Insights
Shipping AI with the same bar as production software
Tax questions invite conversational AI: dense rules, recurring patterns, and users who want a clear next step. In our tax-assistance work, the demo always looked ready. Production asked harder questions: who owns the answer when rules change, when content is wrong, or when the model is confidently mistaken?
That gap is familiar across enterprise AI. A capable model in a controlled setting is not yet a feature teams can operate. Live behavior needs named ownership, measured quality before scale, and a designed way to pause when outputs are wrong in a way that matters.
The challenge
When nobody is named to define “good,” handle bad output, and decide when to stop, the assistant becomes an unowned step inside someone else’s workflow. A chat window beside the work is also different from a governed step inside the work.
Without evaluation gates—and without a designed fallback—teams discover risk only after users already depend on the feature. Security and compliance constraints arrive late, as stickers on an architecture that was never shaped for them.
What we built
We apply the same production discipline used for other customer-facing services: named owners per use case, evaluation before scale, and reverse paths designed in from the start. On tax assistance, that meant pairing the conversational agent with management services for tax content, instructions, documents, and returns—so answers sit on knowledge teams can maintain.
Architecture choices follow governance early: what can be updated without a redeploy, what must be reviewed, how quality is observed, and how the assistant is paused or corrected without abandoning users. The mobile and admin surfaces are edges of that operable backbone—not a free-floating chatbot.
The outcome
Product and risk owners get an AI feature they can operate. End users get guidance that sits on something maintainable. Teams retain a way to pause, fall back, and explain what happened when quality slips.
Related
Related insights
Insights
Modernizing the platform without pausing the business
A cross-border logistics rebuild showed that continuity is the real delivery risk. How we sequence platform change so obligations stay correct through cutover.
Insights
Moving point-of-care results into the hospital system of record
Point-of-care analyzers only help when results land in HIS/LIS with failures operators can see. How we built a recoverable bridge—HL7, CDS export, and operable sync.
Insights
Proxy buying, consignment, and warehouse truth on explicit models
Cross-border logistics mixes buy-for-me, consignment, and package movement that cannot share one status field. How we rebuilt order kinds, lots, and settlement for China–Vietnam operations.
Insights
Receivables operations that match how money is worked
Installment and credit portfolios live on the store floor—disbursement, collections, and bad debt. How we built receivables software around business dates, auditable status, and exception-heavy queues.
Insights
Field product digitization that becomes a system of record
Shelf reality has to become operable product data. How we built mobile capture—photos, measurements, SKUs, and sync—so field digitization writes back instead of dying as a campaign gallery.
Insights
Grounding tax assistance in managed knowledge
A tax assistant needs more than a model. How we paired conversational help with managed content, returns workflows, and production controls teams can supervise.