AI Doesn't Lack Knowledge — It Lacks Trusted Context

Recently, OpenAI announced that ChatGPT for Healthcare can connect to Epic's electronic health record system, and access official medical data sources like PubMed, ClinicalTrials.gov and DailyMed through plugins.

This means AI no longer only answers general medical questions. With proper authorisation, it can combine a patient's visit records, lab results, medication changes and physicians' notes to help medical staff organise information — and point to the original records supporting each conclusion.

News link: OpenAI: ChatGPT can now connect to healthcare sources

This news reminded me that today's large models actually no longer lack general knowledge.

A model can explain what diabetes is — and it can also explain what cash flow, VAT or accounts receivable are. But it doesn't know what happened in the most recent lab test of the patient in front of it, and it doesn't know which customer a payment received last month by the business in front of it actually came from.

What truly limits AI's value is often not that it knows too little — it's that it doesn't understand the real object being handled right now.

AI doesn't lack knowledge. It lacks trusted context.

In healthcare, that context includes electronic health records, lab results, medication history and authoritative medical references.

In business finance, it includes bank transactions, invoices, receipts, contracts, customer and supplier information, historical accounting treatment, and the tax and accounting rules of the country where the business operates.

Without this context, no matter how smart the model is, it can only give generic answers.

It can tell a business what VAT is — but it cannot judge whether this specific invoice should charge VAT. It can explain how a certain type of expense is usually booked — but it doesn't know whether this payment is a business expense, a personal expense or a shareholder loan. It can describe cash-flow management methods — but it doesn't know which payments the business needs to make next week.

This is also the fundamental difference between SnapLedger and a general-purpose chatbot.

What SnapLedger needs to do is not just connect to a more powerful model, but connect the real information scattered across different places: the transactions in the bank account, the receipts photographed on a phone, the invoices received by email, the documents kept in My Vault, the ledger that has already taken shape — and the specific rules that apply to this business.

Only when this information forms trusted, continuously updated context can Snappy truly understand:

Which invoice this bank transaction probably corresponds to; Which document is still missing its payment record; Which expense needs the user to add an explanation; Which revenue might affect the VAT registration threshold; And what the user should handle first the next time they log in.

But "connecting more data" is not the goal in itself.

In healthcare, not everyone should see the full medical record. Likewise, in finance, not every agent, employee or external service provider should have access to all of a company's information.

Trusted context must satisfy several conditions at once: the data source is clear, identities are verified, permissions are controlled, and every conclusion can be traced back to original evidence.

This is why SnapLedger runs an owner check after documents are uploaded — to keep one person's receipts out of another person's books as much as possible. It is why a document, once it has entered a formal accounting relationship, cannot be deleted at will. And it is why every judgement the AI gives should be able to point back to the relevant invoice, bank transaction and applicable rule.

The capability gap between large models may shrink in the future — but business context will not become less important because of it.

Quite the opposite: the more commoditised models become, the more the truly valuable part is likely to come from the data structures, business relationships, rule systems and workflows a company has accumulated over the long term.

I've come to believe more and more that competition among AI-native SaaS is not just about who connects to the strongest model.

More importantly, it's about who can organise scattered information into trusted context — and then let AI, under the right permissions and rules, turn that context into action.

Knowledge lets AI answer questions. Trusted context lets AI truly participate in work.