Risk and Cost Governance for AI Agents in Regulated Institutions  

This interview evaluation is sponsored by Zafin and was written, edited, and printed in alignment with our Emerj sponsored content guidelines. Learn extra about our thought management and content material creation companies on our Emerj Media Services page.

Regulated establishments are deploying AI brokers into actual workflows. This requires the governance, management, auditability, and price self-discipline wanted to allow brokers to work together with delicate programs, knowledge, and choices with out introducing new operational, regulatory, or monetary danger.

Deployment is already outrunning oversight. A Cloud Security Alliance survey of 228 IT and safety professionals found that 85% of organizations now run AI brokers in manufacturing environments, but 68% of those self same organizations can not clearly distinguish agent exercise from human exercise inside their very own programs.

Regulators are more and more responding to the governance challenges created by AI adoption in monetary companies.The Financial Stability Board’s June 2026 session proposed organization-wide AI governance and AI-specific danger administration practices, emphasizing board and senior administration accountability for AI deployment choices.

The U.S. Government Accountability Office has warned that AI use in monetary companies poses dangers, together with biased lending outcomes, privateness considerations, cybersecurity threats, and rising dependence on third-party expertise suppliers, necessitating stronger mannequin danger and oversight practices.

The Treasury Department has since moved to shut that hole: in February 2026, it released a Financial Services AI Risk Management Framework alongside a shared AI Lexicon, aiming to provide establishments a typical, risk-based customary for governing AI deployment.

Layered on high, a Cloud Security Alliance analysis word on 235 large-enterprise safety leaders found that 92% lack full visibility into their AI identities, and that 71% report AI programs have already got entry to core enterprise platforms — ERP, CRM, and monetary programs — whereas solely 16% govern that entry successfully.

Emerj’s Yolandi de Weerdt not too long ago hosted a dialog with Shahir Daya, Chief Product & Technology Officer at Zafin, to look at how agentic AI is reshaping price buildings, workflow execution, and operational self-discipline throughout monetary establishments.

This article distills actionable insights for leaders accountable for scaling AI brokers in regulated environments:

• Workflow‑degree governance for scalable agent oversight: Define controls on the workflow layer so each agent motion carries constructed‑in authorization, human judgment, and proof that may be surfaced immediately fairly than reconstructed weeks later.
• Control‑aircraft infrastructure for actual‑time agent orchestration: Insert a centralized working layer between human intent and agent execution so autonomous work strikes by delicate programs with deterministic transitions and full auditability.
• Variable‑compute price self-discipline for sustainable agent deployment: Govern mannequin alternative and token spend on the process degree in order that agent workloads keep inside predictable financial bounds fairly than ballooning into uncontrolled compute liabilities.

Listen to the total episode beneath:

Episode:  Risk and Cost Governance for AI Agents in Regulated Institutions – with Shahir Daya of Zafin

Guest: Shahir Daya, Chief Product & Technology Officer at Zafin.

Expertise: Enterprise Technology Strategy, Cloud & Hybrid Cloud, Financial Services Technology, Product & Technology Leadership

Brief Recognition: Shahir Daya is Chief Product & Technology Officer at Zafin, following a 27-year profession at IBM, the place he most not too long ago served as IBM Distinguished Engineer and Chief Technology Officer for IBM Consulting in Canada. At IBM, he held senior structure and expertise management roles throughout cloud, enterprise transformation, and monetary companies. Daya has co-authored three IBM Redbooks on microservices and hybrid cloud integration and is an inventor with a number of issued U.S. patents. He additionally mentors by the University of Toronto Engineering Alumni Mentorship Program and WISE. He holds a B.A.Sc. in Computer Engineering from the University of Toronto.

Workflow‑Level Governance For Scalable Agent Oversight

Shahir Daya identifies a governance hole that arises when brokers are inserted into workflows designed for people. Pricing modifications, supply outcomes, disclosure updates, fraud opinions, and lending workflows all carry regulatory questions that assume a human made the choice.

Once an agent participates, establishments should be capable to present who licensed the motion, what knowledge the agent used, the place human judgment occurred, and what proof stays. Daya notes that almost all banks can not reply these questions as a result of their working fashions have been by no means constructed to account for agent‑pushed choices.

Daya notes that AI didn’t create this governance drawback, however is accelerating it. As brokers unfold throughout enterprise items, establishments face rising strain to elucidate choices, reveal accountability, management prices, and fulfill regulators, all utilizing working fashions not designed for agent-driven work.

He distinguishes how establishments govern brokers and how regulators consider danger. Regulators give attention to the workflow:

• which workflows are in scope
• what the chance floor seems like
• what controls apply at every step
• the place human judgment matches
• what proof is retained.

Shahir emphasizes that these questions apply throughout monetary companies and that present working fashions can not produce speedy solutions when an agent is concerned.

He additional explains why agent‑degree controls fail. One agent may match throughout a number of workflows, whereas a single workflow might contain a number of brokers. Because regulators consider determination processes fairly than particular person applied sciences, workflow‑degree controls present a extra scalable basis for accountability and explainability.

The sensible implications observe instantly from Daya’s statements:

• Define the workflow because the unit of governance earlier than deploying brokers.
• Map the factors the place human judgment should stay seen and auditable.
• Specify the controls that apply at every workflow step fairly than on the agent degree.
• Separate workflow ideas from agent ideas to maintain controls steady as brokers multiply.
• Establish a constant workflow taxonomy early so governance controls stay steady as agent adoption expands.

Daya describes what regulators need to perceive:

“Regulators don’t need to speak about your brokers — they need to discuss in regards to the workflow. They need to know what the chance universe is, which workflows are in scope, which controls apply at every step, and the place human judgment really matches. If you construct your controls round particular person brokers, they won’t scale as a result of one agent can work throughout many workflows and many brokers can run one workflow.”

– Shahir Daya, Chief Product and Technology Officer, Zafin

Control‑Plane Infrastructure For Real‑Time Agent Orchestration

Daya describes the management aircraft because the working infrastructure that sits between human intent and agent execution. It coordinates agent exercise by deterministic workflow transitions whereas sustaining visibility into authority, software utilization, and execution historical past. Beyond orchestration, it additionally supplies the guardrails and proof wanted to manipulate agent exercise at scale.

To make clear how orchestration works, Shahir outlines the mechanics of the management aircraft:

“The management aircraft just isn’t a committee that meets on Wednesdays — it’s working infrastructure. It sits between human intent and agent execution, coordinating each motion by deterministic transitions and full visibility. Just like a management tower is aware of each plane’s id, clearance, route, and precedence, the management aircraft is aware of each agent, what it’s doing, what instruments it’s calling, and beneath whose authority it’s appearing.”

– Shahir Daya, Chief Product & Technology Officer at Zafin

Daya additionally argues that brokers ought to be held to the identical governance requirements as workers. Rather than granting broad permissions, establishments ought to apply zero-trust and least-privilege ideas so brokers obtain solely the entry required for a selected process. This limits pointless publicity to delicate programs and helps danger groups confirm that agent exercise stays inside permitted boundaries.

Daya additionally emphasizes the significance of “proof of labor.” By recording prompts, actions, and execution historical past, establishments can evaluation how an agent accomplished a process fairly than reconstructing choices after the actual fact.

The operational necessities observe instantly from Daya’s description:

• Position the management aircraft between human intent and agent execution so autonomous work follows ruled workflows.
• Use deterministic workflow transitions to keep up constant and auditable execution paths.
• Apply zero-trust and least-privilege controls in order that brokers entry solely the programs required for a selected process.
• Maintain visibility into brokers’ identities, authorities, software utilization, and workflow standing.
• Capture proof of agent actions, choices, and execution historical past as proof of labor.
• Provide a typical working layer that coordinates brokers throughout fashions, distributors, and environments.
• Treat orchestration, governance, and proof assortment as steady operational features fairly than periodic opinions.

Shahir’s broader level is that scaling brokers requires governance infrastructure, not simply automation. The management aircraft supplies the coordination, visibility, and proof wanted to function agent-driven workflows in regulated environments.

​

Variable‑Compute Cost Discipline For Sustainable Agent Deployment

Agentic AI modifications the economics of labor by shifting AI spending from a predictable software program expense to a variable compute price. Daya argues that many establishments proceed to manipulate AI as a set finances merchandise despite the fact that consumption grows with workflow quantity, mannequin utilization, and agent exercise. As deployment expands, prices change into a property of operational execution fairly than software program procurement.

To illustrate the shift, Shahir factors to latest trade examples:

“Agentic AI modifications the economics of labor in a means most establishments haven’t internalized. Uber burned by its whole 2026 AI coding finances by April, and Safe Software went from $20,000 a month to $100,000 a month in six months — and these are disciplined corporations. As subsidies disappear, enterprise AI payments will rise one other 30 to 50 %, which suggests AI turns into a variable compute price that have to be actively ruled.”

– Shahir Daya, Chief Product & Technology Officer at Zafin

For Daya, the problem just isn’t merely rising prices however restricted visibility into the place spending originates, which fashions generate it, and whether or not the ensuing outcomes justify the underlying compute consumption.

Shahir’s examples reveal a special type of self-discipline establishments should undertake as agent workloads broaden:

• Recognize AI as variable compute, not a set software program line merchandise.
• Set process‑degree spending limits so utilization can not quietly escalate.
• Control mannequin choice to keep away from pointless excessive‑price inference.
• Track token consumption to detect rising price spikes early.
• Build actual‑time price visibility into orchestration programs so spending displays precise workflow conduct.
• Build price fashions that account for altering pricing buildings as AI suppliers transfer towards consumption-based economics.
• Anchor agent deployment to sturdy price boundaries, not exploratory budgets.

Daya’s broader level is that price governance have to be built-in into agent governance. As agent exercise scales, establishments want the identical degree of visibility into compute consumption that they already anticipate for operational danger, compliance, and workflow efficiency.