AI integration services

AI integration that does real work in your business

Most businesses don't need their own AI model. They need AI in the right places: reading incoming email and documents, sorting requests, drafting replies, answering questions from their own systems. We build those integrations with structured output, clear rules and a person in control wherever it matters.

  • Structured output that lands straight in your systems
  • Human review wherever AI output reaches a customer
  • Cost per task known before you roll out

Where AI fits in a process

  1. Input

    Emails, documents, photos, form entries or records from your database.

  2. Understand

    A model extracts fields, classifies intent and summarises, returning structured data.

    • extract
    • classify
    • summarise
  3. Check

    Rules validate the result. Anything uncertain goes to a person.

  4. Act

    A record is updated, a reply drafted, a task created or an agent takes the next step.

  5. Measure

    Accuracy, exceptions and cost per task, tracked from the first day.

The pattern behind the AI in our own products.
What we can build

Where we put AI to work

Not AI for the sake of it. Each of these replaces a task people do by hand today, and each one returns data your systems can use.

Email and document processing

Incoming email, PDFs and photos read by a model, with the fields that matter extracted into your CRM, ERP or database.

Classification and routing

Requests, leads and tickets sorted by intent and urgency, then sent to the right person, queue or follow-up action.

Drafting in your tone

Replies, product texts, captions and summaries written in your style and in several languages, ready for a person to approve.

Assistants on your own data

Let Claude or ChatGPT search and use your own systems through an MCP connector, with the same permissions your team already has.

Agents for multi-step work

Agents that research, prepare and queue work across several tools, with an approval step before anything is sent or changed for good.

AI features in your product

AI your customers use inside your product, with usage limits, a known cost per run and clear labels when content is generated.

Proof, not promises

AI at work in our own products

Every photo uploaded to Twelve Lumens or Salestreak is analysed by a vision model before anything is generated. It returns structured data, not chat: the room or product type, a caption and alt text in five languages, and flags for collages or photos that already carry an AI label.

In our sales system, replies are classified into ten categories, and each category triggers its own next step on the server. Postcrest drafts captions and comment replies and summarises what's working per account. In every case, a person stays in charge of what goes out.

  • Forced function calls with a JSON schema, so model output always fits the data model
  • Descriptions written natively in five languages, with an instruction never to invent facts
  • Opt-outs suppressed across every campaign as soon as a reply is classified
  • Photos sent at 1024 pixels and low detail, which keeps the cost per analysis down
  • A person approves every outgoing email an AI has drafted
Twelve Lumens property page with uploaded photos sorted into bathroom and bedroom sections, each with a short description written by a vision model
A vision model sorts every upload by room and writes a description of each photo (Twelve Lumens).
languages written natively by AI
5
reply categories classified automatically
10
room types recognised by a vision model
21
of AI-drafted emails approved by a person
100%
Who it's for

Who we build AI integrations for

SMEs exploring AI

You see the potential but not the first step. We find the task where AI saves the most time and build that one first.

Software companies

AI features your customers will pay for, built into your product with limits and costs that fit your pricing.

Service and operations teams

Incoming email, documents and requests handled faster, so people spend their time on the cases that need them.

Organisations with strict data rules

Clear data flows: what is sent to which provider, what is stored and for how long, agreed before anything goes live.

How it works

How we integrate AI

One measurable task at a time, with the guardrails designed in from the start.

  1. 1

    Pick one task

    A concrete, frequent task where AI clearly helps. Not an AI strategy, but a job you can measure.

  2. 2

    Test on your own examples

    We run candidate models on real samples from your business and compare accuracy, speed and cost per task.

  3. 3

    Design the guardrails

    Structured output, validation rules, confidence thresholds and the moments where a person steps in.

  4. 4

    Connect it to your systems

    Your inbox, CRM, product or database, so the result lands where the work happens.

  5. 5

    Roll out gradually

    Start with suggestions a person approves, then automate the cases that prove reliable.

  6. 6

    Monitor and improve

    Track accuracy, exceptions and costs, and switch models when a better or cheaper one arrives.

Models and tooling

We work with models from OpenAI, Google and others, chosen per task, and connect them to your systems with the same care as any other integration: background jobs, retries, logging and tests.

  • OpenAI
  • Google Gemini
  • Function calling
  • JSON Schema
  • MCP
  • Claude
  • ChatGPT
  • Ruby on Rails
  • Solid Queue
  • PostgreSQL
Cost & timeline

What does AI integration cost?

There are two budgets: building the integration, and running the models afterwards. We estimate both. They depend on:

The task
Classifying a short email is cheap per run. Reading long documents or photos costs more.
Volume
How many items per day or month. That sets the running costs and whether batching pays off.
Accuracy needed
The higher the bar, the more testing, validation rules and review steps.
Integration
Connecting to your inbox, CRM, product or database, and the permissions involved.
Data rules
Which data may leave your systems, which providers are allowed and what may be stored.
FAQ

Questions about AI integration

How much does it cost to integrate AI into our business?

It depends on the task, the volume and how accurate the result must be. After a free intro call you get a written scope and estimate, including the expected running cost per task.

How long does an AI integration take?

That depends on the task and the systems involved. We start small: one task, live first as suggestions a person approves, then more automation step by step. The estimate after the intro call includes a timeline per phase.

Where should we start with AI?

With one concrete, frequent task, such as sorting incoming email, extracting data from documents or drafting replies. Pick something you can measure, prove it works, then expand.

Do we need our own AI model?

Almost never. Existing models from providers such as OpenAI and Google, combined with your own data and good guardrails, cover most business needs.

Is our data safe with AI?

We design the data flow upfront: what is sent to which provider, what is stored and for how long. With business APIs such as OpenAI's, data sent through the API isn't used for training by default, and we strip personal data that a task doesn't need.

Will AI replace our team?

That's not how we build it. AI takes the repetitive part of a task; people handle decisions, exceptions and anything that reaches a customer.

Can AI work inside our existing software?

Yes, through an integration or an MCP server that lets assistants such as Claude and ChatGPT use your software within the permissions you set.

Have a task AI should be doing?

Tell us what the task is, how often it happens and where the data lives. You'll get an honest view on whether AI is the right tool and what it would take.

martin@balk.dev

Typically a reply within one business day.

Useful to mention

  • What you want to build, in a few sentences
  • Who is going to use it
  • Any deadline or launch date
  • Tools and systems it should connect to