Agentic AI in Regulated Firms: Controlled Delegation, Sourcing, and Cost Discipline
AI agents raise the governance bar. Treat them as higher-risk by default.
Everything your AI does today is passive. It drafts. It summarises. It suggests. What comes next takes actions. That is a different governance problem entirely.
Agentic AI is not a future consideration for most regulated firms. It is already arriving in the platforms and vendor roadmaps that many Jersey FS organisations are actively evaluating. The question is not whether to engage with it. The question is whether your governance model has caught up with what it actually requires.
Gartner projects the AI services market will reach $1.1 trillion by 2029, with the strongest growth in AI embedded into existing software and service relationships — not in standalone AI platforms (Gartner, 2026d). That trajectory means the governance and sourcing decisions your firm makes about agents in the next twelve months will shape your operating model for the rest of the decade.
Why Agents Change the Management Problem
The management challenge with passive AI capabilities is fundamentally about output quality and evidence. Does the AI produce reliable output? Does a qualified person review it? Is that review documented? Those are tractable questions.
Agents introduce a different set of questions. Who authorised the action the agent took? What access did it have? If it made a mistake, what happened downstream before the mistake was detected? And — critically — how do you stop it? These questions shift the governance problem from quality assurance to operational accountability. An agent that takes actions without clear authority, documented controls, and a functioning kill switch is not an efficiency tool. It is an uncontrolled delegation of authority.
The Jersey AI Playbook's principle of human-AI partnership has specific operational meaning here: for any decision with regulatory, client, or financial consequence, human oversight must be more than aspirational (Government of Jersey et al., 2026).
The Eight Questions Every Agent Deployment Must Answer
Agents should be treated as Tier 3 by default until demonstrated evidence supports a reclassification. The table below presents the eight pre-deployment questions as a board sign-off instrument.
Table 8 — Agent Deployment Checklist: Eight Questions Before Go-Live
| Pre-deployment question | What this controls | Status |
|---|---|---|
| What is the agent's explicit scope and what are its prohibited actions? | Prevents unauthorised access or actions beyond intended function | |
| What is its identity and access model? Is least privilege applied? | Constrains blast radius if agent behaves unexpectedly | |
| What data boundary rules apply — what can the agent read and write? | Prevents data exposure across repositories or systems | |
| Where are the human checkpoints at decision points with material consequence? | Ensures human accountability is retained for regulated decisions | |
| What does the monitoring and evidence architecture look like? | Enables detection of anomalous behaviour before consequences accumulate | |
| What is the kill switch, and can it be demonstrated? | Guarantees ability to pause and revert without uncontrolled downstream effects | |
| What are the cost caps and unit economics? | Prevents consumption-based cost escalation without value evidence | |
| What is the pre-deployment test plan and the regression framework? | Confirms the agent behaves as designed before live deployment |
The Sourcing Question Nobody Is Asking Carefully Enough
The governance conversation about agents is inseparable from a sourcing question most firms are not yet engaging with at the right level. Gartner's CIO research argues that sourcing discipline should be driven by differentiation and risk, not vendor convenience, and recommends sprint-based co-design rather than static RFPs — allowing the organisation to test business fit, delivery capability, and governance maturity before committing to contract terms (Gartner, 2026d). The build/buy/extend framework below provides the decision structure.
Table 7 — Build / Buy / Extend: The Sourcing Decision Matrix for Regulated FS
| Sourcing option | When to use | Governance implication | Jersey FS risk profile |
|---|---|---|---|
| Consume (buy off-the-shelf) | Capability is commodity. Low differentiation. Low risk. Widely available. | Vendor assurance and contractual data protections must be verified — not assumed. | Acceptable for low-risk, non-client-facing workflows. Verify data residency against JFSC and JOIC requirements. |
| Extend (vendor platform + your guardrails) | Domain specificity matters. Vendor provides scale; your firm provides constraints, context, and oversight. | You own the guardrails you build, even if you do not own the underlying model. | Suitable for most regulated FS use cases. Most onerous governance work sits in your layer, not the vendor's. |
| Build (in-house or bespoke) | Mission-critical workflow. Strategically differentiating. Or sensitive data handling that cannot be assured through vendor terms. | Full lifecycle management, ModelOps, and evidence architecture is your responsibility entirely. | Reserve for workflows where data sovereignty, JFSC compliance, or IP sensitivity cannot be met by vendor contractual terms. |
Sourcing choices for agents have consequences beyond the immediate capability: lock-in risk, ongoing cost profile, assurance complexity, and data governance obligations all change based on where the capability sits. These are board-level considerations, not procurement ones.
Contracting for AI: The Problem Most Firms Are Not Yet Addressing
Traditional contracting models are becoming misaligned with AI economics in a way that directly affects regulated firms renegotiating service relationships. Gartner's CIO research highlights a specific commercial problem: time-and-materials pricing allows vendors to capture AI-enabled efficiency gains themselves, while the regulated firm continues paying as if the work were still done at human pace and cost (Gartner, 2026d). For Jersey FS firms reviewing service contracts, this is a commercial discipline question, not a procurement detail.
Gartner recommends that AI contracts should include outcome-based or hybrid pricing models rather than pure effort-based arrangements; explicit data-use clauses covering training data, retention, and secondary use; portability and exit rights to prevent strategic lock-in; consumption cost controls for token-based or usage-based pricing; and performance warranties with explicit quality thresholds (Gartner, 2026d). In a JFSC-regulated environment, data sovereignty assurances and audit logging obligations should also be confirmed in contractual terms, not left to vendor policies.
TRISM Governance for High-Risk AI Use Cases
Gartner recommends what it terms TRISM-style governance for agentic and high-risk AI deployments: cross-functional oversight from IT, legal, risk, finance, and business owners, with explicit rules for data use, audit logging, human review, portability, and exit rights (Gartner, 2026d). For Jersey FS firms, this cross-functional structure maps directly onto the first and second lines of defence — business owners, compliance, and risk — with a defined board escalation path for any agent deployment that affects client-facing or regulated workflows.
The Cost Architecture of Agent AI: What the Numbers Actually Look Like
Abstract FinOps principles land differently when they are attached to real numbers. The table below maps five common agent archetypes to their cost and indicative value profile, based on Microsoft Copilot Studio's published credit pricing (Microsoft, 2025). These figures are illustrative for a regulated FS firm of 20-30 users — the methodology is transparent and the cost assumptions are entirely based on public licensing data.
Table 10 — Agent Cost Architecture: Five Archetypes for a Regulated FS Firm (Based on Microsoft Copilot Studio Pricing)
| Agent type | Interaction type | Credits per interaction (Microsoft, 2025) | Indicative annual operating cost (20-30 user firm) | Indicative annual value |
|---|---|---|---|---|
| FAQ / policy lookup | Classic answer | 1 | ~£200 | £3,000–£4,200 (time saved on lookups and training queries) |
| Client data support | Generative + agent actions | ~7 | ~£750 | £25,000–£37,500 (RM data lookup time: 30 mins → <2 mins) |
| Compliance monitoring | Graph-grounded + generative | ~12 | ~£650 | £14,500–£20,800 (automated compliance monitoring hours) |
| Advanced reasoning | Deep reasoning + generative | ~10 | ~£2,200 | £20,000–£30,000 (IC prep reduction; consultant fee avoidance) |
| Process automation | Autonomous triggers + actions | ~30 | ~£650 | ~£25,000 (daily manual operational checks eliminated) |
| Total — all five agents | — | — | ~£4,450/year | £88,000–£118,000 (20–26× ROI on agent operating cost) |
The overall picture is compelling: total annual agent operating costs for all five archetypes combined — approximately £4,450 per year — represent less than the cost of one professional advisor day in many regulated firms. The question is not whether agents are affordable. The question is whether your governance architecture makes their deployment responsible.
Cost Governance: The Design Decisions That Determine Whether Agents Stay Affordable
The table above masks a critical risk. Within a single agent platform, the cost differential between the cheapest and most expensive operation type is not marginal — it is significant. Microsoft's public pricing shows that deep reasoning capabilities can cost up to 50 times more per interaction than a standard generative answer (Microsoft, 2025). An advanced reasoning agent designed without appropriate constraints can consume the same credit budget as all four other agent types combined.
Two design decisions above all others determine whether agent operating costs remain within the business case: first, using agent flows (13 credits per 100 actions) rather than discrete actions (500 credits per 100 actions) — a 97% cost reduction on high-frequency operational tasks; second, limiting deep reasoning capabilities to explicitly approved use cases with daily usage caps, rather than making them available as a general feature. Neither of these is a technical configuration decision. Both are governance decisions that belong in the agent deployment checklist before go-live.
The monitoring architecture should match the cost architecture. Department-level credit allocations with monthly consumption reporting against budget, escalation triggers for variances above 20%, and quarterly efficiency reviews at AI governance level are the minimum cadence for any agent deployment with autonomous or deep reasoning capabilities. The firms that discover runaway consumption at the end of a quarter have not been governing their agents — they have been running them.
What Controlled Delegation Actually Looks Like
Deploying an agent is an act of controlled delegation. The governance obligations that apply to human delegation — clear authority, appropriate access, documented oversight, accountability to a named owner — apply equally to agentic delegation. The firms that govern agents well are not the ones that avoid agents. They are the ones that treat the deployment decision with the same deliberateness they would apply to hiring a new team member and giving them access to client data and internal systems.
Closing question: If your firm deployed an agent tomorrow in a client-facing or compliance-relevant workflow, could you explain its authority, its access, its oversight checkpoints, its evidence trail, and its kill switch to a regulator — with artefacts rather than explanations?
Next in the series — Part 6: AI shockwaves. Why internal efficiency is the first wave and why the second wave reshapes markets, pricing, and client expectations entirely.
Sources: Gartner CIO Value: From AI Tools to Managed Enterprise AI Services (2026d); Gartner Business Quarterly 1Q26 (2026a); Microsoft Copilot Studio Licensing and Billing Documentation (2025); Jersey AI Playbook, March 2026.
Professional disclaimer: The views expressed are my own and are shared for general informational purposes. This content is intentionally non-confidential and does not include proprietary, client-specific, or employer-sensitive information. It is not legal, regulatory, investment, or professional advice; readers should seek appropriate advice for their circumstances.
Financial services leader (20+ years) | Lean Six Sigma Green Belt | London School of Economics (LSE) AI Leadership Accelerator
See where your firm stands
The free diagnostic scores your firm across the twelve operating-model instruments in about eight minutes.
Take the diagnostic →A version of this article was first published on LinkedIn as part of the Sequence body of work. This is the canonical version.
The views expressed are the author's own and shared for general information. This content is intentionally non-confidential and does not include proprietary, client-specific or employer-sensitive information. It is not legal, regulatory, investment or professional advice. © 2026 Paul Soares. Sequence is the trading name of Sequence Advisory Limited (in formation) · hello@sequenceadvisory.com