The Defensible AI Assessment

Which AI to deploy, which to pilot, which to hold. And proof it survives an examiner.

A fixed-fee diagnostic, 2–4 weeks, answering one executive question: where can this institution deploy AI to create real value, in a way it can defend to a regulator, an auditor, and its own board?

Who it's for

Institutions where every decision has to be examinable

Banks, fintechs, and lenders putting AI into lending, payments, AML, and compliance. If an examiner can ask you to reconstruct a decision, this assessment was built for you.

Not in finance? The same assessment runs for nonprofits, legal teams, and other organizations under scrutiny: same spine, different regulatory hooks.

What it does

Scores every use case, returns a verdict

Every AI use case is scored on business value and on data & technical feasibility. Regulatory defensibility is applied as a gate, never a score. Each use case leaves with a verdict: deploy, pilot with controls, or remediate first.

AXIS 1 · SCORED

Business value

What the use case is actually worth: revenue, cost, risk reduction, measured against your numbers.

AXIS 2 · SCORED

Data & technical feasibility

Whether the data beneath the use case can support it today, and what it costs to close the gap.

A GATE, NOT A SCORE

Regulatory defensibility

Can you explain, evidence, and reconstruct the decision for whoever examines it? If not, value doesn't matter yet.

DEPLOY PILOT WITH CONTROLS REMEDIATE FIRST
Every use case travels this path and leaves with exactly one verdict.

When you need this

  • A vendor is pitching AI for a regulated decision, and compliance is nervous.
  • The board wants an AI plan, and nobody trusts the data underneath it.
  • A pilot has been stuck in review for months, and nobody can say exactly why.
  • AI spend is rising, and nobody can point to the decision it improved.
How it runs

The 2–4 week arc

1
WEEK 1

Discovery & context

Stakeholder interviews, regulatory context, and where effort is bleeding.

2
WEEK 1–2

Data foundation review

The core: lineage, entity resolution, reconciliation, and controls.

3
WEEK 2–3

Scoring & prioritization

Every candidate scored on value, feasibility, and defensibility.

4
WEEK 3–4

Synthesis & readout

Opportunity map, roadmap, and an executive decision readout.

What we need from you: a few hours of stakeholder time across the engagement, read access to the systems in scope, and one point of contact.
What you get

Six deliverables, one decision

Everything here exists so you can move, and show your work when someone asks how you knew.

Executive decision snapshot: deploy / pilot / remediate, with the reason
Prioritized AI opportunity map (value × readiness)
Data-foundation scorecard: strengths and the gaps that block deployment
Effort-bleed findings: where time and risk are lost
Sequenced implementation roadmap
Proposal for the build, where warranted
SAMPLE · EXECUTIVE DECISION SNAPSHOT
Internal policy & procedure assistant
DEPLOY
Approved repositories, cited sources, no regulated customer decision.
AML alert triage & case summarization
PILOT WITH CONTROLS
High value, but keep it decision-support with logged human approval.
Reconciliation exception classification
PILOT WITH CONTROLS
Classify and route breaks, but don't auto-clear or post adjustments.
Credit decision automation
REMEDIATE FIRST
Lineage, explainability, and adverse-action evidence not yet sufficient.
Illustrative; every engagement is tailored to the client's data and regulatory context. Not legal advice.
Signature diagnostics

Failure modes a generic review misses

Three named diagnostics, drawn from live systems. Each earned its place by catching something real.

01

The capacity-blind origination engine

A lending vehicle budgeted origination against available cash, with no throughput constraint tied to facility size. The modeled book ran to roughly $57M on a $13M facility before anyone asked where the ceiling was.

LESSON
The constraint that matters is often the one nobody put in the model.
02

Hidden exposure along unmodeled dimensions

A market-neutral equity book was hedged on every named factor and still bled P&L. Regressing P&L on factor returns found an unhedged energy exposure sitting outside the model.

LESSON
The model answers every posed question correctly. The loss comes through the unposed one.
03

Aggregate performance that hides the consequential cases

A pre-LLM toxic-language classifier posted a strong headline metric, carried by the easy 99%. The false negatives pooled exactly in the subtle-harassment sliver the system existed to catch.

LESSON
With fat-tailed consequences, validate the error distribution and weight sampling by consequence, never by volume.
Engagement & timeline

Three tiers, all fixed fee

Focused

$10–15K
2 weeks · fixed fee

One team or one use case. The fastest way to a defensible answer.

Comprehensive

$40–50K
4 weeks · fixed fee

Institution-wide and board-ready.

Fixed scope, fixed fee, delivered by a senior principal who does the work.
Tony Daggett, founder and principal of ARDInsights
Delivered personally by Tony Daggett

Every engagement, start to finish: the interviews, the data review, the verdicts. 20 years building data and model systems across banking, funds, AI, and the nonprofit sector. LinkedIn

Fixed scope. Fixed fee. 2–4 weeks.

A free 30-minute intro call. You'll leave knowing whether the assessment fits, even if we never work together.

Book a free intro call