Umbra Penance

Services

AI integration · UK

Use AI inside the workflows your team already relies on.

We put AI where it saves real time: reading documents, sorting requests, qualifying leads, drafting replies and answering questions from your own knowledge. It runs inside your systems, with guardrails and a human where it matters.

Tell us where AI could help
  • Free and no obligation
  • Pick a time straight after the form
  • You speak to one of the two founders
  • Private data stays controlled
  • Human-in-the-loop by design
  • Measured against real outcomes
Document intakeexample
  1. Document uploadedPDF, email, scantrigger
  2. ClassifiedType and ownerAI
  3. Fields extractedValidated against rulesAI + rules
  4. Low confidence → humanreview
  5. Filed & system updatedCRM · storage

You’ll speak with Adel or Mehdi, the founders who design, build and test every project. About us

Project

Already running in production.

Client project · Education & training · client name kept confidential

AI answers from each learner's own course

An AI assistant that answers from the course, not from the internet.

Problem
Learners often waited a day or more for answers from trainers, and the teaching team was overloaded with routine questions.
What we built
An AI assistant that answers from each course's own material, reads learners' project files and handles admin questions, inside the learner platform.
The hard part
Keeping each course's material strictly separate, reading 3D project files, and keeping AI keys out of the browser.
Result
Routine questions are answered immediately from the right course; trainers focus on the rest.
Safeguards
Per-course retrieval isolation enforced server-side. Model and API keys held server-side behind a proxy; never exposed to the browser.
Read the full project

Where AI projects usually stall

  • The team uses ChatGPT in a browser tab, but nothing is connected to your systems.
  • A proof of concept impressed in a demo and broke on real documents.
  • Nobody is sure what data can safely be sent to an AI provider.
  • An "AI agent" was proposed for a job a simple rule would do better.
Two colleagues working together at a laptop by a window

Tested on your data

We prove it works on your documents before it goes live.

Every AI step is tried on real samples from your business, measured, and given a human review path for the cases it isn't sure about.

Production AI we implement

  • AI document processing

    Extract structured data from invoices, contracts, applications and forms, validated and routed into your systems.

  • Classification & triage

    Sort enquiries, tickets and emails by type, urgency and owner the moment they arrive.

  • Lead qualification

    Score and enrich inbound leads against your criteria so sales spends time on the right conversations.

  • Assistants that answer from your own documents

    Answer questions from your policies, documentation or knowledge base, with sources, access control and no training on your data.

  • Summarisation & drafting

    Meeting notes, case summaries, replies and reports drafted for a person to approve.

  • Agentic workflows, where justified

    Multi-step AI that takes actions, only when the task genuinely needs it and with strict permissions.

AI that's safe to connect to your systems

An AI step with access to your data and tools is a new attack surface. We design for that.

  • Data minimisation: only what the task needs is sent to a model.
  • Providers and settings that don't train on your data.
  • AI steps get only the permissions they need, and we test them against attempts to make them take unintended actions.

When AI is, and isn't, the answer

The right call when

  • The input is unstructured: documents, emails, free text.
  • The task needs judgement that's consistent but hard to write as rules.
  • A human can review the output where mistakes would be costly.

What we won’t recommend

  • A deterministic rule would be cheaper and more reliable.
  • There's no way to check the output is right.
  • The data can't be processed safely by any available provider.

Technology

Model-agnostic: we choose the provider on accuracy, cost and data requirements.

OpenAIAnthropic ClaudeMistralVector databasesPostgreSQL + pgvectorn8nMakePythonTypeScript

Questions we get asked

Is our data used to train AI models?
We use providers and API settings where your data is not used for training, and we minimise what is sent in the first place. Data flows are documented for your GDPR records.
How accurate is AI document processing?
It depends on the documents. We test on your real samples before launch, measure accuracy, and route low-confidence results to a person so errors are caught rather than posted.
Can the assistant answer from our own documents?
Yes. The assistant searches your own documents and answers from them, citing its sources (this is often called retrieval-augmented generation, or RAG). Answers stay grounded in your information rather than the model's general knowledge.
Do we need an AI agent?
Usually not. Most value comes from a single, well-placed AI step inside a reliable workflow. We recommend agents only when the task genuinely needs multi-step reasoning and actions.

Tell us which task you'd hand to AI if you trusted it.

On a 20-minute call we'll tell you whether AI is the right tool, how we'd make it reliable, and what it would take.

Describe the task instead
  • Pick a time in the next step
  • Free, no obligation
  • 20 minutes on Google Meet