Skip to main content

How we think about the work.

What to expect from AI and from the software you pay for, in a few short essays, each followed by a working example you can try in your browser.

Two short articles, then three working examples you can try on sample data.

What an owner should demand from an AI system

Five questions to ask before AI touches your company's data or decisions.

Most AI projects are judged on the demo, and the demo is the easy part. The real test is a normal working day, when the system reads a document it has never seen and nobody is checking.

First, ask where your data goes. Not in general, but for this system: which provider, which region, which account, and whether anything you send can train someone else's model.

Second, ask who approves. An AI system should propose and a person should decide, at least wherever money, personal data or a customer is involved. Approval should be one clear step, not a habit people skip when they are busy.

Third, ask what is checked. A model can misread a total or a date and sound completely sure. Rules that check its work and say plainly what is wrong catch those mistakes.

Fourth, ask what is recorded. When something goes wrong, you should be able to see what the system read, what it proposed, who approved it and which version of the model was in use.

Fifth, ask what happens when the model changes. It will. Build the system so you can switch provider without rebuilding it.

None of this slows AI down. It makes it something you can rely on.

Owning your software

Why the code, the accounts and the knowledge should sit with you from the first day.

Many companies learn too late that they do not own the software they paid for. The code sits with a supplier, the cloud account is in a developer's name, and the only person who understood the system has moved on.

Ownership is settled at the start: whose name is on the accounts, where the code lives, how decisions are recorded and who could take over tomorrow.

Owning your software means three things. The code and the intellectual property are assigned to your company. The repositories, cloud accounts and domains carry your company's name. And the system is documented well enough that a capable team you have never met could run it.

It also means being told plainly when something built earlier is reused, with your rights to it set out in the contract.

An owner who holds all of this can change supplier, raise capital or sell the company without asking anyone's permission.

Working examples

Three small systems on sample data. Each runs in your browser and says what is simulated.

Your team wants AI on real emails. This is what the AI should receive.

Nothing you type here leaves this browser: edit our email, or paste one of your own. Names, Emirates ID numbers, IBANs and phone numbers are replaced before anything could be sent. Each gets a numbered placeholder, so the AI can still tell people apart, and the originals stay on this device so the answer can be put back together. Switch categories on or off: what gets removed is a policy you choose.

What the AI receives

Please draft a polite reply to this email. Hi [Name 1], This is [Name 2] from the leasing office. Your deposit of AED 5,000 was received from IBAN [IBAN 1]. Before we hand over apartment 1204, please confirm your Emirates ID [Emirates ID 1]. We have your mobile as [Phone 1]. You can reach us on [Phone 2] or at [Email 1]. Regards, [Name 2]
Remove
Kept on this device7
PlaceholderOriginal
[Name 1]Omar
[Name 2]Layla Haddad
[IBAN 1]AE07 0331 2345 6789 0123 456
[Emirates ID 1]784‑1990‑1234567‑6
[Phone 1]+971 50 000 0001
[Phone 2]+971 4 000 0001
[Email 1]layla.haddad@example.ae

Runs in your browser. Nothing is sent. Names are found from context, so a name with no cue can slip through, which is why real systems pair rules with review. The people and phone numbers are fictional; the ID and IBAN pass their checksums but belong to nobody.

When the AI misreads a total, who catches it?

A sample tax invoice in Arabic and English, and what an AI would read from it. Approve it, or press “Introduce a reading error” and watch the checks stop it before it reaches the books. Every check says in plain words what is wrong.

Sample

Tax invoice

فاتورة ضريبية

Example Office Supplies LLC

TRN 100234567800003

مثال للوازم المكتبية ذ.م.م

الرقم الضريبي 100234567800003

Invoice no.رقم الفاتورة
INV-2026-0418
Dateالتاريخ
14/09/2026

Bill toفاتورة إلى

Example Logistics FZEمثال للخدمات اللوجستية م.م.ح

TRN 100398765400003

Description
الوصف
Qty
الكمية
Amount
المبلغ
Office chair
كرسي مكتب
41,800.00
Standing desk
مكتب بارتفاع قابل للتعديل
21,900.00
Delivery and assembly
التوصيل والتركيب
1500.00
Subtotalالمجموع الفرعي
4,200.00
VAT 5%ضريبة القيمة المضافة 5%
210.00
Total AEDالإجمالي بالدرهم
4,410.00
Sample for this page. Not a real invoice.نموذج لهذه الصفحة، وليست فاتورة حقيقية.

What was read

Format
  • Read
  • Checked
  • Read
  • Checked
  • Checked
  • Checked
  • Checked
  • Checked
  • Checked
  • Checked

Checks

  • The TRN has 15 digits.
  • The lines add up to the subtotal.
  • VAT is 5% of the subtotal.
  • The total is the subtotal plus VAT.
  • The date is not in the future.

Every check passes. Approve when you agree with the reading.

Record

Nothing has changed yet.

In a real system a language model reads the document. Here the reading is prepared for this page, so no AI service is called. The checks are real and run in your browser. The companies, numbers and tax registration numbers are fictional.

Change it, sign it, then try to tamper with it.

The record below is a sample, but the controls are the ones an inspector looks for. Try to break it. Change a value and the record will not save without a reason. Sign it as reviewer, then approve it under another name; edit it afterwards and the signatures stop matching. Then tamper with the trail and see exactly where the chain breaks. Entry 2 is the one an inspector would open first: it turns a reading outside the limit into one inside it.

Balance calibration record

CAL-0147

Sample data

Status
Draft
Record hash
e58e5e29

Any name will do. It goes into the trail with each change you save.

BalanceMaster data, under change control

QC-BAL-03

Test weight (g)Master data, under change control

100.0000

Acceptance limit (± g)Master data, under change control

0.0010

ResultCalculated

Within the limit

Signatures

Reviewed by

Not signed

Approved by

Not signed

Sign this record

Meaning

No password here. A real electronic signature that is not biometric needs at least two identity components, such as a user ID and a password (21 CFR 11.200).

A signature shows the printed name, the date and time, and what it means (21 CFR 11.50). It is linked to the record it signed (21 CFR 11.70; EU GMP Annex 11, section 14). Here that link is the record's hash, so any change breaks it.

Audit trail

Each entry is sealed with a SHA-256 hash of its own contents and the previous entry's hash, so the entries form one chain.

  1. Entry 1
    Time
    2026-10-26
    08:12:40 +04:00
    User
    S. Karim
    Action
    Created the record
    Hash
    a129ce19
  2. Entry 2
    Time
    2026-10-26
    08:19:05 +04:00
    User
    S. Karim
    Action
    Changed a value
    Field
    Reading
    Old value
    100.0040
    New value
    100.0004
    Reason
    Typing error, corrected from the balance printoutResult moved from outside the limit to within it
    Hash
    60268e18
Try to change the trail

Tampering goes around the application and rewrites a stored entry directly, as someone with access to the database could, and leaves the hashes as they were. Here it rewrites entry 2 so the first reading looks as if it had passed. A real system should also keep the latest hash outside the database, so that recomputing the hashes would not hide the change either.

Runs in your browser.

A sample record and a demo identity. No password is asked, nothing is stored or sent, and a reload clears it. 21 CFR Part 11 (11.10(e), 11.50, 11.70) and EU GMP Annex 11 (sections 9 and 14) expect controls like these; a validated system also needs procedures, access control and validation evidence. This shows the record-keeping part.