AI Integration

For teams that read documents manually, sort requests in email or plan using last year's spreadsheet, we connect AI to your ERP and document archive. The model prepares a draft; decisions and data remain with you.

Who it is for

Integrating AI into business processes connects a model to your ERP, email and document archive to support a specific step. Enterprise AI solutions design user access, permitted data and human approval together. The model reads, classifies or suggests; operations requiring an authorised person's decision are not automatically completed.

When is it useful?

  • Sales support: email and PDF orders are entered into the ERP manually.
  • Accounting and warehouse: delivery notes and invoices are compared with orders one by one.
  • Technical teams: the information needed from specifications is lost among folders.
  • Planning and finance: demand and cash forecasts use last year's spreadsheet.

When is it unnecessary?

  • If a rule can be written, such as an amount threshold or approval sequence, it is cheaper and easier to audit.
  • e-Invoices and e-Delivery Notes are already XML; no model is needed to read them.
  • If data is scattered, the first step is to bring the process into one system: custom software development.
AI or rules?

Scope

We agree the scope during discovery. The example below is a typical first release taking one scenario, such as reading incoming documents, live from end to end.

  1. Data connections

    Email, document folders, ERP and CRM.

  2. Model layer

    Reading or forecasting, independent of the provider.

  3. Validation rules

    Output is compared with ERP data.

  4. Approval screen

    The draft, its source and uncertain fields side by side.

  5. Evaluation dataset

    Examples with correct answers; every change is measured against them.

  6. Monitoring and logging

    Correction rate, cost and audit trail.

Outside this example's scope

  • Training a language model from scratch
  • Automated decisions about people, such as recruitment or credit
  • Model and cloud fees (charged to your accounts)
  • General AI training for employees

Assumptions

  • Digital input: email, PDF or ERP API.
  • Historical examples with correct answers are available.
  • The process owner defines approval rules.
  • The first release covers one scenario.

Scenario and data suitability are clarified during discovery: What is a discovery sprint?

Example scenarios

Explore the use cases

In every scenario, the model handles one step while a person retains the decision.

  1. Incoming document inbox

    Role
    Purchasing and accounting specialist
    Input
    Email attachments containing PDFs or scanned delivery notes and invoices
    Action
    The document type is identified, fields read and compared with the order.
    Approval
    A specialist corrects uncertain fields.
    Output
    A matched ERP record and a discrepancy list for the relevant team
  2. Internal knowledge assistant

    Role
    Proposal engineer
    Input
    Documents, specifications, procedures and a question
    Action
    Only authorised documents are searched; the answer is based on those sources.
    Approval
    For information only; the engineer decides on the proposal.
    Output
    A sourced answer or “not found”
  3. Demand and cash forecasting

    Role
    Planning or finance specialist
    Input
    Sales, consumption or collections history
    Action
    Forecasts by product, facility or account, compared with the current method.
    Approval
    The specialist uses the forecast or changes it with a recorded reason.
    Output
    A forecast with confidence intervals and a deviation alert

Integrations

Integration details

The model works within the systems your teams already use, rather than in a separate application. Everything it writes to the ERP remains a draft until approved.

Systems that can be connected and typical data flows
SystemExamplesTransferred dataDirection and frequency
ERPLogo, Netsis, Mikro, SAPOrder, account and product historyPeriodic reads; writes saved as drafts
Email and documentsMicrosoft 365, Google WorkspaceMessages, attachments and scanned documentsFrom designated inboxes as documents arrive
CRM and supportSalesforce, HubSpot, Dynamics 365Requests and customer historyClassification and draft responses written back
Metering and financeMeters, sensors, banks and data warehousesConsumption, sales and collectionsHourly or daily; forecasts shown in the dashboard
Model providerCloud APIs or an open model on your own serverRequired fields, maskedEvery call logged; provider can be changed

Where is data processed?

Only fields needed for the task reach the model; unnecessary personal data is masked beforehand. The processing region or your own server is selected with your legal adviser in light of KVKK requirements. Operation IDs, model versions and approval status are recorded for auditability; raw documents and personal data are retained only where necessary, with defined retention periods and access permissions.

Questions for the provider: data and KVKK questions

Product names are examples and do not imply a partnership.

Process

Steps in our process
  1. 01

    Discovery

    We choose the scenario, assess whether data samples are sufficient and define success criteria.

    1–2 weeks
  2. 02

    Definition & Prototype

    An evaluation dataset is prepared; the approval screen is tested with real users in week 3.

    2–3 weeks
  3. 03

    Build

    Weekly working-software demos; every change is re-evaluated against the dataset.

    Two-week sprints,
    first release target: weeks 6–12
  4. 04

    Grow & Operate

    We monitor correction rates and costs, then add new scenarios.

    Ongoing

This is an illustrative schedule assessed once access and pilot scope are clear; the final duration is determined during discovery. Step-by-step details: how we work

Budget and timeline

We do not publish a price list; scope determines cost and duration. Model usage is a separate cost item that varies with volume and is charged to your account.

Data preparation
Fragmentation, example answers and masking.
Integrations
Number of systems and API availability.
Approval flow
Who approves, and at what threshold.
Evaluation
Dataset size, metrics and acceptance threshold.
Model usage
Monthly volume, document length and selected model.
Security and compliance
KVKK, data location and your own server.
What determines the cost of custom software?
Discovery sprint
2 weeks Fixed scope, fixed fee
First release
6–12 weeks A first release scoped together: one scenario end to end
New scenarios
In stages Each is added after a separate evaluation

Delivery terms

Delivery and maintenance details

You can switch providers and remain independent of us: you hold the code, prompts and records.

  1. Code and prompts

    Transfer of project-specific code and prompts is defined separately from third-party model licences in the contract.

  2. Evaluation datasets

    Examples, correct answers and measurement scripts are in your repository.

  3. Models

    Models trained specifically for you are held in your account.

  4. Accounts

    Cloud, model and repository accounts are in your name; call and cost limits are defined in the application.

  5. Documentation

    Data flows, masking and approval points are documented.

  6. Handover and support

    Handover through joint sessions; model monitoring, maintenance scope, support hours and new scenarios are agreed separately.

Which deliverables support a decision at the end of a pilot?

A working demo alone does not justify live use. For a single-scenario pilot, we agree how to prepare these outputs alongside the scope:

  • Comparison: the current method and proposed flow are evaluated against the same held-out test examples, including errors, rejected records and human corrections.
  • Access and approval: document the sources the model can read, operations it can perform and steps requiring human approval; test prohibited-access scenarios too.
  • Operations: record total cost per operation, call and retry limits, the stopping procedure and the responsibility of the person who continues manually.
  • Handover: include the model and prompt versions, evaluation scripts and outstanding risks in delivery. Model changes go through the same checks.

Example scenario: if a model prepares document drafts quickly but misreads critical amounts, do not expand the pilot. First assess the correction workload and cause of errors. Instead of promising a success rate or savings, we define pilot acceptance and stopping criteria together.

Frequently asked questions

What should we prepare to discuss an AI pilot?

Describe one task you want to improve, how it is done today and the expected output. Share the known frequency, unacceptable errors, the role checking the output and any limits on where data may be processed. You do not need to invent durations or success rates you have not measured.

Do not send a real document archive or access key at first contact. We will agree sample data and sharing conditions during the discussion. The pilot evaluation guide shows how baseline measurement, test data, human review and stopping decisions are distinguished.

What is the difference between an AI agent and rules-based automation?

In rules-based automation, steps are predefined: a document arrives, fields are extracted, rules check them and a person approves. In an AI agent, the model chooses which tool to use and in what order to achieve a goal. We assess a fixed workflow first, scoping an agent only if variable steps require one. Accessible data and tools are restricted; external messages, payments and record changes require human approval.

How are usage costs and repeated operations limited?

We define call, time and output limits per operation, concurrent-job limits and daily usage quotas. Retries use increasing intervals and a limited number of attempts; persistently failing jobs move to human review. Receiving the same job ID again does not create another side effect. A provider budget alert may not stop spending by itself, so the application's stop rule and the procedure for disabling model access are tested too.

Does every task use a large language model (LLM)?

No. Language models are useful for free text and documents. For demand, consumption and cash forecasting or anomaly detection, time-series methods and traditional machine learning are often more accurate, cheaper and easier to explain. A step that can be handled by a rule never reaches the model.

How is accuracy measured, and how are errors caught?

We do not promise an accuracy rate in advance. We measure your current processing time, error rate and human review time for comparable work, then compare the pilot on a held-out test set. A model's confidence is not validation: critical fields are assessed with rules and human review. Incorrect, rejected and reprocessed records remain in the measurements. Pilot success criteria consider total cost per operation together with the decision to expand, revise or stop.

Will our data be used to train the model?

It should not be; we verify this together in the provider's contract: are inputs retained, for how long, and are they used for training? If data must never leave your company, an open model can run on your own server or in your cloud account.

Will we become dependent on one model provider?

Model calls pass through one layer; you retain prompts, evaluation datasets and records. The model version is pinned so a provider update does not change behaviour without notice. When a provider retires a version or a new model is tried, the candidate is measured against the same dataset and deployed only if the result is sufficient.

Does an internal assistant expose every document to everyone?

No. An internal assistant or chatbot follows users' existing file and system permissions; it does not search or use documents they cannot access. Answers with sources point to the relevant document; if there is no source, the system says so. Logging scope, masking and retention periods are defined separately.

Can we see your reference projects?

We present product concepts instead of a customer reference list: interface designs for fictional companies showing how a problem could be solved. To limit risk, we recommend a single-scenario pilot, measuring the model against an evaluation dataset prepared from your data before going live.

Next step

Tell us about the task you want to improve and the data you have. We will assess your request and clarify the next step for a discussion, data review or limited pilot.

Technical research

A measured starting point for a document pilot

We compared two local OCR models on 24 synthetic documents using the same rules. Accuracy, error types and processing times, with downloadable data and code.

Read the methodology and results of the Turkish document extraction test

When assessing a pilot's economics, calculate using your current workload and automation assumptions. Include remaining human review time in the reduction rate, and model and maintenance expenses in monthly operating cost; not all time released is a cash saving.