Standard Bank, Mercy Corps and ShepHertz deploy agents with human oversight
Standard Bank names Amazon Bedrock its enterprise platform and cites roughly 20 per cent coding gains, Mercy Corps routes crisis research through five agents with analyst review, and ShepHertz offers a customer-hosted stack on India-trained models.
Artificial Intelligence··Evening
Standard Bank bets on Bedrock across four business lines
Standard Bank Group has named Amazon Bedrock its enterprise AI platform and said about a third of its technology employees now use AI-enabled coding tools. In early use cases the bank attributes productivity improvements of roughly 20 per cent to those tools; the figure is the bank's own and it has not said how it is measured or over what sample. The disclosure adds detail to interim results published the previous week and puts a number on the coding gain for the first time. The bank describes the Bedrock choice as a deliberate multi-model strategy and says it has prioritised four areas: relationship management, servicing, payments and lending.[1]
Mercy Corps chains five agents, then sends analysts in
The humanitarian NGO Mercy Corps has built VERA, an agent arrangement in which one agent works out what information is needed, four specialist agents gather the data, and a fifth writes a referenced baseline assessment. It went live at headquarters in June after six months of piloting and has since reached teams in Sudan and Colombia. VERA runs on Amazon Web Services, was written with Cloudera's AI Studios tools and employs Anthropic's Claude models. The four gathering agents draw today on conflict data from ACLED and ACAPS, displacement information from the International Organization for Migration's Displacement Tracking Matrix, humanitarian reporting through ReliefWeb and local news from field teams in Colombia and Sudan. Analysts verify the output, place it in context, flag inconsistencies and complement it with interviews with humanitarian workers. Assessments have also been produced for teams in Ukraine and the Democratic Republic of Congo. Nayid Orozco, the organisation's AI solutions and delivery manager, says analysts previously spent days or even weeks collecting information spread across websites in different languages. Over the next 12 months the aim is to deploy VERA in the other 30 or more countries where the organisation operates; Orozco says scaling will be slow because each context brings its own difficulties.[2]
ShepHertz keeps agents inside the customer's own cloud
Gurugram-based ShepHertz Technologies has launched AgentAnywhere, a platform that lets banks, insurers, hospitals, governments and outsourcing firms run AI agents inside their own infrastructure, cloud environments and encryption keys. It is built on seven model families all trained in India, of which only the coding model Taksha is generally available. According to the company, Veil masks and de-identifies personal, financial and regulated data before it reaches a model; Kavach screens requests and responses for prompt injection and jailbreaks; an Agent Universal Gateway governs model calls, tool calls and agent-to-agent messages through a single control point; and Custodian maps the platform to SOC 2, ISO 27001 and Reserve Bank of India controls. Six families are planned through 2026 — Manthan for general reasoning, Kuber for banking and insurance, Seva for service operations, Tatva for edge deployments, Astra for defence and Sanjaya for critical telemetry, with Astra and Sanjaya offered on a partner-scoped basis. Each family will come in Fast, Pro and Max tiers and run inside the customer's virtual private cloud under the same governance layer. The company says every model must pass an execution-graded testing battery before release and that two candidate models were rejected after failing those tests. Customers can also run open models such as Mistral, Qwen, GLM, Gemma, GPT-OSS and Llama within the same framework. All product and security descriptions are the company's own; no independent verification is reported.[3]
