Not AI that knows everything. AI that knows your company.

Public AI has read the internet and sounds confident about everything. A business needs the opposite: an AI that has read your documents, knows your rules, and says so when it does not know. The Back Office is built that way from the first screen: the libraries you approve are its world, the verifier checks every statement against them, and the source is shown beside the answer.

• Your libraries, ranked by you, are the only sources
• Deep reasoning on any file or record you point it at
• “Not in our sources” is an answer — routed to the person who owns the topic

How verification works →

Libraries: what the company approved, with priorities and readers. This is the AI’s world.

Types, fields and names configured your way: a patient, a member, a tenant, a dealer, a subcontractor.

It learns your company — and keeps learning it

Your vocabulary, your record types, your procedures, your templates, your tone. The launch loads them; every correction your people make afterwards is remembered. The third report sounds like your company. The tenth automation knows the exceptions the first one met.

• Industry vocabulary and your own terms
• Persistent organisational memory of rules and preferences
• Zanus maintains the platform; you maintain your rules and approved knowledge

Automation that can read

A supplier emails a revised price list. The flow fires; the AI step reads the list, compares it with the agreed terms on the supplier record, finds two items above the ceiling, drafts the objection with the clause cited, updates the lines that are fine — and puts the objection on Output for purchasing. Exceptions go to a human; routine work keeps moving.

• A visual canvas your process owners can read
• Failures in plain words, retry in one click
• Every run logged with what it read, decided and did

Repeatable work, automated →

Trigger, sources, AI step, actions. Green ran; red is explained.

New AI work: objective, sources, format, schedule. Anyone who can write a request can define it.

Give Zanus the objective, not every instruction

“Compare these three offers against the drawing.” “Every Friday at 16:00, the stock-out report to the buyers.” “List every deadline in these files with the page.” Four steps, and the AI does it — once or on a schedule — verifies it, and delivers it to Output. The people who used to do it read the result.

• One-off analyses on any set of files
• Recurring reports, reviews and reconciliations
• Results as documents, tables or record updates

Give the AI a job →

Output: everything the AI produced, with a preview and a status. Awaiting your decision, or sent under your rule.

Your people decide

Everything the AI produces lands on one page with a preview and its sources. Approve with a click, correct, or let the types you trust go on their own. Permissions say who may approve what; the audit log keeps every decision. Nothing leaves without a decision — yours, or a rule you wrote.

• Approve, edit, reject, reassign
• Auto-send per type, reversible
• Every action — human or AI — on record

Approve before it goes out → · Who sees what →

The procedures are the training. Update one and the other follows.

People change. Your company’s intelligence remains.

When a key manager leaves, the company usually loses how things were done. Here the procedures, the rules, the corrections and the history are in the system: the new person is trained from them by the AI, the assistant answers from them, the automations enforce them. Business continuity is not a binder — it is the Back Office itself.

New staff trained by the AI →

Six steps between a question and an answer you can sign

The verifier is not a filter added at the end — it is how the AI works. RETRIEVE the relevant passages from your approved libraries. BUILD the answer only from those passages. VERIFY each statement against them. CITE the passage next to the statement. ACT only when everything is grounded. ESCALATE to a person when it is not.

• Works the same in chat, in scheduled work, in automations and in email drafts
• Library priorities decide which document wins a conflict
• Section-level citations: document, section, page

The index behind the verifier: every approved page, searchable by meaning, not only by keyword.

The result lands on Output with its status: awaiting your decision, or already sent under a rule you set.

A worked example

A supplier writes asking for payment terms “as agreed last year”. The AI opens the supplier record, finds the 2025 agreement in the contracts library, quotes the clause — net 45, early-payment discount 2% — drafts the reply with the clause cited, and puts it on Output for the person who owns purchasing. If no agreement existed, the draft would say so and ask, instead of proposing terms.

• The clause is quoted, not paraphrased from memory
• The draft names what it could not find
• A person decides before anything leaves

One chat, three kinds of business

A consultant asks for every commitment made to a client across three years of correspondence and gets a dated list with the emails cited. A production manager drops a supplier’s spec sheet and asks whether it meets the drawing’s tolerances. A warehouse lead asks which orders are waiting on the same missing item. Same chat, same verifier, different libraries.

• Reads and reasons on any file you drop in: PDF, spreadsheet, image, audio, video
• Searches by meaning across libraries, records and mail
• Long answers arrive as a document you can keep or send

What the chat knows is what you put in the libraries — and what each user is allowed to see.

Every conversation and every piece of AI work is searchable later — the company stops losing what it already figured out.

Assistants that remember how you work

An assistant is a chat with a job description: which libraries it reads, which skills it uses, how it writes, what it checks each time. It keeps a memory of your corrections, so the third report sounds like your company and not like a first draft. Presets make the common requests one click: “summarise this file for the board”, “compare these two offers”, “draft the reply in our tone”.

• Skills and presets per assistant
• Persistent organisational memory of your preferences and rules
• Assistants shared with a team or kept private

Every kind of file, one library

Contracts and their amendments. Procedures and the training video that explains them. Drawings, spec sheets and the supplier’s certificate. Meeting recordings. The AI reads PDFs and Office files, transcribes audio and video, reads images — and files each into the library you routed it to, ready to be cited.

• PDF, Word, Excel, PowerPoint, images, audio, video
• Routing rules from intake to library
• Versions kept; the current one wins by priority

New files arrive by mail, upload or connector and are routed to the right library with handling instructions.

Generated documents land on Output as paper-style thumbnails: preview, approve, send or file.

From sources to a finished document

Ask for the monthly board summary, the comparison of two supplier offers, the answer to a tender question, the client letter about a contract change. The AI drafts it in your template, cites what it used, and places it on Output. Approve and it goes where you said — to a record, a library, a mailbox.

• Your templates: letterhead, structure, tone
• Sources listed inside the document
• Scheduled documents: the same report every Monday, unattended

What a morning looks like

A client sends a signed contract: filed on the client record, the renewal date set in the calendar, a thank-you drafted. A supplier sends an invoice: matched to the order, the difference flagged, the question drafted. A form arrives with three attachments: each routed to its library with the handling note. A complaint arrives: the record opened, the history summarised, the reply drafted from the policy — and held for a person.

• Read, understood, attached to the right record
• Documents routed with handling instructions
• Replies drafted with the sources cited

By 08:00 the night’s mail has been read, filed and drafted. The board shows what is left for a person.

The client record after the mail was processed: the message, the document, the draft, the date — all in one place.

Nothing lost between inbox and archive

The usual failure is silent: an attachment saved to a desktop, a request in a folder nobody watches. Here the pipeline is a board everyone can see. A failed step is red and retryable; a document routed to the wrong library is moved in one click; every action is on the record and in the audit log.

• Visible states, one-click retry
• Daily send ceilings so no mailbox is suspended
• Failures explained in plain words

Three jobs, one wizard

A law practice asks for every deadline in a folder of new matters, listed with the source page. A manufacturer asks for the three supplier quotes compared line by line against the drawing. A distributor asks for the weekly stock-out report every Friday at 16:00, sent to the buyers. Each is a four-step definition; each result arrives verified.

• One-off analyses on any set of files
• Recurring reports and reviews
• Results as documents, tables or record updates

The works list on a normal day: scheduled reports, one-off analyses, reviews waiting on Output.

Finished work surfaces on the dashboard’s Act Now feed — the human part of the job is a decision, not the work.

A fleet of managers at the keyboard

Scheduled work is where the back office changes shape. The reports that used to consume Monday mornings run at dawn. The reviews that waited for a quiet week happen every week. The people who did them read the result, correct what needs correcting — and the AI remembers the correction next time.

• Runs nights and weekends; the results wait for the morning
• Corrections become memory for the next run
• Every run recorded, searchable and auditable

The supplier email, end to end

A supplier emails a revised price list. The trigger fires. The AI step reads the list, compares it with the agreed terms in the supplier record, finds two items above the contract ceiling, drafts the objection citing the clause, updates the two purchase lines that are fine, and puts the objection on Output for purchasing. Purchasing reads a draft with the clause beside it — and clicks.

• Reads the document, not only the subject line
• Compares against your records and contracts
• Acts on the routine part, escalates the exception

The automation starts where the work arrives: an email with an attachment enters intake and the flow takes it from there.

Work types are the vocabulary of your automations: the kinds of work your company does, named your way.

Owned by the business, not by a developer

The canvas uses your words: your work types, your client types, your libraries. The person who owns the process can read the flow and say “that is not how we do it”. Every run is logged with what it read, decided and did — and the AI step carries the same verifier as everything else.

• Visual, readable, in your vocabulary
• Every run logged with sources and decisions
• Our people build the first flows with you at launch

What a reviewer sees

The draft on the left, the passages it was built from on the right, the recipient and the channel at the top. The reviewer reads, maybe changes a sentence, clicks. Ten reviews take the time one used to. Corrections are remembered, so the same kind of draft needs fewer next month.

• Preview, sources and destination on one screen
• Edit in place before approving
• Corrections feed the next run

The result and its sources on one screen. A review is a read, not a re-do.

Every approval, edit and send is in the audit log with who, when and what changed.

Graduating a type to auto-send

Week one, everything waits for a person. By week four the order acknowledgements have been approved unchanged two hundred times; you flip the rule and they send themselves. The complaints, the contract letters, the anything-with-money stay with people. The rule is yours, per type, reversible in a click.

• Auto-send per type, reversible
• Full audit of decisions
• Escalation when a result fails verification

A briefing that knows the company

“Overnight: 214 emails processed, 31 filed to client records, 6 waiting for you. The Friday report ran; margin on two orders is below your floor — see Output. Supplier X has not answered the RFQ due yesterday. Three contracts renew within 30 days.” Then you ask: “which three?” — and the answer comes with the documents.

• Written from records, libraries, mail and calendar
• Ask follow-up questions in the same place
• Delivered on screen and, if you want, by email

A brief written from the company’s own records and libraries — not a template with numbers pasted in.

The signals come from the modules: an order stuck here becomes a line in Act Now there.

Seeing across the modules

The value of one system is that it sees everything at once. A supplier’s late reply, an order waiting on the same missing item, a deadline in a contract and a task nobody picked up are four facts in four modules — and one warning on the dashboard, with the people who own them named.

• Cross-module signals, ranked by impact
• Owners named on every warning
• Trends you can open into the underlying records

The calendar the company writes

A contract filed on Monday puts its renewal on the calendar eleven months out, assigned to the account owner, with a reminder rule. A supplier confirmation sets the delivery date and the inspection task that follows it. A regulatory library sets the recurring filings for the year. Nobody typed a date.

• Dates extracted from contracts, mail and rules
• Owner and reminder rule per kind of deadline
• Recurring obligations from your procedures

A date on a record is a date on the calendar — with the document it came from one click away.

Scheduled AI work is on the same calendar as the people’s tasks, so a change in one shows in the other.

When the week changes

A delivery slips three days. The AI lists what depends on it — the inspection, the customer notice, the invoice date — proposes the new dates and the message to the customer, and waits for a person to confirm. Confirm, and the calendar, the tasks and the draft notice update together.

• Dependencies found across tasks and records
• Reschedule proposed, confirmed by a person
• Notices drafted for the people affected

The record that fills itself

A client emails: the message and its attachment are on the record before anyone reads them. A meeting note is dictated: transcribed and filed. A signed contract arrives: filed, its dates on the calendar. The AI summary at the top of the record is written from all of it — and answers when a colleague asks “where are we with this client?”

• Mail, documents and media filed automatically
• AI summary and open items per record
• History searchable across all records

Mail is linked to the record the moment it arrives; attachments are filed on it through intake.

Agreed terms live on the supplier record where automations can read them.

Suppliers with terms that mean something

Approval status, agreed prices, payment terms, certificates and their expiry dates sit on the supplier record. When an invoice or a price list arrives, the automation reads the record first: a price above the ceiling, an expired certificate or an unapproved supplier becomes an exception for purchasing, with the clause cited.

• Approval status and certificates with expiry
• Agreed terms enforced by automations
• RFQs and orders on the same record

Three RFQs, same desk

A property manager asks three contractors for a roof repair and gets a comparison that flags one bid excluding disposal. A manufacturer asks five suppliers for a machined part and gets unit price, tooling, lead time and certification lined up against the drawing. A municipality asks for office supplies and gets the compliant bidders ranked, the non-compliant explained. One desk, one document each.

• Spec-driven line items
• Exclusions and non-compliance named
• Ranking with the reasons

The RFQ desk is an automation: request out, quotes in, comparison built, decision requested.

Every figure in the comparison points at the line in the quote it came from.

What the comparison checks

Each quote is read against the request: missing lines, changed quantities, alternative parts, delivery terms, validity dates. Each supplier is read against its record: approval status, agreed ceilings, certificates in date. The document says what it found and what it could not find, cites every figure, and proposes — the person awards.

• Terms checked against the supplier record
• Validity dates and lead times on the calendar
• Award recorded on the RFQ and the supplier

One missing item, four orders

Four orders are waiting; each looks like its own problem. The AI sees they wait on the same SKU, that the supplier’s confirmed delivery slipped in yesterday’s mail, and that two of the four customers have a delivery commitment on the calendar. One warning on the dashboard: what is blocked, why, who to tell — with the customer notices drafted.

• Cross-order and cross-record reasoning
• Supplier mail read against open orders
• Notices drafted, held for a person

The AI sees across orders: the same missing item blocking four of them is one warning, with the supplier already named.

Receipt updates stock; stock updates the reorder logic; the proof of delivery closes the loop on the record.

From award to proof of delivery

An RFQ awarded on Output becomes a purchase order. The confirmation sets the delivery date. Receipt is scanned and stock moves. The sales order ships with a carrier label; the checkpoints show on the map; the proof of delivery files itself on the client record and the invoice date is set. Nobody re-typed a number.

• Purchase orders from RFQ awards
• Barcode receipt and movements
• Carrier rates, labels, checkpoints, proof of delivery

Found by what is in it

A contractor searches “kitchen before demolition, Maple Street” and finds the photo, dated, on the project record. A clinic manager finds the recording of the staff briefing about the new procedure. A distributor finds the video of the damaged pallet the carrier disputed. Nobody named a file.

• Automatic descriptions of images, video and audio
• Search by content across records
• Versions and rights noted on the asset

The AI writes the description of every asset — what, where, who, when — and that description is what you search.

Posts are results like any other: on Output with a preview, a channel and a date.

Publishing without a marketing department

The AI drafts from what happened: the project completed, the new certification, the event you hosted, the product that arrived. Your tone, your images, your channels. A calendar of proposed posts waits for approval; approve the month and it publishes on schedule; a comment that needs a person comes back as an exception.

• Drafts from real events on your records
• Monthly plan approved once or per post
• Replies that need a person are escalated

Knowledge that stays when people leave

The way you handle a claim, close a month, receive a delivery or greet a client is written or recorded somewhere. Put it in a library and it becomes training, help and verification at once: the new hire learns it, the assistant answers from it, the automation enforces it. When the person who knew it best leaves, the company still knows it.

• Procedures, documents and videos become the curriculum
• One source for training, help and automation
• Updates propagate; nothing is trained twice

The training content is the company’s own procedures and recordings. Update the procedure and the training updates.

After the path, the same assistant answers the day-to-day questions — with the source, so the answer teaches.

Days, not months, to be useful

A new operator finishes the path in the first days and then works with the assistant beside them: how do we handle this, where is that, what did we agree with this client. Every answer cites the procedure, so the answer teaches. Managers see where the questions cluster and fix the procedure, not the person.

• Assistant as day-to-day help after onboarding
• Questions clustered by topic for managers
• Refresher sessions when a procedure changes

Governing the AI like a colleague

The AI works under the permissions of the user or the automation that asked. A purchasing assistant reads purchasing libraries; a clinic’s care fields are visible to clinical roles; an automation that files documents cannot export them. What the AI retrieved, built, cited and sent is in the log beside the human decisions — readable by your auditor.

• AI acts under the permissions of who asked
• Per-record disclosure view: who saw this
• Logs exportable for audit under the export right

AI actions are logged with the same detail as human ones: what it read, what it cited, what it did.

The structure is yours to shape; the platform underneath is maintained and updated by Zanus.

Platform kept by us, rules kept by you

Two-factor authentication, snapshots, signed updates, encrypted storage and isolated tenants are the platform’s job. The organisation tree, the user types, the access hours, the approval rights and the approved libraries are yours. Zanus maintains the platform; you maintain your rules and approved knowledge.

• 2FA, snapshots, signed updates
• Isolated tenant; private environment on AI PRIVATE
• Same governance on a unit you own

Read as a source, write under a rule

The accounting system is a source: the AI reads the open invoices when a supplier asks about payment. The ERP is a destination: an approved order is written to it as the last action of the flow. Writes follow the same rules as any Output result — approved per item or allowed per type — and every call is in the log.

• Connected systems as sources for verification
• Writes as automation actions under approval
• Every call logged with payload and result

A connected system is a source for the AI step and a destination for the action step — under the same approval rules as everything else.

The API turns the Back Office into a service other software can ask — with the same verifier and the same permissions.

Your software asks the company AI

A quoting tool asks whether a component is in stock and in spec. A portal asks for the status of a customer’s case, in the customer’s language. An internal script asks for the summary of a file. Each gets an answer built from your sources, verified and cited, under the permissions of the API key that asked.

• OpenAI-compatible endpoints where useful
• Permissions and rate limits per key
• Same verifier, same citations

Three processes no product ships

A field-service company gets an inspection app: photos, checklist, measurements, the report generated and sent. A legal practice gets a case-qualification app: intake questions, conflict check against the records, the engagement letter drafted. An accounting firm gets a period-close robot: statements pulled from the connected systems, reconciled, exceptions listed, the pack assembled for review.

• Apps for the people, robots for the steps
• Your records, libraries and rules underneath
• Results on Output like everything else

The same building blocks make an inspection app, a case-qualification app and a period-close robot.

A robot is a user with a job description: it works under permissions, logs everything and stops where the rule says.

Built to order, governed like the rest

You describe the process; we write the scope; you approve it; we build. The app or robot runs inside your tenant with your permissions, uses the verifier for every AI step, delivers to Output, and logs every action. When the process changes, the scope changes — not the platform.

• Written scope, quoted once
• Same verifier, permissions and audit
• Changes as scoped work, platform unchanged

A customer call, followed inside

A customer calls; the Front Office answers, books the visit and promises a document. Inside, the Back Office files the transcript on the record, puts the visit on the calendar with a preparation task, finds the document in the library and drafts the email with it attached — on Output for the account owner, or sent under the rule you set.

• Transcript and intent on the record
• Booking to calendar and task
• Promised documents found and drafted

After the call: the transcript, the booking, the task and the draft are on the record before the customer hangs up.

Front Office signals — a complaint call, a hot lead, a booking change — appear on the Back Office dashboard with the owner named.

Sold apart, better together

Each product stands alone and is priced on its own page. Connected, they share one approved knowledge, one set of records and one calendar. Start with the one that hurts most; add the other when you are ready. Nothing is configured twice.

• Separate products, separate pricing pages
• Shared knowledge, records and calendar when connected
• Add the second product without re-doing the first