Theodoros Ampas
Join the newsletter

Back office

Clear the busywork that sits between your people and their real jobs.

Operations & Internal AI builds the systems behind the scenes: internal knowledge assistants, document and data processing, reporting, workflow orchestration, repetitive back-office processes, and the connective tissue that reduces handoffs between people and tools.

Start a project

Best when a capable person is carrying repeatable work that has never been written down as a process.

By Last updated

In one sentence

Operations & Internal AI is an engagement in which Theodoros Ampas builds internal AI systems for a company’s back office: knowledge assistants that answer from company documents, document and data processing pipelines, automated reporting, and workflow orchestration that reduces manual handoffs, all running inside the company’s own tools and accounts.

Your best people are doing your most repeatable work.

The operations lead knows where every document is, which spreadsheet is the real one, and what the answer to the question everyone asks on Monday is. That knowledge lives in their head and their afternoon.

Every handoff between a person and a tool, or between two people, is a place where a detail gets lost, a status is not updated, and someone has to ask again.

Internal AI is not about replacing that person. It is about giving them a system that answers the Monday question, processes the Friday report, and moves the record without them being the bridge.

What gets built.

Knowledge that answers

Internal assistants over your policies, procedures, contracts and history, with sources shown, so the team stops asking one person for everything.

Documents and data that process themselves

Intake, extraction, classification and validation of the documents and records the business handles every day, with exceptions routed to a person.

Reporting and orchestration

Reports assembled from the systems of record on schedule, and workflows that move work between tools without a human copy-pasting it across.

From one handoff to a running process.

  1. Week 1

    Pick the handoff

    One process where a capable person is the bridge. Document it as it really runs, with time per step and where it fails.

  2. Weeks 2 to 4

    Build the system

    Assistant, pipeline or orchestration built inside your tools, with exceptions and approvals routed to the right person.

  3. Weeks 5 to 6

    Run beside the team

    The system handles the real volume while the person who used to do it reviews. Edge cases become rules or become documented exceptions.

  4. After

    Hand over

    Runbook, monitoring, named owner. The person who was the bridge now directs the system instead of being it.

Internal does not mean unguarded.

  • Access scoped to what each assistant or pipeline actually needs. Nothing reads the whole drive because it was easier.
  • Answers carry sources. If the assistant cannot find one, it says so instead of guessing.
  • Exceptions and low-confidence cases go to a person, always.
  • Every processed document and every change is logged and reversible.

What you leave with.

  • An internal assistant the team actually uses, answering from your own documents.
  • Document and data pipelines that handle the daily volume with exceptions routed to people.
  • Reports that arrive on schedule from the systems of record.
  • Hours per week returned to the people who were the bridge, measured against week one.

This is for you if

  • One person is the bridge between two tools, or between the team and the answer.
  • The same report, document or status update is assembled by hand every week.
  • Knowledge lives in people’s heads and leaves with them on holiday.

Questions founders ask

What tools can this connect to?

Anything with an API or a file export: Google Workspace, Microsoft 365, Notion, Slack, your ERP, accounting software and shared drives. The design phase lists exactly what is connected and with which permissions.

Is our internal data safe?

The system runs in your accounts with scoped access. Data boundaries, including whether anything leaves your infrastructure to reach a model provider, are decided and documented in design.

How accurate are the answers?

Answers are grounded in your documents and show their sources. When the assistant cannot find a source it says so. An evaluation set is built from real questions so accuracy is measured, not assumed.

How long does it take?

A first internal system usually takes four to six weeks from picking the handoff to a stable process the team runs.

Give the bridge back their afternoon.

Tell me which person everyone asks and what they are asked. That is usually the first system.

Start a project