Bonsay Tech

AI & automation · built for production

AI that reaches production. And stays there.

We find the one process where AI pays back, build it into your real systems — ERP, CRM, inbox, documents — and measure it. Agents, knowledge search, document AI, voice and automations, with a person approving anything risky.

Book a process reviewCheck your readiness

Open tools, your environment, everything handed over. No vendor lock-in.

Workflow trace · example run · quote request
  1. 0.0sNew email · “Quote for 120 custom mats, delivery Łódź”
  2. 0.6sIntent: quote · product: mats · qty 120 · confidence 0.96
  3. 1.3spricing.lookup(sku, qty=120) · margin check
  4. 2.1spdf.render(quote) · crm.upsert(contact)
  5. 2.8sPolicy: value > €5k → waits for human approval
  6. —Approved by sales · sent · logged
Illustrative · every action logged · alert on failure
process and data audit before anything is built
2 weeks
from audit to a pilot in your environment
2–4 weeks
of agent actions logged, risky ones approved by a person
100%
lock-in: you own workflows, prompts, code and credentials
0

What we build

From one workflow to a full AI platform

Model-agnostic, inside your tools, measured against the process before it.

  • AI agents

    Agents that work in your real tools through APIs and MCP, follow the steps of a process and stop for approval where money, clients or data are at stake.

    • Tool calling
    • MCP
    • Human-in-the-loop

  • Knowledge & RAG

    Private search over documents, tickets and recordings that answers with sources, respects access rights and updates itself.

    • Hybrid search
    • Re-ranking
    • Access control

  • Document AI

    Invoices, orders and contracts parsed into your ERP or spreadsheet, with a review queue for anything uncertain.

    • OCR
    • Vision models
    • Review queue

  • Voice & calls

    Call transcription and scoring, voice notes turned into tasks and emails, voice agents for routine calls with a clean hand-off to people.

    • Speech-to-text
    • Scoring
    • Hand-off

  • Workflow automation

    n8n and code-based workflows that connect CRM, ERP, email and messengers, with retries, alerts and monitoring.

    • n8n
    • APIs
    • Monitoring

  • Private & open-weight models

    When data must stay inside, models run on your own servers or EU cloud. We pick the cheapest model that does the job reliably.

    • Self-hosted
    • EU data
    • Model-agnostic

What pays back first

Processes we build, so you can recognise your own

Results depend on your volumes. We count the manual hours before we build and measure again after.

  • Enquiry handling

    Email and WhatsApp requests read, qualified and answered with the key details collected, so no lead waits until Monday.

    • LLM
    • WhatsApp API
    • CRM

  • Quotes and proposals

    A brief goes in, a branded PDF quote comes out. Prices live in a database, margins recalculated on the server.

    • PostgreSQL
    • PDF
    • Approval step

  • Document processing

    Invoices, orders and contracts into your ERP or spreadsheet, with a person reviewing what the model is unsure about.

    • OCR
    • ERP API
    • Review queue

  • Company knowledge search

    A private search over your documents and recordings that updates itself and answers with sources.

    • Embeddings
    • Vector DB
    • Drive

  • Content pipelines

    One source text turned into posts, articles and localised versions, with a person approving before publication.

    • Multi-model
    • CMS
    • Human review

  • Back-office admin

    Voice notes and messages turned into tasks, calendar entries and emails, confirmed in the same chat.

    • Speech-to-text
    • Telegram
    • Calendar

Quality, not demos

Models are rented. Harnesses are owned.

Switching to a newer model rarely fixes an agent. The engineering around it does. That is what we build, document and hand over.

  1. 1ToolsWhat the agent may call — and nothing more. Minimum access by default.
  2. 2ContextThe right data at the right moment: your prices, your policies, your customer history.
  3. 3MemoryWhat it remembers between runs, and what it must forget.
  4. 4VerificationA written definition of done. Checks that the work is finished, not just reported as finished.
  5. 5Logs & alertsEvery action traceable. Someone is told when something fails.

How a project runs

One process. Measured. Then scale.

Two ways to start: a dedicated automation engineer in your team, or a fixed-scope project with the price agreed before we start.

Guardrails we ship by default

  • Minimum access: the workflow touches only the tools and data it needs
  • A person approves anything risky: payments, refunds, client emails
  • Every action logged, alerts when something fails
  • Everything documented and handed over: workflows, prompts, credentials
  • EU AI Act-ready: AI disclosure, human oversight, audit trail
  1. 1

    Process audit1–2 weeks

    Map the workflow, count the manual hours, pick what pays back. Agree what the automation may touch and what a person must approve.

  2. 2

    Pilot2–4 weeks

    One process automated end to end, in your environment, with a before-and-after measurement.

  3. 3

    Scale outongoing

    The proven pattern extended to other teams and processes.

  4. 4

    Run and ownalways

    Monitoring, error handling and full handover. You own everything we build.

Who can work with you

Engineers who have shipped automations to production

Remote, full-time or part-time. Full CVs on request.

  • Slava · AI business automation specialist

    7+ years · English B2 · Poland

    • n8n automations end to end: mapping, build, monitoring, handover
    • AI pipelines for document sorting and reply drafting, with human review
    • n8n
    • OpenAI API
    • Python
    • Django
    • AWS

  • Serge · n8n and full-stack developer

    5+ years · English B2 · Warsaw

    • Automations plus the web tools around them: React front ends, NestJS APIs
    • Built an AI invoice parser for ERP systems
    • n8n
    • React
    • TypeScript
    • NestJS
    • AWS

  • Lazar · Senior Python backend and AI engineer

    5+ years · English C2 · North Macedonia

    • APIs, automation and AI integrations from database design to production
    • Integrates OpenAI and Anthropic models into real products
    • Python
    • FastAPI
    • LLM APIs
    • Docker
    • AWS

  • Alexander · Solution architect, automation and integrations

    15+ years · English B2 · Georgia

    • Designs the integrations and event-driven systems automations plug into
    • 7 client projects delivered through Bonsay since 2018
    • Java
    • Spring Boot
    • Node.js
    • BPMN
    • Kubernetes

Free · 60 seconds

Is your AI project heading to production — or to pilot purgatory?

Five questions we ask before anyone builds anything. Answer honestly; you get an instant verdict and the next step.

Question 1 / 5

Is it one clearly defined process — clear inputs, clear outputs?

Results

Automations already running

Measured against the same process before automation. Details and references on the call.

  • Customer service AI

    Service business · anonymised

    Every enquiry answered in under 5 seconds

    An agent reads every enquiry, collects the details, books the slot and sends a payment link. Staff approve exceptions.

    direct bookings
    +20%
    operations hires avoided
    2
    first reply
    <5 s
  • Sales AI

    B2B sales team · anonymised

    100% of sales calls scored, not a random few

    Every call transcribed and scored against the team's playbook, with coaching notes for each rep the same day.

    sales conversion
    +17%
    customer satisfaction
    +23%
    of calls reviewed
    100%
  • Knowledge AI

    Operations team · anonymised

    A knowledge base that updates itself

    A private search over company documents and recordings that answers with sources and keeps itself current as documents change.

    manual work on transcription and document updates
    −70%

FAQ

Questions we hear first

Which models do you use?

We are model-agnostic: GPT, Claude, Gemini, Qwen, Llama, Mistral and others. We pick the cheapest model that does the job reliably and keep the option to switch. When data must stay inside, we run open-weight models on your servers.

What if the AI makes a mistake?

Risky actions — payments, refunds, client emails — wait for a person's approval. Every action is logged, failures trigger alerts, and uncertain cases go to a review queue instead of being guessed.

Is our data safe?

You contract with an EU company under GDPR with a DPA. Workflows get minimum access, and nothing is sent to a model your contract does not allow.

Do we depend on you afterwards?

No. You own the workflows, prompts, code, credentials and documentation. We can run it for you, or hand it over to your team.

How do you measure success?

Before we build, we count the manual hours, response times or error rates of the process. After the pilot we measure the same numbers again. That comparison decides whether we scale.

Are you ready for the EU AI Act?

We build AI disclosure, human oversight and audit trails into every system and help you classify the use case early. For formal legal opinions we work alongside your counsel.

Related

Often comes up in the same conversation

Next step

Which task takes the most hours in your team?

Book 30 minutes and describe it. You get a scoped proposal: what gets built, on which stack, by whom, and what it is expected to save — or an honest “not worth automating”.

  • Process and data review on the call
  • Fixed scope and price before any build
  • You own everything we build
Book a 30-min call WhatsApp Telegram

You will talk to

Gleb Zamyslovsky

Executive Officer · Warsaw

One reply within a business day: questions, a fixed price, or an honest “not worth it”. No account managers in between.