Practical notes from the engineers who build agents, copilots and AI products.
AI development in 2026 ranges from about $3,000 for a discovery sprint to six figures for platforms. The cost drivers are data readiness, the number of integrations, the level of autonomy and evaluation depth, not the model. Here are real ranges in USD and INR.
It depends on the problem and the signal; assess in a discovery sprint and use pretrained or LLM approaches where data is thin. A strong signal with clean labels needs hundreds of examples, a weak one needs many thousands, and the honest way to find out is a learning curve on your own data rather than a rule of thumb.
HIPAA-aligned AI keeps PHI in controlled environments, signs BAAs, logs access and never sends patient data to unvetted models. The rules were written for records systems and apply unchanged to a model that reads a chart, transcribes a consultation or answers a patient's call; the work is vetting vendors.
GDPR-aligned AI needs a lawful basis, DPIAs for high-risk use, processor terms, transfer safeguards and data subject rights support. Most of that is architecture: where data flows, which vendors see it, how it is deleted from every store, and whether a person can be told why the system decided what it did.
A platform exposes what your product does through APIs and webhooks so partners build on it; design for it early, build it when demand exists. The sequence that works is an internal API first, webhooks for the events partners ask about, then a public launch with keys, quotas, docs and a partner programme.
Migrate when performance, integrations or ownership block growth; keep the validated product, rebuild the foundation properly. The no-code app has proved demand and taught you the workflow; the migration moves that knowledge into code you own, with a real data model, API and tenancy, in waves that keep customers live.
Migrate by clustering historical tickets, connecting account data, launching in shadow mode and expanding autonomy per intent. Keep the old bot running until the agent beats it on each intent, retire flows one at a time, and measure resolution rather than deflection so the upgrade proves itself before it takes over.
Fraud models score transactions in real time, validated on hold-out periods, with precision tuned to the cost of a false positive. The hard parts are labels that arrive weeks late, fraud patterns that shift as soon as you block them, and a review team that has to trust the score enough to act on it.
Fixed price for scoped programs, time and materials for open-ended research; the mistake is fixed price without a scope lock. AI adds a twist: accuracy cannot be promised before it is measured, so the contract should fix the scope, the date and the evaluation method, and let the discovery or POC sprint set the target.