Review my process
DerKIArchitekt / Navigation
Appearance
Language
Motion
About meGuidesProcess first.
Then the architecture.
Illustrative architecture / Document processing

From document to a checked CRM record.

Read, check and correctly match details from known document types. Unclear values are reviewed manually.

Example architecture, not a client project.

ARCHITECTURE: DANIEL FRANKUPDATED: 7 SEPTEMBER 2026RAW / RULE / AI / HUMAN
01Starting point

What is done manually today.

Documents arrive as files or email attachments. A person finds the relevant information, copies values and matches the file to a CRM contact. Different formats and missing identifiers make this time-consuming.

The example is limited to known document types and defined target fields. It does not describe universal document AI. Unknown or unreadable files need different handling.

02Target state

One case.
One clear path.

Only checked details are matched to an unambiguously identified CRM record. The original remains accessible through a reference. Unclear information receives a visible review task.

RESULT

Validated record + unique document reference

03Trigger

A file reaches an authorised intake channel. Its source, format, size and document reference are checked before processing.

04Input data
  • Original document from the agreed source.
  • Known document types and required fields.
  • CRM identifiers and matching rules.
  • Checking rules for values, totals and required details.
05The example workflow

Every step has a purpose.

EXAMPLE WORKFLOW / INTERACTIVESelect a process step
  5. 05 / PROCESS STEP

    Match entity

    RULE
    Input data
    Validated identifying details
    Action
    Finds the matching company, person or transaction.
    Checks
    Unique identifiers take precedence over name similarity.
    If something fails
    Ambiguous matches require human assignment; no speculative automatic creation.
05 / PROCESS STEP

Match entity

RULE
Input data
Validated identifying details
Action
Finds the matching company, person or transaction.
Checks
Unique identifiers take precedence over name similarity.
If something fails
Ambiguous matches require human assignment; no speculative automatic creation.
RESULTValidated record + unique document reference

Example architecture, not a client project. The implementation depends on your process and available system access.

06Decisions & details

What happens at each step.

01

Document intake

RAW
Input data
File or attachment from a defined source
Action
Captures the document with a unique reference.
Checks
File type, size and permitted source are checked before processing.
If something fails
Unsupported or damaged files enter a manual queue.
02

Classify

AI
Input data
Readable document content and permitted document classes
Action
Assigns the content to a known document type.
Checks
Only agreed classes are permitted; unknown types are not forced into a category.
If something fails
An unclear document type is classified by a person.
03

Extract

AI
Input data
Document content and field schema for its type
Action
Extracts relevant details with source references.
Checks
Missing values stay empty. The model must not fill in missing master data.
If something fails
Illegible or conflicting details are flagged for review.
04

Validate

RULE
Input data
Extracted fields with source references
Action
Checks required fields, data types, value ranges and relationships.
Checks
Totals, dates and identifiers are checked against fixed rules.
If something fails
Invalid records do not reach the CRM.
05

Match entity

RULE
Input data
Validated identifying details
Action
Finds the matching company, person or transaction.
Checks
Unique identifiers take precedence over name similarity.
If something fails
Ambiguous matches require human assignment; no speculative automatic creation.
06

Map fields

RULE
Input data
Validated data and an unambiguous CRM entity
Action
Maps source fields to the defined data model.
Checks
Field ownership, format and permitted overwrites are predefined.
If something fails
A schema change stops the affected transfer and produces a reviewable error.
07

Save

RULE
Input data
Validated field mapping and document reference
Action
Writes the data once to the target system.
Checks
Checks the target response and reconciles the saved record.
If something fails
Temporary errors receive limited retries. Already saved data is detected.
08

Review exception

HUMAN
Input data
Open cases, cause and original document
Action
The responsible person corrects or rejects the exception.
Checks
Correction and restart remain linked to a specific case.
If something fails
Unresolved exceptions stay visible and are escalated after the agreed period.
07Ausnahmen

Ausnahmen

  • Damaged or unreadable files are referred for manual handling.
  • Unknown document types are not forced into a convenient category.
  • Conflicting details are flagged rather than imported without checking.
  • Where several contacts may match, a person decides the match.
08Checks

Checks

  • Format and data type are checked, followed by the business information.
  • Where possible, values are compared with the original and master data.
  • Existing fields are overwritten only according to agreed rules.
  • The CRM response is checked; repeated files must be recognised.
09Systems involved

Systems involved

An authorised file or email intake and a CRM with suitable access. Additional storage is needed only if statuses and document references cannot sensibly be maintained in the existing systems.

10What we measure

What we measure

  • Manual time per document, including rework.
  • Accuracy of imported fields against suitable test cases.
  • Manual reviews grouped by cause.
  • Duplicate files, incorrect matches and pending review tasks.
11What to clarify before operation

What to clarify before operation

Instructions inside the document remain document content. They must not change recipients, permissions or workflow steps. Reading, checking and writing are separate tasks.

Access, retention and deletion of originals are defined. If the CRM data model changes, the integration must be checked before more data is written.

The metrics show what could be measured in a future project. They do not describe achieved client results or guaranteed savings.

Assess data quality and suitabilityHow AI results are checkedHow an automation project is scoped

Does this sound like your process?

Review my process

Describeyour project.

A workflow, a website or an internal application: I assess the possibilities.

Project enquiry