What does Prima Partners do?

Prima Partners is the advisory practice of Prasanna Walimbe. It operates in three modes: AI & IT strategy advisory for CXOs, corporate AI enablement and delivery, and angel investment plus mentoring for early-stage founders in fintech, supply chain and applied AI.

Who is Prasanna Walimbe?

Prasanna Walimbe is a deep-tech operator with 20+ years of experience across AI evaluation and safety, AI/ML architecture, and financial research. He is an alumnus of IIT Kharagpur and IIM Calcutta and a Certified Independent Director advising boards on cross-border AI investments, operations and compliance.

Which sectors does Prima Partners focus on?

Fintech and BFSI, supply chain, manufacturing, and applied AI across enterprise use cases. Assurance work is mapped to the EU AI Act, ISO 42001 and the NIST AI Risk Management Framework.

How do I engage Prima Partners?

Email pw@prima-partners.com to start a conversation. Engagements run globally on a remote-first basis from Dubai.

What is AI governance consulting and what frameworks do you use?

AI governance consulting maps an organisation's AI systems, decisions and controls to a recognised risk framework — typically the EU AI Act, ISO 42001 and the NIST AI Risk Management Framework — and closes the gaps. Prima Partners runs checklist-based evaluations, model-risk reviews and board-level assurance for financial services, supply chain and applied-AI use cases.

Do you serve boards as an Independent Director?

Yes. Prasanna Walimbe is a Certified Independent Director and advises boards on AI strategy, cross-border investment and operations, and compliance obligations across jurisdictions — with a specialisation in fintech, supply chain and applied AI.

Do you offer corporate AI training and enablement?

Yes. Corporate AI enablement includes executive workshops, leadership AI literacy, hands-on programmes for delivery teams, and practical playbooks — AI realism vs hype, ML deployment, GTM for AI products, and the founder's mindset inside corporates.

Do you mentor or invest in startups?

Prima Partners writes angel and pre-Series A cheques and provides hands-on mentoring. Current positions span defence technology, artisan-led premium design and manufacturing, AI-powered franchisee onboarding for D2C brands, and memory-module hardware and software, alongside fintech, supply chain and applied AI. Support covers technical architecture, GTM, commercial modelling and board readiness.

Which AI model should an enterprise standardise on?

The Stanford 2026 AI Index shows the performance gap between top models has collapsed, so standardising is now a switching-cost and security decision rather than a capability one. Reframe long-running evaluations around integration risk, data-egress controls and exit paths instead of another round of benchmarks.

Should companies slow junior hiring because of AI?

The Index's data on entry-level task exposure argues for redesigning the junior role, not deleting it. Cutting the pipeline solves an immediate headcount problem but removes the bench that becomes your senior engineers and managers by 2028.

Where do AI agents actually belong on the enterprise roadmap?

The Index's agent-benchmark results locate a narrow band of workflows where agentic systems clear the reliability bar today — mostly bounded, tool-using tasks with human review. Broader deployments across open-ended enterprise workflows are a 2027 conversation and should be scoped as such.

Can enterprises trust reported AI productivity gains?

The Index places self-reported productivity uplift directly beside realised-output figures and the two diverge sharply. Boards should measure cycle time, defect rate and unit economics on specific workflows rather than treat satisfaction survey scores as evidence of P&L impact.

Who should sign off on enterprise AI governance?

The Index's transparency scores show foundation-model vendors are disclosing less each year, so control cannot be inherited from the vendor stack. Governance sign-off belongs with an internal owner running a checklist mapped to the EU AI Act, ISO 42001 and the NIST AI Risk Management Framework.

When Everything Works and It Still Fails

2026-08-29 · Enterprise AI · 14 min read

When enterprise AI fails, it usually fails above the model — in trust, accountability and the messy business of actually running the thing.

The analysis

The thought experiment is deliberately unfair to the usual explanations. Assume the data is secure, privacy handled, the model accurate, bias tested and drift monitored, and take cost off the table. The agentic programme still fails. That is not hypothetical — it is roughly where the market is. The technical layer is the part enterprises already know how to buy: necessary, and nowhere near sufficient.

The evidence converges from five independent vantage points. Gartner expects more than 40% of agentic AI projects to be cancelled by end-2027. MIT's Project NANDA found around 95% of enterprise generative AI pilots delivered no measurable P&L impact, and was explicit that the divide is driven by approach rather than model quality or regulation. McKinsey describes the same gap as the gen AI paradox, and names the shift that makes agentic different: from systems that might say the wrong thing to systems that can do the wrong thing. Accenture found only about 13% of organisations reporting significant enterprise-level value — and that firms with responsible-AI governance across the lifecycle were roughly 2.7 times more likely to create it. Infosys puts the responsible-AI gold standard at about 2% of firms, with only around 14% having scaled agentic AI at all.

In regulated industries the gap has a shorter fuse. A wrong answer does not arrive as product friction; it arrives with a regulator attached. India's Reserve Bank made this concrete with its FREE-AI framework, placing explainability and accountability at the centre of AI in financial services. When an agent declines a loan, the customer has a right to reasons — a requirement a black box that is usually right cannot satisfy.

Trace each failure downward and stop only when the answer is no longer 'buy a better component'. What remains is a small set of structural fault lines. Accountability has been separated from the act: the agent takes a consequential action, and the human principal the law requires neither made nor could have foreseen it. Justification debt accumulates: the decision path was non-deterministic and state-dependent, so accuracy is available but a stable, contestable account of the decision is not. Tacit rules go unwritten: the agent follows the explicit instruction and violates the unwritten one nobody knew needed writing down. And tail risk governs trust — months of good behaviour end with a single egregious, often irreversible action, because autonomy raises the ceiling of harm per action and removes the human-paced friction that used to catch errors before they scaled.

The practical implication is a reordering of the work. The controls that matter are the ones that bind an answerable actor to each consequential action, produce a durable record of why an action was permitted, make the organisation's unwritten rules explicit enough to encode, and cap the blast radius of irreversible actions before scale rather than after an incident. None of that is bought from a model vendor, and none of it is a compliance overlay applied at the end. It is the operating design of the programme itself.

Full essay on Substack: When Everything Works and It Still Fails.

More on Enterprise AI

Discuss this with Prasanna: pw@prima-partners.com · book a 30-minute call.