Real time is a negotiation, not a requirement
I have heard “we need it in real time” in almost every data program I have led. In banking the request came from digital channels, treasury, wealth advisors and risk, and each group meant something different. The most productive question I learned to ask was simple: what decision changes if this information is an hour old? A day old? The answers were rarely what the first request implied.
The same discipline applies when Claude or any enterprise AI assistant answers questions from business data. A monthly management review can use an approved snapshot. A fraud or operational exception may depend on the last few minutes. A CFO's board materials need a fixed reporting cutoff far more than they need speed. Ask the business owner to define the acceptable age and cutoff for each workflow you plan to support.
Separate three timestamps: when the source record changed, when it was extracted and when the assistant produced its answer. They describe different events. Treating them as interchangeable is how a confident answer ends up describing yesterday.
A reconciled baseline plus visible change
At a bank I helped grow from community to national scale, online and mobile banking depended on a master data management platform that reconciled overnight. When it struggled during the day, customers could not see their own information. Rebuilding every enterprise layer was not an option. We kept the nightly reconciliation as the trusted baseline and streamed intraday changes through an event-driven design into a separate real-time store that the channels read.
That work taught me something I now apply to AI answers. Freshness is not a single property. It is an agreement between a trusted baseline, the changes layered on top of it and the honesty of the interface about which one you are looking at.
Map the sources that must agree
A question that spans several systems inherits several refresh schedules. A live CRM record and a nightly ERP snapshot can both be correct while describing different moments. When timing affects interpretation, the answer should say so. Faster access to one source does not make the combined answer current.
Document which sources are mandatory for a question and which add optional context. If a mandatory source is stale or unavailable, the workflow should stop or qualify the answer. The business owner, not the engineer, approves that distinction.
A successful connection is not proof of freshness. A service can respond in milliseconds with a week-old snapshot. Monitoring has to separate availability from age and completeness, because they fail independently. When we moved petabyte-scale data from on-premises platforms to a cloud data platform, the hardest defects were never outages. They were jobs that finished on time and quietly delivered incomplete data.
Decide what happens at the boundary
Decide in advance what the assistant does when the freshness requirement is not met: present a clearly dated historical view, ask the user to retry or explain that current evidence cannot support the decision. Users should never have to infer the limitation from a timestamp buried in a log.
Test with deliberately stale and partially refreshed data. An answer that acknowledges old data in its second paragraph and then gives an unqualified recommendation in its conclusion is still misleading. Include a source that updates mid-conversation, and decide whether follow-ups use a fixed snapshot or a fresh view. Mixing periods makes comparisons meaningless.
Set the agreement to fit the institution
Set the agreement around the decision. A monthly report may need a reconciled snapshot, while a customer-facing balance may need intraday changes. Eligibility, claims and lab results also have different timing requirements. For each workflow, record the acceptable age, the approved cutoff and what happens when the source misses it. Then test those conditions with the people who will rely on the answer.
Assign a source owner who understands business timing and an operational owner who can diagnose extraction failures. Give users one clear route to report an answer that contradicts what they know happened this morning.
Revisit the cost. A tighter refresh window adds infrastructure and support load that may not improve the decision at all. When a source misses its window, record the affected question types and temporary guidance for users, not just the failed job. A freshness agreement tells users what period an answer represents, tells operators what to watch and tells reviewers when the system should qualify or withhold a response.