Is Your Data Ready for AI? Fix the Records Before the Model
Your data is ready for AI when the records the AI will read are accurate, complete, consistent, current and properly controlled. Most businesses find they are not, and that is normal. AI tools answer from your data, draft from it and act on it. If customer records are duplicated, product names are spelt five ways or policies exist in three versions, AI repeats those errors faster and in more convincing language.
The fix is not a company-wide data project. It is cleaning the specific records your first AI use case will touch, at the source, and stopping new mess from entering.
The short answer
- Start from the use case: what will the AI read, write or decide?
- Check five things in those records: accuracy, completeness, consistency, currency and access.
- Measure with a sample of real records before you spend anything on AI.
- Fix errors where the data is created, not in a copy made for the AI.
- Arabic names, spellings and digits need their own rules.
- Decisions made by automated processing carry legal duties in Oman, including human review on request.
“AI doesn’t fix messy data. It repeats its mistakes at speed, in fluent sentences. Cleaning the records is the least glamorous part of any AI project, and the part that decides whether it works.” — Abdulqader Wahiba, founder and CEO of ZAWAT
What “good data” means
The international standard for data quality, ISO/IEC 25012, defines 15 characteristics, viewed from two angles: qualities of the data itself, and qualities that depend on the system holding it. For a business preparing for AI, five of them do most of the work:
| Characteristic | What it means | Typical symptom | Typical fix |
|---|---|---|---|
| Accuracy | Records match reality | Wrong prices, old addresses | Correct at source; assign an owner |
| Completeness | Required fields are filled | Products without categories, customers without phone numbers | Make key fields mandatory at entry |
| Consistency | The same thing is recorded the same way | One customer under four spellings | Deduplicate; standardise formats |
| Currentness | Data is up to date for its use | Superseded policies still in the shared drive | Archive old versions; set review dates |
| Access and confidentiality | The right people see the right records | Payroll files readable by everyone | Permissions before any AI reads them |
Why AI raises the stakes
A person who sees a strange record usually notices. An AI assistant drafting a reply from the customer record does not hesitate: it uses the wrong address, the old price or the duplicate account, and writes it up fluently. An AI that answers staff questions from company documents will quote the outdated policy as readily as the current one.
The more you automate, the less often a human looks at each record before it is used. That is the point of automation, and it is why data quality has to be fixed before, not after. An AI assistant for your own staff shows how this plays out with documents.
A practical routine, in eight steps
1. Start from the use case
Write down exactly what the AI will do: answer staff questions from policies, draft customer replies from the CRM, extract data from supplier invoices, or summarise sales. That defines which records matter. Where to start with AI in a small business helps choose the first use case.
2. List the sources and owners
For each kind of record the use case touches, note where it lives and who is responsible for it. If two systems both claim to hold the customer list, decide which one is the source of truth.
3. Measure with a sample
Abdulqader Wahiba suggests a simple first measure: pull 100 records at random and count how many you would trust in front of a customer. Then note what is wrong with the rest. It takes an afternoon, and it replaces opinions with a number.
4. Fix at the source
Correct errors in the system where the data is created, not in an export prepared for the AI. Otherwise the next export brings the old errors back.
5. Standardise the formats
Phone numbers in one format with the country code, dates in one format, one unit for each product, one way of writing each branch. Small rules, applied everywhere, remove a large share of inconsistency.
6. Write down the definitions
What counts as an “active customer”? Is a “sale” recorded at order, at delivery or at payment? Two departments answering differently will give an AI two different truths. A short data dictionary settles it.
7. Stop new mess at entry
Required fields, drop-down lists instead of free text, duplicate checks when a new customer is created, and validation on formats. Cleaning without prevention is a job you will repeat every year.
8. Control access before the AI reads anything
An AI tool can see whatever its connected account can see. Fix permissions first, so that confidential HR or finance files are not quietly made searchable to everyone. A policy for staff using AI tools covers the people side.
Arabic data needs its own rules
Businesses in the Gulf face data problems that English-only guides ignore:
- Name spellings and transliteration. One person can appear in Arabic script and in several Latin spellings. Decide which fields hold which script, and link the variants to one record.
- Letter variants. Arabic text often varies in small ways, such as ة and ه at the end of a word or different forms of alif with hamza. Search and duplicate detection must treat these variants sensibly.
- Digits. Arabic-Indic and Western digits can both appear in the same column. Normalise them before any calculation or matching.
- Mixed-language fields. Product names and addresses often mix Arabic and English. Keep separate fields for each language where you can.
Building bilingual Arabic-English software covers these choices at the system level.
When AI makes decisions about people
If an AI system makes or shapes decisions about individuals, such as screening applicants, setting credit terms or flagging customers, data quality becomes a legal matter too. Under Oman’s Personal Data Protection Law, as amended by Royal Decree 68/2026, the controller using automated processing must take appropriate measures to protect privacy and avoid harm. The person concerned has the right to object to decisions resulting from automated processing, and the controller must introduce a human element when reviewing them. Wrong data in, wrong decision out, and now a reviewable one. Oman’s PDPL amendments sets out the wider changes.
Frequently asked questions
Do we need a data warehouse before using AI? Not for a first use case. You need the specific records that use case touches to be clean, current and properly permissioned. A warehouse may come later, when several AI and reporting uses share the same data.
How clean is clean enough? Clean enough that the error rate in your sample is acceptable for the use. A draft reply that a person reviews tolerates more than an automated decision about a customer. Set the bar by the consequence of an error.
Can AI help clean the data? Yes. AI tools are useful for suggesting likely duplicates, standardising formats and flagging gaps. A person should approve the changes, and the results should be checked against a sample before they are applied everywhere.
Who should own data quality? The department that creates each kind of record, not IT alone. Sales owns customer records, purchasing owns supplier records, and finance owns the chart of accounts. IT provides the tools and controls.
How long does it take to get data ready? For a single, well-chosen use case, often weeks rather than months. The sample measure in step 3 tells you early whether the job is small or large.
Sources: arc42 Quality Model — ISO/IEC 25012 data quality model (summary) · ISO/IEC 25012:2008, Software engineering — Software product Quality Requirements and Evaluation (SQuaRE) — Data quality model · Personal Data Protection Law, Royal Decree 6/2022, as amended by Royal Decree 68/2026, Article 14
This article states the position under the laws in force on 9 November 2026, the date of publication, and is not legal advice.
Planning an AI project and unsure whether your data will hold up? ZAWAT builds AI and automation on top of clean, well-controlled business data, and will tell you if the data work has to come first. Book a call.