What does using AI in a process mean?
Integrating AI into business processes means connecting a model to your company's systems to support a defined task. It takes inputs from an inbox, document archive, or ERP, performs a specified job, and writes results back into those systems, often for human approval. What data it can see, what it may change, and how each step is logged are defined upfront. We describe scope and delivery conditions on our AI integration page.
Buying enterprise licenses for a chat tool is different work. A license helps employees write and summarize faster individually; the process stays the same. If individual productivity is the goal, licensing is often enough and cheaper. Integration changes the process itself: order entry, routing a request to the right team, or feeding forecasts into planning.
“AI” also does not mean only large language models. Language models stand out for text and document tasks; for demand forecasting and anomaly detection, conventional statistics and machine learning are often more suitable, cheaper, and easier to explain.
Six scenarios in B2B operations
For each scenario, be clear about what the model receives, what it produces, and who retains the decision.
Draft orders from email and PDF orders
Input: A customer's order in free text or an attachment.
Processing: The model reads the text and attachment and matches product names to the catalog.
Output: An ERP order in “draft” status, with uncertain fields flagged.
Human approval: A customer representative reviews and approves it.
Required data: Historical orders, product catalog, and product names used by customers.
Document reading: dispatch notes, invoices, certificates
Before selecting a model, first define ground truth for fields and how errors will be measured. In the OCR study of 24 synthetic documents, downloadable examples and raw outputs show how a misread label becomes a missing field. This small test does not replace a pilot on your own documents.
Input: A scanned or PDF document.
Processing: The model extracts fields; rules compare them against the purchase order.
Output: Extracted line items and a list of differences.
Human approval: Accounting or the warehouse approves discrepant items.
Required data: Example documents and corresponding orders. e-Invoice and e-Dispatch Note already arrive as structured XML (UBL-TR); AI is not needed to read them.
Request classification and draft responses
Input: Support emails and portal requests.
Processing: The model classifies topic and urgency and drafts a response from current procedures.
Output: A request routed to the relevant team and a draft response citing sources.
Human approval: The response is sent only when the representative reads and sends it.
Required data: Labeled historical requests and current procedures.
Demand, consumption, and cash flow forecasting
Input: Historical sales, consumption, or collection data; known factors such as promotions, holidays, and production schedules.
Processing: A time-series model forecasts using historical patterns and these factors.
Output: A forecast and confidence interval by product, facility, or account.
Human approval: The planning or finance team uses the forecast or adjusts it with a reason.
Required data: Consistently maintained history, spanning several years if there is seasonality. Compare the forecast against today's simple method, such as the same week last year; do not launch it if it cannot outperform that baseline.
Permission-aware document search and internal assistant
Input: Specifications, contracts, procedures, and an employee's question.
Processing: The system finds relevant passages in documents the user is authorized to access; the model writes the answer only from those passages.
Output: A short answer or specification summary identifying the source document and page; “not found” if no source exists.
Human approval: The answer is informational; quotation and contract decisions remain with the person.
Required data: Current document versions and access permissions.
Anomaly alerts
Input: Meter, payment, price, or stock movements.
Processing: The model learns expected behavior for each source and measures deviations.
Output: A record deviating from expected behavior and an alert to the relevant team.
Human approval: The reviewer marks the alert as “real” or “false alarm”; these labels improve the model.
Required data: Historical records of normal periods and incidents that occurred.
Each scenario needs a one-sentence metric: the share of drafts or documents approved without correction, requests routed to the right team first time, forecasting error compared with the current method, answers verified against their sources, or false alarms. The Industrial Energy Monitoring product concept shows how consumption forecasts and waste alerts could appear on an energy team's screen, using the fictional company X Enerji.
What happens if an employee's access to document search is revoked?
Access control cannot be left to an instruction given to the model writing the answer. Identity must be verified on the server, and only authorised document passages should reach the model. Define separately how the search index, cache and links in previous answers will be updated when permissions on a source file change. If current permissions cannot be verified, the safe default is not to generate an answer from that document.
For example, a sales employee may access the current product specification but not a confidential agreement in the finance folder. During pilot acceptance, ask the same question as two different users; check that the document title, quotation and download link respect the access boundary as well as the answer. Then remove group membership and try again. Measure the maximum delay before revocation takes effect, document deletion behaviour, and whether shared caches prevent data from passing between users. These are example test conditions, not results from a completed client project.
Preparation requires document owners, source systems, a list of groups and roles, confidentiality classifications, update frequency and permitted sample questions. A custom assistant may be unnecessary if existing enterprise search meets the need. The number of sources, complexity of the access model and time needed to update permissions affect development and maintenance scope.
Product behaviour depends on the system used: Microsoft's document-level access control documentation explains permission filtering at query time; the example of transferring SharePoint permissions into a search index notes that permission changes may not propagate automatically to downstream systems. These examples do not imply a product partnership or a commitment to ready-made integration with every system.
How do you choose the first scenario?
The first scenario should be measurable and allow errors to be caught, rather than simply being the most exciting. Three questions are enough:
- Volume. Does the work recur every day? High volume makes both benefits and measurement possible.
- Cost of errors. What happens if the model is wrong? An error in a draft order can be caught during approval; an incorrect price sent automatically to a customer cannot.
- Data access. Are the inputs digital and accessible, and are there historical examples of correct answers?
Fixed workflow or agent?
In fixed automation, you define the steps: an email arrives, the model extracts fields, rules validate them, a draft is created and a representative approves it. The model runs at a single step; the workflow can be tested and its cost predicted. With an agent, the model decides which tools to use and in what order. This is powerful when steps cannot be specified in advance, but behaviour is harder to test and cost varies per task. Our recommendation: build the first scenario as a fixed workflow. Move to an agent only when the steps truly cannot be defined in advance; even then, grant read access and require approval for every action that changes records.
When data is missing or scattered
If orders sit in personal inboxes and correct answers are not recorded anywhere, the first step is to bring the process into one system, rather than start an AI project. We cover this in our article on moving from Excel to an operations platform. Every well-maintained record becomes evaluation data for the next step.
AI or rules?
If a rule can be written, that rule is cheaper, faster and easier to audit.
| Task | More suitable approach |
|---|---|
| Send payments above a specified amount for a second approval | Rule. The threshold and approval chain can be defined. |
| Read fields from e-invoices and electronic delivery notes | Read XML directly. The data is already structured. |
| Send an alert when stock drops below a critical level | Threshold. A threshold table if it varies by season. |
| Extract orders from emails written differently by each customer | AI. The format cannot be known in advance. |
| Identify deviations in consumption that varies by shift and season | Statistical model. A fixed threshold generates too many false alarms. |
Some tasks should not be automated: decisions with adverse consequences for individuals, such as applicant screening, credit refusal or performance reviews; irreversible actions, such as payment instructions or prices sent to customers; and work whose output no one can check. In these cases, a model should at most make recommendations. Article 11 of Türkiye's Personal Data Protection Law (KVKK) also gives individuals the right to object to adverse outcomes arising from analysis of their data exclusively by automated systems.
Data and KVKK questions
In a scenario involving personal data, the data flow must be clear from the outset. Two publications from Türkiye's Personal Data Protection Authority provide a useful starting point: Recommendations on Personal Data Protection in the Field of Artificial Intelligence and Guide to Generative Artificial Intelligence and Personal Data Protection (15 Questions) (checked: 11 October 2026). According to the guide, when personal data is transferred through a provider established abroad, the transfer must comply with Article 9 of the Law and the regulation on transfers abroad published in 2024.
Questions for the provider and your team
- Which fields are sent to the model? Is unnecessary personal data masked before sending?
- In which country is the data processed? If it is transferred abroad, what is the legal basis?
- Does the provider retain inputs and outputs, and for how long? Are they used for model training?
- Are the data controller and data processor roles defined in the contract?
- Do search and assistant functions respect users' existing permissions?
- Can transactions and model versions be audited? If raw inputs and outputs must be logged, are masking, access and retention periods defined?
- Can the same task be performed without personal data?
This list is not legal advice; assess every scenario involving personal data with your legal adviser.
Pilot, costs and risks
Example scenario: we want to create draft orders from emails and PDFs received by a spare-parts distributor. This plan is not a completed client project or a measured result. The pilot is limited to product code, unit and quantity; pricing decisions and final order confirmation are outside the model's authority:
- Baseline measurement. Record the time and corrections required for a representative to have an order accepted correctly by the ERP using the current method. Compare the pilot on work with the same distribution of difficulty.
- Separate test set. Prepare answers verified by the business owner; include examples with missing product codes, box/unit distinctions, illegible quantities and duplicate emails. Do not mix documents used in development into the final test set; mask personal data where required.
- Acceptance rule. The product, unit and quantity must match the source document; missing fields must be referred to the representative rather than guessed. In this example, a document belonging to the wrong customer or a confirmed order created without human approval is a failure that blocks acceptance. Set time and cost thresholds after measuring the baseline.
- Human review. The representative sees the source document alongside the draft. Corrections, rejections and approvals are recorded separately. A repeated email does not become a new order; model output enters the workflow only with validated fields and an authorised person's approval.
- Cost and stopping conditions. Calculate total cost per accepted order, including failed calls and review effort. When the per-task call limit or daily quota is reached, new model tasks stop and requests join the existing manual-entry queue. The pilot owner controls restarting.
A controlled test covers more than the successful path: test attempts to bypass approval, documents from the wrong customer, duplicate submissions and quota overruns too. A pilot that cannot stop safely in these conditions should not be expanded, even if its average accuracy is high.
How do we know whether an AI pilot is successful?
Measure today's method first. On a sample with comparable document types and workloads, record elapsed time from arrival to completion, staff time and errors requiring correction. Reserve a test set that is not used during development, rather than repeatedly improving the same easy examples. Include difficult cases such as missing documents, illegible fields and incorrect product codes.
| Criterion | How is it compared? |
|---|---|
| Completion time | Measure queue time, human approval and ERP acceptance alongside the model response. Do not let delayed tasks disappear into an average. |
| Errors and corrections | Track the proportion of all reviewed tasks that contain errors or omissions. Assess critical errors, such as an incorrect quantity, separately from spelling mistakes. |
| Human review workload | Add up staff time spent reading, validating, correcting and resolving exceptions. Moving review work to another team does not count as a saving. |
| Cost per accepted task | Divide model and tool calls, retries, operations and human review costs by the number of successfully accepted tasks. Costs of rejected tasks remain in the total. |
Show one-off development and data preparation costs separately from this operating calculation. Adapt the results to your own volume; do not turn unachieved time savings or accuracy into sales promises. Automated quality metrics can support review but cannot by themselves replace human judgement for critical business decisions. Microsoft's evaluation metrics documentation explains using different quality and safety checks for different tasks (checked: 13 October 2026).
Scale, improve or stop
- Scale: Expand scope in small steps if the agreed time, error, human review and cost limits are met, and the permission and stop tests have passed.
- Improve and measure again: If the issue concerns a particular document type or data quality, narrow the scope and evaluate the change on the reserved test set.
- Stop or return to the previous workflow: Do not expand automation if there are critical errors, unauthorised data access, uncontrolled usage costs or unresolved review workload. Assign an owner for returning to the baseline method.
What determines the cost?
Most providers charge for models according to the amount of text processed; monthly cost depends on volume, document length and the model selected. Total cost also includes data preparation, an evaluation set, ERP and email integration, monitoring and maintenance. To control model spending, send only the necessary content, select the smallest model that passes the evaluation set, batch non-urgent tasks where providers offer lower prices, avoid sending cases that rules can resolve, and review the provider's budget tools. A budget alert does not always stop spending: the application also needs per-task call and time limits, concurrency limits, a daily quota and a rule that stops work when necessary. Limit retries; running the same task twice must not create another payment, email or record.
Risks and precautions
- Incorrect output. A model can sound confident and still be wrong: use human approval, flagged uncertainty and source references.
- Model changes. A provider update may change behaviour: pin the model version and rerun the evaluation set for every change.
- Provider dependency. Keep model calls in one layer; retain ownership of prompts, evaluation sets and logs.
- Instructions within the input. An email may contain instructions aimed at the model: let it read only, and require rules and approval for every step that changes records.
We explain the scope, process and delivery terms for this work on our AI integration service page. If you are unsure which scenario fits your business, a discovery sprint answers that question by examining your data.