Clean data before agents.

When an AI agent gives a customer the wrong answer, the instinct is to blame the model. In our experience it is usually the data. The agent found two records for the same customer and picked the older one. It read a product price that finance changed last quarter in another system. It addressed a regional director by a title they left two years ago.

Agents do not fix bad data. They act on it, quickly, and in front of customers. That is why clean data is the first step of any agent programme, not a later improvement.

Why does data quality matter more with agents?

A person reading a messy record applies judgement. They notice the duplicate, ignore the stale field, call a colleague. An agent has no colleague to call. It treats what it reads as true. HubSpot makes the same point about Data Hub: it exists so that AI tools have “complete, accurate, and up-to-date information” (HubSpot).

Scale makes it sharper. A rep with a bad record sends one wrong email. A Prospecting Agent working from the same data can send many. A Customer Agent reading an outdated policy gives the same wrong answer every time it is asked.

What is Data Hub?

Data Hub is HubSpot's data management product. On its product page, HubSpot lists Data Studio to “blend first and third-party data”, two-way sync with cloud data storage platforms, “100+ integrations that sync in real-time”, data quality monitoring, and automation to “fix formatting issues and create custom rules to keep your data clean and consistent”. Professional adds AI-created datasets; Enterprise adds advanced datasets and data warehouse connections (HubSpot).

HubSpot's knowledge base describes the data quality tools in more detail: an overview with recommended actions and property insights, duplicate management for contacts and companies, formatting fixes, enrichment coverage, and a weekly data quality digest. Some advanced features, such as anomaly alerts, need Data Hub Professional or Enterprise (HubSpot Knowledge Base).

What should you fix first?

Not everything. Start from the first agent's job and clean the data it will read and write. We work through five steps.

1. Profile. Measure the current state of the objects the agent will use: how many records lack an owner, an industry, a market or a lifecycle stage; which properties nobody fills in; which records have not been touched in a year. Property insights show where each property is used.
2. Merge. Resolve duplicate contacts and companies. In Southeast Asia, duplicates often come from local-language and English versions of the same company name, group companies entered as separate accounts, and contacts with several phone numbers across WhatsApp and email. Decide merge rules before you merge.
3. Standardise. Fix formats for names, phone numbers, countries and dates. Replace free-text fields with picklists where the agent needs to filter or decide. Write the rules once and automate them.
4. Sync. Connect the systems that hold the truth for prices, stock, contracts and invoices, so the agent reads current data rather than a copy. Decide which system wins when they disagree.
5. Govern. Give every important object and property an owner, set validation rules, and review the weekly data quality digest in a standing meeting. Clean data decays without owners.

How clean is clean enough?

Clean enough is when your best employee would trust the record without checking another system. Test it. Take a sample of records the agent will rely on and ask the people who know those customers to grade them. Where they would hesitate, the agent will be wrong.

Set a small number of thresholds for go-live, such as owner coverage, duplicate levels and the age of key fields, using your own baseline. Then keep measuring after launch. Handoffs and poor answers from the agent are an early warning that data has slipped.

Resist the urge to clean the whole database before the first agent. That project never ends. Clean what the first use case needs, prove it, and extend the same routines to the next.

How does this connect to context?

Data is half of what an agent needs. The other half is context: brand voice, policies, product knowledge and rules. We cover that half in context before agents. Both need the same thing: one source of truth, an owner and a review rhythm.

Is there a compliance angle?

Yes. Under Malaysia's amended data protection law, data portability, breach notification and processor obligations are all easier to meet when each customer has one record and you know which systems hold their data. Merging duplicates and mapping integrations is good data protection practice as well as good AI practice. We set out the detail in PDPA amendments for CRM teams.

Where does the Data Agent fit?

HubSpot's Data Agent turns CRM data, calls and documents into insight, and HubSpot reports it makes teams 10x faster to research a prospect or prepare for a call (HubSpot). It is a good example of the principle. The faster an agent works, the more the quality of its inputs decides the quality of its output.

What should leadership ask?

Before approving any agent, ask three questions. Which records and properties will it read? Who owns them? When were they last checked? If nobody can answer, the agent is not ready, and neither is the data.

Put those three questions to the owner of the first agent you plan to launch.