — Moves through the repository like a file system, with your permissions.
- **Resolve names** — Turns “Müller GmbH” into the one company in the knowledge graph — including spelling variants.
- **Neighbours of an entity** — Shows what a person, company or project is connected to — with the supporting documents.
- **Connecting paths** — Finds the path between two entities: which intermediate steps link them.
- **Proof required** — Names the document behind every statement — and says so when he found nothing.

### Read and prepare

He reads what he has found and turns it into a result you can work with.

- **Read a document** — Fetches a document with details, attachments and full text.
- **Resolve the source** — Finds the retrievable address — SharePoint, OneDrive, Exchange attachment.
- **Fetch an attachment** — Recognises from the mail hit which attachments belong to it and loads them specifically.
- **Fetch the whole mail** — Pulls the whole mail as a file straight from the Exchange server.
- **Save as ZIP** — Exports several hits together, flat or with folder structure.
- **Redact** — Makes personal passages unreadable in PDF, Office and text — through the local service, never by script.
- **Write Word** — Creates, reads and changes Word documents.
- **Write Excel** — Creates spreadsheets with formulas, reads and changes existing ones.
- **Write PowerPoint** — Creates and changes slides including notes.
- **Create PDF** — Creates, reads and changes PDF files, fills in forms.
- **Interpret images** — Reads a photographed receipt or a screenshot with an error message.
- **Summary instead of the whole file** — Every document carries an AI summary in a selectable length together with the entities found — he reads that instead of fetching the whole file.

### Organise and tag

He brings structure into the repository — across whole result sets at once.

- **Create a tag** — Creates a tag or changes its name, visibility and colour.
- **Tag result sets** — Adds or removes tags across a whole result set — inverted as well: all but these.
- **Remove a tag** — Deletes a tag together with its assignments.
- **List tags** — Shows which tags exist and who can see them.
- **Merge entities** — Spelling variants of the same person or company belong to one entry. The proposals arise on their own; they are confirmed in curation — not by the agent.

### Mailbox, appointments, contacts

He works in your mailbox and calendar — under your sign-in, with your permissions.

- **Write and send mail** — Composes the mail with attachments and sends it through your mail server.
- **List mailboxes** — Shows which mailboxes he is allowed to access.
- **List folders** — Shows the folder structure of a mailbox.
- **Create a folder** — Creates a new mailbox folder.
- **Move mails** — Moves the mails he found — “all invoices from supplier X from 2025 to Accounting/2025”.
- **Mark** — Sets read and unread, follow-up and categories.
- **Move to the bin** — Tidies up without deleting for good.
- **Create and change appointments** — Creates appointments, moves and cancels them — in shared calendars too.
- **Maintain contacts** — Creates contacts and changes existing ones.

### Reachable and remote-controlled

You reach him while you are out and give him instructions there — the same agent, the same session as at your desk.

- **Signal** — Instruction goes in as a message, the answer comes back in the same channel. Remote control, not just notification.
- **WhatsApp** — The same in the WhatsApp chat — with a source for every statement.
- **SMS** — Text message to a phone number.
- **Chat on your website** — A dedicated element embeds a chat bubble — its own access and its own profile per embedding.
- **Module in the searchit client** — Embedded at your desk; the session survives navigating away.
- **Notification at your desk** — He reports in when an instruction is finished or a question is pending.
- **One session across all channels** — What you discussed in the messenger, he knows at your desk.

### At the desk and in the browser

He operates your computer and every web application — including where there is no interface.

- **Open a file** — Opens a document or a mail in its default application.
- **Show in the explorer** — Shows the storage location in File Explorer or Finder.
- **Put on the clipboard** — Puts the file itself there, not only its path.
- **Read and write files** — Works in the folders and drives you have access to.
- **Search files** — Searches directories on the computer by name and content.
- **Run commands** — Runs commands and short programs — under your sign-in.
- **Operate the screen** — Clicks and types where there is no other interface.
- **Drive the browser** — Its own visible window with your signed-in session: open, read, fill in fields, submit, download. Marked in red, every address recorded.

### Stay on it and learn

He remembers how you work, keeps to schedules and takes on procedures of your own.

- **Memory at your desk** — Remembers preferred ways of working, recurring constraints, stated preferences — as a readable file on your machine.
- **Find earlier instructions** — Searches his own conversations in full text: “what did we have on this customer last week?”
- **Scheduled instructions** — The new tenders every Monday, the mailbox review every day — create, pause, trigger at once.
- **Split large jobs** — Hands sub-tasks to subordinate agents with their own working context.
- **24 ready-made procedures** — Brings work instructions in seven groups — documents, research, search and graph, contacts, mail, messenger, media.
- **Take on your procedures** — A recurring procedure is stored once — from then on he can do it.
- **Asks instead of guessing** — When something is ambiguous he asks back instead of inventing an answer.
- **Visible task list** — Multi-step work is out in the open — you see where he is.
- **Weekly planning** — A look back at what was promised and what is left, a plan for next week.
- **Tasks from minutes** — Pulls evidenced tasks and deadlines out of documents and meeting notes.

Every task runs under your sign-in and with your permissions.

### What always applies

- **Under your sign-in** — He sees and does only what the person who instructs him is allowed to.
- **Everything on record** — Every tool call and every browser step is recorded with user, time and result.
- **Enabled individually** — You decide which tool group a role receives — and block what is not needed.
- **Credentials stay out** — Password, one-time code and PIN fields are removed before any text reaches the language model.
- **Openly connected** — You connect the AI assistant you approve, or your own application, through the open Model Context Protocol (MCP) — with the role you assign for it.
- **AI output is marked** — Every field an AI produced carries model, version, time and prompt identifier — machine-readable, in exports and through the open interface as well.
- **Where it runs is your choice** — At the workstation on Windows, Linux or macOS; centrally on a Windows terminal server with its own agent per logged-in session; or in the cloud on a virtual machine of its own. On every path it acts as the logged-in user.

## Search yourself whenever you like.

Your staff ask for themselves and see the answer at once: the document in the browser, the passage highlighted, the company behind it right beside it. The same memory and the same permissions as the agent, only with their hand on the wheel. Some questions are asked faster than they are delegated.

- Result list and preview — the document in the browser, without a download.
- Summary, signals and concepts for a contract — and below them, which model produced them.
- People, companies, places and amounts from exactly this document — with their relations.
- The entity workspace: every company in your holdings, each with its evidence.
- Images are searched too — the AI describes what can be seen on them.
- Tagging in one go — one tag onto the whole result set.
- Reports on what has been indexed — per source, per index, over time.

- **See the document at once** — View documents, spreadsheets, images and e-mails in the browser — with highlighted passages, or open them in the original program with one click.
- **Jump to the company behind it** — People, companies, places and amounts from exactly this document. One click filters the search to everything they appear in.
- **Browse through all sources** — Folder tree and path as in a file manager — across drives, SharePoint and mailboxes.
- **Take whole result sets with you** — Bookmark, tag, export as ZIP or spreadsheet — also for many results at once.

## Hand off work.

The biggest item is rarely search time. It is work that people do by hand today because no business system holds the necessary facts in one place:

- Collecting documents for a case from several systems, checking and filing them.
- Producing the same report again every week because nobody has stored it.
- Retyping data from a web application that has no interface.
- Evidencing an access request with its deadline by hand.

An agent takes over such work — under the employee's sign-in, with the employee's permissions — and does several things at once. A measured case shows what that looks like.

### A bid for a public tender

Measured in a customer project, 2026

Sales manager, construction company with around 200 employees

Review criteria and standards, evaluate additional documents, prepare the bid, adopt clauses and bring in subcontractors by email.

Sources used

- existing bids from the document library, found via the graph search
- service and product catalogues
- general terms and conditions
- clause and contract components compared

**Effort before:** around one week

**Effort with searchit:** around one person-day

Measured in this company, with these sources — not an average for every tender. What it looks like for you is shown by a run on your own data.

## Prove what the AI wrote.

The transparency obligations of the EU AI Regulation (EU AI Act) have applied since 2 August 2026. searchit is the provider of the system, you are the deployer — and you get the evidence a deployer needs without building it yourself.

- **Origin on every AI field** — Summaries, entities and signals carry model, model version, time and prompt identifier — separately per enrichment run, so several runs on the same document stay distinguishable.
- **Stays with the document** — The origin survives being passed on: in the CSV output it appears as its own column, AI columns visibly carry the suffix “(AI generated)”, and it travels through the open interface too.
- **The AI says that it is one** — Chat and quick answer disclose at the first interaction that an AI system is replying. In the chat bubble on your own website that disclosure is fixed in the component and cannot be configured away.

## These organisations use searchit.

Allianz · Samsung · Deutsche Leasing Austria · Österreichische Gesundheitskasse · MANZ · Universität Hamburg · Technisches Museum Wien · TÜV Austria · Landesamt für Digitalisierung, Breitband und Vermessung · Stadt Krems · Nationalpark Donau-Auen · TSA Traktionssysteme Austria · PARK Wirtschaftsstrafrecht · Dittrich Software · SATZ Software & Consulting · Dr. Spang · WFB Wirtschaftsförderung Bremen · Regens Wagner

### What they use searchit for — four examples

- Allianz: full-text search in the intranet across databases, file systems, websites and Active Directory.
- Österreichische Gesundheitskasse (Austrian Health Insurance Fund): search in the intranet, started at the Lower Austrian regional health insurance fund.
- Deutsche Leasing Austria: search with text recognition in scans and images and in the Exchange mailboxes.
- Samsung: searchit as an appliance with the data protection module, collecting documents spread across six machines.

## See every item up front.

Four points make your bill predictable.

- **Pay by data volume** — The price follows the size of your data. You pay per workstation only for the agent — from 50 seats it gets cheaper.
- **Separate software and model** — The agent is charged per workplace — regardless of how you obtain and pay for the language model. Both appear separately in the price calculator.
- **Search without a meter** — Within the booked tier your staff search without limit — no fee per query and no consumption points.
- **Cancel monthly** — The monthly subscription can be cancelled monthly. If you commit for a year, the annual subscription gives you a lower monthly price.

### Rather try it first?

A trial on a section of your data is possible — one department, one project, one document type. Which section fits and what it costs, we settle in the demonstration.

### Let us show you.

In around 45 minutes we demonstrate searchit and answer your questions. Tell us beforehand which question the appointment should answer.

- Name

- Company

- Phone (optional)

- Preferred date (optional)

- Which question should the session answer? (optional)

- The session takes around 45 minutes and is held online.
- A human from sales gets back to you, usually on the next working day.
- We use your details to arrange the appointment and for nothing else.
- The button opens your e-mail program with the message ready — it is only sent once you send it.

- Your sender address is in the mail your program opens — you do not have to enter it here.

---

- Price calculator: <https://searchit-agent.ai/pricing/>
- Model prices: <https://searchit-agent.ai/pricing/modellpreise/>

This version is generated from the page's language files (scripts/gen-agent-textfassung.py). The page itself is authoritative.
