Skip to content

From source data to checked documents

Gather data from your systems, check it against rules and produce documents from templates, with approval and source control.

Next: a short form and a discussion of your task. No files need to be uploaded now.

Process schematic

Data → document → outcome

  • Source of record: every field has an owning system, and it wins on a conflict
  • No invented terms: prices, deadlines and volumes come from the source, never phrased by the model
  • Mandatory-field validation before assembly — a gap halts the document
PDFDOCXXLSX and CSVHTML and e-mailAPI payload
A process schematic, not a screenshot of a finished system. Built for your workflows.

One client report: before and after

Manual process

9 steps
  1. Open the first source and export the data
  2. Open the second source and export the data
  3. Merge the exports into a spreadsheet by hand
  4. Check that periods and metric definitions actually match
  5. Copy the values into the document template
  6. Write the findings and recommendations
  7. Send it to the owner for review
  8. Apply the corrections and rebuild the document
  9. Send it to the client and remember to log it in the CRM

Automated process

4 steps
  1. Data is collected from the sources and checked for completeness
  2. The document is assembled from the template with current values
  3. The owner adds the findings and approves
  4. The document goes to the client, and the CRM record updates itself

This shows the number of process steps, not time saved: any claim about effect requires a fixed baseline, a measurement window and a stated method, and is published only alongside a case.

A single document is assembled by hand from several systems — and the whole process rests on whoever knows where everything lives.

Cuts the time spent producing documents and reports and moving data between systems.

Capabilities and details

Data is pulled from your systems, checked against rules, placed into the template and delivered — with mandatory-field validation and a hard block on inventing terms that are not in the source.

Collects data from its sources, checks it, and produces the document from your template.

  • PDF
  • DOCX
  • XLSX and CSV
  • HTML and e-mail
  • API payload
  • extraction and normalisation of data from inbound documents

What usually gets in the way

  • The data for a single document sits across several systems
  • It is copied into the template by hand
  • Versions drift apart and nobody knows which one is current
  • Delivery slips because it waits for approval
  • The process rests on the one person who knows where everything lives
A close-up of a binder packed tight with documents.

Which documents

  • Proposals
  • Client reports
  • Contract packages
  • Price list normalisation
  • Instructions and internal standards
  • Invoices and accounting documents
  • Internal management summaries

How it works

The route from input to outcome

Nodes expand: input, processing, output, control.
  1. 01Source

    Sources

    Expand
    Input
    CRM, ERP, spreadsheets, ad platforms, price lists.
    Processing
    Data is collected according to the document definition.
    Output
    A field set where every value carries its origin.
    Quality control
    Every field has a source of record: on a conflict it wins, not whatever was found first.
    Integration
    CRM, ERP, databases, external APIs
  2. 02Business rule

    Validation

    Expand
    Input
    The collected field set.
    Processing
    Mandatory fields, formats, ranges and internal consistency are checked.
    Output
    Either a set ready for assembly, or a list of problems.
    Quality control
    A missing mandatory field halts assembly. No document goes out with a gap in it.
  3. 03AI step

    Transformation

    Expand
    Input
    Validated data plus unstructured fragments: descriptions, comments, findings.
    Processing
    Values are normalised; text blocks are written strictly from the data.
    Output
    The document's content.
    Quality control
    Commercial terms come only from the source. Below the confidence threshold a block is flagged for manual editing.
  4. 04Automated step

    Generation

    Expand
    Input
    The content and your template.
    Processing
    Assembly in the required format, preserving your brand formatting.
    Output
    A PDF, DOCX, XLSX, message or API payload.
    Quality control
    The template version is recorded with the document — you can see which version produced it.
  5. 05Human checkpoint

    Approval

    Expand
    Input
    The finished document and the list of anything that needed attention.
    Processing
    The owner reviews, edits and approves.
    Output
    An approved version.
    Quality control
    Mandatory for anything carrying commercial terms. Who approved it and when is logged.
  6. 06Target system

    Delivery

    Expand
    Input
    The approved version and its recipient.
    Processing
    Sent by e-mail, uploaded into the client's system, or passed over an API.
    Output
    The document with its recipient, and a delivery record in the CRM.
    Quality control
    A failed delivery is not counted as done — the event moves to the exception queue.
    Integration
    E-mail, CRM, document storage
  7. 07Monitoring

    Archive

    Expand
    Input
    The delivered version and all of its data.
    Processing
    The version, the origin of each value and the approval history are stored.
    Output
    A reconstructable history: what was sent, when, and on what basis.
    Quality control
    Retention and deletion follow an explicit policy rather than accumulating indefinitely.

Feedback: monitoring returns data to the rules — the loop closes rather than ending at the last step

Fact control

The main risk in automated document generation is an invented term that exists nowhere in the source. Five mechanisms close exactly that.
Hands with a pen over a printed contract, turning a page before signing.
Header fields
Source of record: every field has an owning system, and it wins on a conflict
Mandatory field
Mandatory-field validation before assembly — a gap halts the document
Text block
A confidence threshold: below it, the block is flagged for manual editing
Terms block
No invented terms: prices, deadlines and volumes come from the source, never phrased by the model
Signature
Human approval for anything carrying commercial terms
An interface schematic, not a screenshot of a running system. The empty slots are deliberate: they are filled from your own sources.

What it needs

  • deal and contact records from the CRM
  • ERP and accounting data
  • price lists and commercial terms
  • ad platforms and analytics
  • spreadsheets and exports
  • previous document versions

What it produces

  • proposals
  • client reports
  • instructions and internal standards
  • contract packages
  • invoices and accounting documents
  • internal management summaries
  • normalised price lists

Output formats

The approved version

  • PDF
  • DOCX
  • XLSX and CSV
  • HTML and e-mail
  • API payload

To the recipient · Delivery logged in the CRM

An interface schematic, not a screenshot of a running system: the product is assembled around each client's own stack, so there is no typical screen.

What can be measured

Categories, not promised values. Numbers appear only in cases, with a baseline and a stated method.
  • time to produce a document
  • share of documents assembled without a manual step
  • rework caused by errors and discrepancies
  • time from request to delivery
  • share of reports delivered on schedule

What it connects to

A specific service is named once the integration is verified.
  • CRM
  • ERP and accounting systems
  • databases
  • e-mail
  • document storage
  • Google Workspace / Microsoft 365
  • advertising and analytics APIs

Where the human stays

A document with commercial terms goes out only after approval. An unfilled field is flagged, not guessed.

Cases

Where this has been applied

  • Marketing and digital agencies · Document & Reporting Automation

    Client under NDA

    The first ten days of the month stopped going into client reports

    Time to prepare a monthly report
    10 days → 5 minutes
    Agency figures: a before-and-after measurement of the process
    Projects handled by the same team
    +30%
    Agency figures: project count compared at unchanged headcount
    Case details

    For the first 8–10 days of the month the team assembled reports instead of doing project work. Approvals dragged, payments moved with them, and every new client meant either overtime or a hire.

    Monthly saving on manual work
    ≈$5,000
    The agency's own estimate from their rates: we saw neither the timesheet nor the calculation, so this is an estimate rather than a measurement
    Read the case
  • EdTech · Document & Reporting Automation

    Client under NDA

    Course materials stopped going stale faster than they could be updated

    Speed of releasing and updating material
    ×10
    Client figures: measured on preparing material against an existing course standard, not on designing a syllabus from scratch
    Share of manual work in content production
    −50%
    Client figures: their own time records for the methodology team
    Case details

    A new module or an update took weeks. The load on methodologists grew faster than the team, and the product aged between releases: a student would see an example on a library version that no longer exists.

    Cost of producing content
    −40%
    The client's own cost model; the composition of the costs was not disclosed to us
    Read the case

Frequently asked questions

Can the system invent terms that aren't in the contract?

That is the first thing we close off. Commercial terms — prices, deadlines, volumes — are taken only from the source of record and are never phrased by the model. Mandatory fields are validated before assembly, and a missing one halts the document. Where confidence falls below the threshold, the block is flagged for manual editing.

Do we have to redo our templates?

Usually not — we work with yours. Rework is needed where a template doesn't say where a value comes from: if the same figure is calculated differently across documents, it can only be automated once a single definition is agreed. Those spots surface during the assessment.

Which output formats are supported?

PDF, DOCX, XLSX and CSV, HTML for e-mail, and an API payload for handing data to another system. The format follows what the recipient actually does with the document: if the numbers get retyped into a spreadsheet afterwards, PDF is the worst of the options.

Let's look at this workflow on your own material

An example of your report will help us discuss sources, checks and approval. We'll check whether the economics work and name the next step.