ITF Digital

Agents for the reading and assembling

Software that does the repeated reading, checking, chasing and pack assembly, and hands a named person the decision.

First agent live
Three to five weeks from signed specification
Approval
A named human signs anything consequential
Audit trail
Every run logged with its inputs and output
Kept with a person
Conformity calls, payments, regulated decisions

An agent here is a scheduled program that reads documents, checks them against your own rules, drafts the output and routes it to a named person for approval. It removes the hours a person spends assembling what somebody else will sign.

A six step agent run running top to bottom: trigger, read and extract, check against rule, draft, a named approver drawn as the live gate, then file and log. A branch sends anything ambiguous to an exception queue, and a return path sends a rejected draft back. An inset shows one entry of the run log.TRIGGERSCHEDULE, INBOUND MAIL OR FILE DROPREAD AND EXTRACTFIELDS PULLED, AMBIGUITY MARKEDCHECK AGAINST RULEREGISTER, DEADLINE, CRITERIADRAFTYOUR TEMPLATE, SOURCES LINKEDNAMED APPROVEREVERY ISSUED OUTPUT PASSES HEREFILE AND LOGOUTPUT FILED, RUN WRITTENEXCEPTIONBACK TO A PERSONRETURNEDEVERY STEP IS WRITTEN TO THE RUN LOGRUN LOG / ONE ENTRYINPUT1 PDF, 2 PHOTOSRULE SETREV 4EXCEPTIONS1 RAISEDDRAFTEDAUTOMATICAPPROVED BYA NAMED PERSONFILED TOTHE RECORDTHIS IS WHAT AN AUDITOR ASKS FORTHE GATE IS DESIGNED INPART OF THE RUN FROM THE FIRST BUILDHANDED TO A NAMED PERSONCONFORMITY CALLSMONEYDECISIONS
One agent runThe approval gate is the point of the drawingSwipe to pan
  1. 01 Enquiry
  2. 02 Systems audit
  3. 03 Specification
  4. 04 Build
  5. 05 Pilot
  6. 06 Deploy
  7. 07 Handover
  8. 08 Support

Agents for the reading and assembling sits at step 4 of the build cycle. You receive a working increment you can use.

01

Why most agent pilots quietly stop

The demo works. Six weeks later nobody is using it.

It usually fails for one of three reasons. The agent was pointed at a task nobody actually repeats, so the saving was imaginary. Or it produced output that still needed rewriting, which is slower than writing it once. Or it acted on something it should have asked about, and after the first bad call the team stopped trusting it.

The fix is boring. Pick a task that runs weekly or more. Define what good output looks like using examples you already have on file. Put a person on the approval, and log every run so a disagreement can be settled by reading the log instead of by argument.

02

What the agent is allowed to touch

Scope is written down before anything is built, because the boundary is the part that gets an organisation in trouble.

CapabilityWhat it means in practiceWho signs
Read and extractPull structured fields out of emails, PDFs, forms and photographs of paperwork.No signature needed
Check against a ruleCompare what it read against your own register, deadline or acceptance criteria, and flag the mismatch.No signature needed
ChaseRemind the owner before a due date, escalate on a schedule you set, and record that it did.No signature needed
DraftProduce the report, the response, the pack or the letter in your format.Named reviewer
IssueSend, file or publish the output to a client, a regulator or a register.Named approver, every time
DecideFindings, conformity calls, money and anything regulated go to a competent person, with the evidence the agent gathered attached.The accountable person
03

What you get to keep

01

The written scope

What the agent reads, what it may do with it, and who approves what. Signed before the build.

02

The run log

Every run with its inputs, its output and its approver. This is the document that answers an auditor asking how a record was produced.

03

The prompt and rule set

In plain text, in your repository. You can read it, change it and take it to somebody else.

04

The exception report

What the agent could not handle and passed back to a person, counted weekly. It is the honest measure of whether this is working.

05

The operations manual

How to turn it off, how to change a rule, and who to call. Written for whoever inherits it.

04

One run, start to finish

  1. 01

    Trigger

    A schedule, an inbound email, a form submission or a file landing in a watched folder. Not a person remembering.

    Automatic
  2. 02

    Read and extract

    Fields pulled out and normalised. Anything ambiguous is marked as ambiguous rather than guessed.

    Seconds
  3. 03

    Check

    Compared against the register, the deadline or the criteria. A mismatch becomes an exception, not a silent pass.

    Seconds
  4. 04

    Draft

    The output assembled in your template, with every claim linked back to the record it came from.

    Under a minute
  5. 05

    Approve

    The named person opens the draft, sees what changed and what was flagged, and signs or sends it back.

    Human, always
  6. 06

    File and log

    Output filed where it belongs, run written to the log, exception counted.

    Automatic
05

Client audit reports, retyped by hand

A facilities contractor in Riyadh receives client audit reports as PDFs, each with ten to thirty findings, and its QHSE coordinator retypes every one into a spreadsheet.

The agent watches the mailbox the reports arrive in. It extracts each finding with its clause reference and the severity the client wrote, proposes an owner from a rule that maps areas to supervisors, and drafts an acknowledgement letter.

The QHSE manager approves the letter and the owner list before anything is sent. The agent then chases each owner on a reminder ladder the manager signed off, and attaches the photographs they upload to the right finding. When every finding in a report has evidence, it assembles a closure pack for the manager to review.

The contractor holds the signed scope, the run log, the rule set in plain text, a weekly exception report and an operations manual that starts with how to switch the agent off.

06

Back-office jobs agents handle well

Each one has the same shape: read, check, draft, and hand the result to a person.

Audit findings that stop living in email

A finding arrives as a paragraph in a report. Somebody retypes it into a spreadsheet, somebody else chases the owner, and the closure evidence ends up in a folder nobody links back to the finding. The agent does the retyping, the chasing and the linking.

  • Findings parsed out of a report into a register with owners and due dates
  • Root cause prompts put in front of the owner instead of a blank box
  • Evidence attached to the finding it closes, so the closure pack assembles itself
  • Effectiveness review raised on a date, not when somebody remembers
  • The finding marked closed by the person accountable for it
07

Questions worth asking before you buy an agent

Q01

Who makes the decisions?

Your people. Agents chase, read, check, assemble and draft. Approvals, conformity calls, money and anything regulated go to a named person, and the scope names that person before the build. If a supplier tells you their agent can close a corrective action on its own, ask them who signs it.

Q02

What happens when it gets something wrong?

It is designed to be wrong visibly. Anything ambiguous becomes an exception and goes to a person, and every run is logged with its inputs and its output. You will see the error rate in the weekly exception count, which is deliberately a number we cannot hide from you.

Q03

Where does our data go?

Each step of a run either executes inside your own environment or calls an external model. The specification maps every step: where it runs, which model provider handles it and which classes of document pass through it. Where a class of document has to stay inside your environment, the steps that read it run there, and the specification records what that means for capability.

Q04

What drives the cost?

Four things: how many distinct document formats the agent has to read, whether your rules are already written down or need drafting with you, how many systems it has to write into, and how much of your own review time the pilot needs. A single well-defined task against one document format is the cheap end. We publish no price list because the honest number comes out of the systems audit, not out of a page.

Q05

Can we take it over ourselves?

Yes, and the handover assumes you will. Prompts and rules are plain text in your repository, the data model is documented, and the operations manual is written for your administrator. A support agreement is something you buy because it is useful, not because you cannot leave.

Q06

Can it read Arabic documents?

Yes, including documents that mix Arabic and English. We test on your own documents at specification and record what it reads well and what it reads poorly.

Q07

Do you need our historical data to start?

Examples, not history. Ten good examples of the output you want tell us more than a year of records, because the work is teaching the agent your standard rather than training a model. If you have no examples at all, the first week produces them with you.

Book a systems audit

Bring the task that eats a person's week. The audit produces a current-state process map and an honest answer on whether an agent is the right tool for it.

Book a systems audit