Agentic Analytics in Production: What Snowflake CoWork Replaces, and What It Doesn’t
Douglas da Silva | Jul 27, 2026
Most executives treat an AI governance framework as the thing that slows AI down. The evidence points the other way. PwC’s 29th Global CEO Survey found that 56% of CEOs saw neither revenue gains nor cost reductions from Artificial Intelligence over the past year — while the leaders whose organizations had built responsible AI frameworks and enterprise-wide integration were three times more likely to report meaningful financial returns. Governance is not the tax on AI adoption. It’s the condition that makes adoption pay.
See what a minimum viable AI governance framework contains, how to tier use cases so oversight matches risk, what changes once agents enter production, and how to measure whether the framework is doing its job.
Why does shadow AI create liability before it creates value?
The pattern in mid-market regulated firms is consistent. Adoption has already happened — informally, through personal accounts, with no inventory of what’s in use and no policy governing it. Leadership discovers the extent of it only when someone goes looking.
That creates two distinct exposures, and they compound. The first is contractual. Enterprise AI licenses carry data protections; consumer accounts don’t. Every prompt containing client information that passes through a personal account leaves the firm’s perimeter with no agreement behind it.
The second is evidentiary. When a firm has no record of which tools were used, by whom, and for what purpose, it has nothing to show a regulator or an opposing counsel. Documented adoption of a recognized framework is what demonstrates reasonable care — and supervisors consistently credit active, documented progress over a perfect program that doesn’t exist yet.
The risk is not hypothetical for firms already moving on agents. McKinsey’s 2026 AI Trust Maturity Survey put average responsible-AI maturity at 2.3 out of 5, with nearly two-thirds of respondents naming security and risk as the top barrier to scaling agentic AI — ranked ahead of regulatory uncertainty. What stays undefined becomes a risk and what gets defined becomes defensible.
Read more: Conversational AI: How to Design and Build Scalable AI-Powered Features
What does a minimum viable AI governance framework include?
The NIST AI Risk Management Framework is the practical default, because it’s recognized, voluntary, and profile-based — you implement the subset that matches your risk, and the documentation of that choice is itself part of the defense. It organizes work into one cross-cutting function and three operational ones.
| Function | What it means for Kingsview | Maps to what Howard already wants |
|---|---|---|
| Map (principles, inventory, document, educate) | AI principles doc; turn his 3–4 week inventory into a living AI use-case registry; interim policy + signed acknowledgment + training video | His inventory + interim policy — validate and formalize it |
| Manage | Who’s in charge: AI steering committee connecting compliance, IT, and the adviser org; decision rights per risk tier | His “roles/hiring guidance for IT and compliance” ask |
| Measure | Evals before rollout, ongoing monitoring, incident reporting, audit trail | His “compliance must audit who uses what, for what” ask |
| Govern | Tone from the top, accountability culture, chargeback discipline | His adviser chargeback model |
- Govern sets accountability: who signs the policy, who holds decision rights at each risk tier, and who owns the outcome when something fails. It informs the other three functions rather than following them.
- Map builds visibility: an AI principles document, a use-case registry that turns a one-time inventory into a living record, an interim policy with signed acknowledgment, and role-appropriate training.
- Measure creates evidence: evaluation before rollout, monitoring after, incident reporting, and an audit trail that answers who used what, for what, and when.
- Manage turns that evidence into control: risk-tiered approval paths and a steering group that connects compliance, IT, and the operating business because a committee made only of one of those three will either block everything or approve everything.
For a firm under 1,000 people, the minimum credible version is not certification. It’s a documented profile: a registry, a risk-tier matrix, a policy v1, a training program, and a committee charter with named owners. Certification can wait. Documentation cannot.
What makes the profile operational rather than decorative is a single intake path. Every AI request moves through the same six gates: register it, assign a risk tier, approve it at the level that tier requires, pilot it against one KPI and its guardrails, scale it under standardized controls, then monitor it against the same measures that justified it.
The registry stops being a document at that point and becomes the front door. It’s also the metering base — the record that lets a firm attribute AI cost to the business unit consuming it, which matters in any organization where teams carry their own P&L.
Read more: AI Use Cases & Applications: How Businesses Are Leveraging AI
How do you decide which AI use cases need human oversight?
Tier by two variables: how sensitive the application is, and how much control the model holds over the decision — whether it supplies an input, sets a default, or decides outright.
That grid produces four oversight levels, chosen by risk and reversibility:
- Assist – high risk, hard to reverse: the model drafts, a human decides. In regulated advisory work, the compliance line and the oversight line fall in the same place, which simplifies the design considerably: prediction can belong to the model, judgment stays with the licensed professional.
- Approve – high risk, easy to reverse: client communications and regulated marketing. The model produces, a reviewer signs off before anything goes out.
- Audit – low risk, hard to reverse: meeting summaries filed to records. Sample-review the output on a fixed cadence rather than gating every item.
- Automate – low risk, easy to reverse: internal search, scheduling, structured data entry against a warehouse.
One tier sits outside the grid. Where the model would decide and the application is highly sensitive, generated portfolio recommendations, for example, the answer is prohibition.
Write the reasoning into the policy, not just the rule. A prohibition people understand survives staff turnover; a prohibition they don’t understand gets worked around.
What controls does agentic AI add to an AI governance framework?
Trust requirements escalate by architecture, and the rungs don’t skip. Generative AI needs review and disclosure discipline. A single agent adds logging, human override, and controllability. Multi-agent systems add coordination protocols, standard operating procedures, and lifecycle management for the agents themselves as creation, permission changes, retirement.
McKinsey’s 2026 data shows why this sequencing matters now: 62% of organizations are at least experimenting with agents and 23% report scaling them somewhere in the enterprise. Governance maturity has not moved at the same rate.
The control layer that makes agents auditable has five components: permissions enforced at the tool level rather than in the prompt, control agents that check outputs inside the workflow, evidence annotation so every answer carries its source, confidence scoring that routes low-confidence cases to a human, and audit logging on by default.
Tool-level permissions carry the most weight, and the reason is architectural. When independent practices share an agent library but legally cannot see each other’s client data, a prompt instruction it’s a request. Enforcing isolation in the permission layer turns a policy commitment into a system property.
Read more: Private Equity is Committed to AI: Here’s Why Most Portfolio Companies Aren’t Ready to Collect
How do you measure whether an AI governance framework is working?
Every AI initiative should deliver with one primary KPI and two or three guardrail metrics across three dimensions: flow (cycle time, throughput), quality (exception rate, audit pass rate), and economics (cost-to-serve, capacity redeployed).
Drop “hours saved” as a headline metric. Saved hours are not a result until they’re redeployed into throughput or capacity — measuring them rewards the appearance of efficiency instead of the fact of it. And a single metric, however well chosen, becomes a target and then a distortion. Guardrails exist to catch that.
The harder truth is that tooling is the smaller half of this work. Strategy clarity, skills, process redesign, and change management carry more of the load than the platform decision does — which is why PwC’s return data separates firms by the strength of their foundations rather than the size of their AI spend.
Why does AI governance succeed or fail on adoption rather than tooling?
Tooling is the smaller half of this work. Prosci’s study of 1,107 professionals found that roughly 63% of AI implementation difficulties traced to human factors rather than technical ones — user proficiency alone accounted for about 38%, against 16% for technical issues. A framework nobody follows produces the same audit trail as no framework at all.
Three things determine whether the policy holds in practice.
Start with the people already using AI well. In most firms a single team, often marketing or operations is well ahead of everyone else. The instinct is to bring them into compliance first. The better move is to make them the reference implementation: govern their existing workflow, document it, and let it become the template. Blocking your most capable users teaches the organization that governance means friction.
Then recruit champions inside each business unit, because top-down mandates fail where teams own their own results. Middle management is where adoption stalls and managers translate strategy into daily behavior, and they can’t do that for a framework they only read about.
Finally, train for fluency rather than compliance. A signed acknowledgment proves someone received the policy. It doesn’t mean they can tell which tier their next task falls into. That judgment is the actual control, and it has to be taught.
The payoff shows up in the returns data: McKinsey’s research indicates that firms seeing significant financial gains from AI are roughly twice as likely to have redesigned end-to-end workflows rather than automating individual tasks. Redesign is organizational work, not a procurement decision.
Read more: Your AI Strategy Has a Data Problem
Governance is what lets you move
Regulated firms don’t choose between speed and control. They choose between governed AI and unlogged AI, and only one of those can scale past a pilot.
If AI is already running in your organization without a registry, a policy, or an audit trail, the exposure exists whether or not the framework does. Book a call with us! We’ll tell you exactly what we’d govern first, and what we’d leave alone.
Or discover where your organization stands on the path from AI experimentation to true business transformation, and what to do next with our free AI Readiness Assessment:
About the author.
