Data readiness for AI: getting your information in order

Why data readiness comes first

When an AI project disappoints, the cause is often the data rather than the AI. A lead-routing automation misfires because half the leads have no service type. A proposal assistant produces generic drafts because past proposals are scattered across personal folders. A knowledge assistant gives wrong answers because three versions of the same procedure exist.

You do not need perfect data. You need the specific data a workflow depends on to be findable, consistent and current. That is a much smaller job, and it usually improves the business even if you never add AI.

It also helps to scope readiness to one workflow at a time. Trying to clean all data everywhere before starting anything is a common reason AI plans stall for a year.

1. One source of truth

Start by agreeing where the master record lives for each core thing the business tracks. For most service firms that means customers or clients, jobs or matters, and money.

Typical answers: client and matter records in Clio or MyCase for a law firm; jobs and customers in ServiceTitan, Housecall Pro or Jobber for a trades business; clients and work in Karbon or Canopy for an accounting firm; properties and tenants in AppFolio or Buildium; patients in Dentrix or Open Dental; deals and contacts in HubSpot; invoices in QuickBooks Online. The specific tool matters less than agreement. If the same client exists in four places with different details, decide which one wins and how the others are updated.

Write the decision down in a short table: record type, system of record, who can edit it, and which other systems copy from it. Share it with staff. It sounds basic, but many firms discover that different teams have different answers.

2. Consistent key fields

Identify the handful of fields your target workflow needs, then check how complete and consistent they are. Pull a sample of 50 recent records and look at:

Fix the rules before fixing the records. If free-text entry caused the inconsistency, cleaning the existing data without adding a drop-down list means you will be cleaning it again in six months.

A 50-record sample takes about an hour in a spreadsheet export and gives a far more honest picture than anyone's impression of how clean the CRM is.

• Completeness: what share of records have the field filled in at all?
• Consistency: are values chosen from a list, or typed freely so that "HVAC repair," "AC repair" and "repair-hvac" all mean the same thing?
• Duplicates: how many customers or contacts appear more than once?
• Currency: are statuses, owners and dates up to date, or did people stop updating them months ago?
• Contact details: are email addresses and phone numbers in usable formats?

3. Documents and knowledge

Many AI uses in service businesses work from documents: past proposals, procedures, templates, contracts and reports. Here readiness means people can find the right version.

Check whether there is a shared, organized place for each type of document, or whether key files live in personal OneDrive or Google Drive folders and email attachments. Look for duplicate and outdated versions of procedures and templates. Decide on a naming pattern and an owner for each important folder. For an internal knowledge assistant, start with a small, curated set of current documents rather than connecting everything.

Past outputs are valuable here. Twenty good, approved proposals or reports in one folder give an AI drafting tool far better examples to follow than a template alone. Mark which ones are good examples, since old files often include mistakes or outdated pricing.

4. Access and permissions

AI assistants connected to your files generally show each user what that user already has permission to see. Microsoft, for example, states that its Copilot only surfaces organizational data the individual user has at least view permission for. That is reassuring, but it also means loose permissions become visible very quickly. If everyone can open the HR folder today, an assistant may summarize it for anyone who asks.

Before connecting AI to shared drives, review who can access sensitive areas such as HR, payroll, partner finances, health information and privileged client files. Remove broad "everyone" sharing where it is not needed, and check external sharing links.

For regulated data, such as health information, taxpayer records or privileged material, check your obligations before connecting any AI tool, and keep those folders out of scope until you have confirmed the tool's terms and agreements are in place.

Fixes that pay back on their own

Most data readiness work is unglamorous and inexpensive. In rough order of value:

None of these needs AI to justify it. Cleaner records make reports accurate, reduce embarrassing errors such as duplicate mailings, and make every future automation simpler and cheaper to build.

Pick the two or three that affect your target workflow and schedule them, rather than trying to do all seven at once.

• Make a few key fields required and use drop-down lists instead of free text.
• Merge duplicate customers and contacts, and set rules to prevent new ones.
• Agree and document which system is the source of truth for each record type.
• Move shared documents out of personal folders into organized team locations.
• Archive outdated templates and procedures so only the current version is easy to find.
• Tighten permissions on sensitive folders and remove stale external links.
• Assign an owner for data quality in each core system, with a monthly check.

How to know you are ready enough

For a specific workflow, you are ready enough when a person new to the business could find the inputs, trust the key fields and locate the current templates without asking around. Measure it: field completeness for the workflow's key fields, duplicate rate, and how long it takes to assemble the inputs for one example.

Data readiness is part of every InTheory engagement's first 30 days, because it often decides which workflow should go first. Sometimes the cleanup is the project, and the AI can wait.

If you only have time for one thing this quarter, pick the workflow you most want to improve and fix its three or four key fields. That focused effort tends to deliver visible results quickly, which makes the next round of cleanup easier to justify.