The routine work, done by agents, signed off by people.

Agents process invoices, prepare the close, draft campaigns and keep your CRM current, inside your systems. Your team approves what matters.

Why routine back-office work is ready for agents

In most companies, finance, marketing and sales still run on people copying data between systems: keying in invoices, rebuilding the monthly report, chasing leads in a shared inbox. None of it is hard, but it never stops, and it grows with the business.

Older automation followed fixed rules and broke the moment a supplier changed its invoice layout or a customer wrote in free text. Agents read messy documents the way a person does, find the matching records, apply your rules and note what they did.

Regulation adds pressure. German businesses have had to accept e-invoices since January 2025, and those above €800,000 turnover must send them from January 2027. The real question is how to automate with controls your auditor accepts, and prove the hours saved.

What we build

  • Invoice intake and posting

    Agents read every invoice, match it to the order and prepare the posting in DATEV or SAP for release.

  • Reconciliation and month-end close

    Bank lines matched, differences explained, the close checklist kept current. Controllers review exceptions, not spreadsheets.

  • Campaign briefs and content

    Briefs, product copy and newsletters drafted in your tone, checked against approved claims, sent to the right reviewer.

  • Campaign reporting

    The weekly report built from your ad, web and CRM data, with a short note on what moved.

  • Lead research and CRM upkeep

    New leads researched and scored, missing fields filled, duplicates merged. Your CRM stays accurate without manual data entry.

  • Quotes and order preparation

    Customer requests turned into draft quotes from your price lists and discount rules, checked for margin, approved by the owner.

How it works

Every agent follows the same five steps, from trigger to an approved, logged result.

  1. A trigger arrives

    An invoice lands, a lead fills in a form, or the month ends. The agent picks it up.

  2. The agent gathers context

    It reads the request and looks up the order, supplier, customer history or price list it needs.

  3. It prepares the result

    A coded posting, a matched reconciliation, a draft quote or a report, with every source logged.

  4. A person approves

    Anything that moves money or reaches a customer waits for approval in Teams, Outlook or your ERP.

  5. Results are measured

    Every run is counted: cases handled, exceptions, corrections and hours saved against the agreed baseline.

How we roll it out

  1. Discovery

    2 to 3 weeks

    We sit with your team, measure the process and pick one or two workflows with a clear, measurable saving.

    • Process map with time baseline
    • Workflows ranked by value
    • Integration and data access plan
  2. Build

    6 to 8 weeks

    Agents built inside your systems run in shadow mode beside your team until accuracy meets the agreed bar.

    • Working agents in your environment
    • Approval flows in your existing tools
    • Accuracy report from shadow mode
  3. Run and extend

    Ongoing

    Agents go live with approvals in place. We review the numbers weekly, then add the next workflow.

    • Live view of volume and hours saved
    • Runbook and handover to your team
    • Roadmap for the next workflows

In production elsewhere

Public examples from other organisations, not unus clients, showing what agents already do in finance, marketing and sales.

Public case studyE.ONEnergy · Germany · 2025 AI agents process invoices inside its SAP environment >95% tax coding accuracy
The situation
E.ON runs over 500 company codes in more than 30 countries. About half its invoices had no purchase order, VAT rules varied by country, and older rules-based automation could not cope.
What they deployed
With Hypatos, E.ON deployed agents in SAP for auto-posting, purchase order matching, ledger and tax coding, and VAT checks. Unclear supplier matches go to a person. Five KPIs set target automation rates of 75% to 95%.
The outcome
According to the vendor's case study, end-to-end auto-posting doubled, tax coding accuracy exceeds 95% across jurisdictions, and purchase order matching reached a 90% automation rate, on target in the first year in production.

Agree measurable targets per task before you start, and clean the supplier master data first, because that is where automation stops.

Source: Hypatos, customer story: E.ON (energy, EU), Undated, accessed November 2026
Public case studyKlarnaPayments · Sweden · 2024 In-house AI tools take over copy, images and translation About $10 million a year
The situation
Klarna's marketing relied on external agencies and stock imagery. New images took around six weeks, too slow and costly for bespoke material for every retail occasion.
What they deployed
Its in-house team generates images with Midjourney, DALL-E and Firefly, checked for brand, quality and legal compliance: over 1,000 in Q1 2024, in seven days instead of six weeks. Copy Assistant is used for 80% of copywriting.
The outcome
Klarna says its sales and marketing spend fell 11% in the first quarter of 2024 while it ran more campaigns. AI accounts for 37% of the savings, about $10 million on an annualised basis, including a 25% cut in spending on external agencies for translation, production, CRM and social media.

Bring routine content work in-house with AI, but keep brand, quality and legal checks inside the flow, because they are what make the speed safe.

Source: Finextra, Klarna press release: AI helps Klarna cut marketing agency spend by 25% and run more campaigns, June 2024
Public case studySiemensIndustrial technology · Germany · 2026 Two AI agents qualify inbound sales leads Response times: days to minutes
The situation
Siemens sales teams received thousands of inbound leads from industrial customers worldwide. First responses could take days, and some leads were never followed up.
What they deployed
On Salesforce Agentforce, two orchestrated agents hold the first conversation with each lead and qualify it against 50 rules. An agent also gives sellers the data to prioritise qualified leads.
The outcome
Salesforce reports that lead response times fell from days to minutes, and that leads which were previously left unworked can now be converted, opening up new customers.

Start where volume outruns your team, write the qualification rules down explicitly, and hand the conversation to a person once intent is clear.

Source: Salesforce Germany, press release: Agentblazer Award 2026, May 2026

These are public deployments by other organisations, summarised from the sources linked. They are not unus clients.

Built for European rules

Data stays where you decide

Agents run in your cloud tenancy in an EU or Swiss region, or on your servers, under GDPR and the Swiss nDSG.

An audit trail for finance

Every proposal, approval and posting is logged with inputs, rule and approver, designed for GoBD, so auditors can trace each booking.

Human approval by design

Payments, master data, quotes and customer messages wait for a named person, so no significant decision is fully automated (GDPR Art. 22).

EU AI Act duties in scope

Back-office agents are rarely high-risk. Customer-facing agents disclose they are AI (Article 50); Annex III uses face high-risk rules from December 2027.

Questions buyers ask

Do we need to replace DATEV, SAP or our CRM?

No. The agents work through the interfaces your systems already offer: DATEV's data services, SAP's Service Layer for Business One or OData and IDocs for S/4HANA, Microsoft Graph for Outlook and Teams, and the HubSpot and Salesforce APIs. Where no interface exists, we use a controlled file exchange. Your team keeps working in the tools it knows.

How is this different from RPA or classic OCR?

RPA follows fixed clicks and OCR templates expect fixed layouts, so both break when a document or request changes. Agents read unstructured input, look up the context they need and apply your rules, and they record the reasoning behind each proposal, which makes exceptions faster to resolve. Where an existing bot works well, we keep it and let the agent handle the cases it cannot.

Who is accountable if an agent makes a mistake?

Your people, as today, because nothing that moves money or reaches a customer is released without their approval. Agents start in shadow mode, where every proposal is compared with what your team actually did, and they take on more only once agreed accuracy thresholds are met. Every action is logged, so an error can be traced and corrected.

How do you measure the time saved?

Before we build anything, we measure the current process: volume, handling time per case and error rate. After go-live, the agents count every case they handle, every exception they raise and every correction a person makes. You see the baseline and the live figures side by side, so the saving is measured rather than estimated.

Which workflow should we start with?

One with high volume, clear rules and a single owner. Incoming invoice processing, bank reconciliation, CRM data upkeep and monthly campaign reporting are typical candidates. We avoid starting with processes that are rare, contested or still being redesigned, because a first agent should prove its value within a quarter.

Does our data leave the EU or Switzerland?

It does not need to. We deploy into your own cloud tenancy in an EU or Swiss region, or on premises, and choose model hosting that keeps processing there. Data processing agreements, access rights and retention periods are agreed with your data protection officer before any live data is used.

Bring one process. We'll bring the loop.