We help companies prove their AI is still earning its keep.
Your agents went live. Somebody signed off on what they were worth. A year on, can you show a board they are still doing it — or prove when they stopped?
A vendor updates the model. A data source goes stale. Somebody edits an instruction without a change record. The reviewer who was reading every recommendation starts approving them in batches. Nothing breaks. The agent still runs — it has just stopped doing what you bought it for.
Drift sends two bills. One lands on the CFO: the agent quietly stops paying for itself. The other lands on the risk committee: the agent quietly starts doing something nobody approved. Same silence, same cause — two different people who find out late.
And with no care, it only ever gets worse. Never better.
The inference paradox. The price of a token keeps falling, so everyone assumes AI is getting cheaper. But agents call other agents, challenge each other and retry until they agree — so the tokens burned to finish one task climb. Your unit price drops while your bill rises. That is why the two lines below move apart instead of together.
Illustrative. The shape is the point: value bends down while the cost of running the same work bends up — and the two lines cross long before anything looks broken. Without a written starting point there is nothing to measure either bend against.
The check
Find out whether you need us.
So — is yours? Five questions, four minutes, scored against the five things that have to be true before anyone can say an AI program is still sound. Answer honestly; the score is only useful if it is true. Nothing is stored, and the score is yours without an email — only the longer read-out asks for one.
Question 1 of 5
—
Your Day-Two readiness
—
Overall readiness0 / 100
Want the full read-out?
All five anchors written up, what each gap is likely costing you, and where we would start. Your score above is yours either way — this is the longer version.
Putting AI to work in a regulated business is three different jobs, done in order. Most firms are sold the first two and discover the third on their own.
Day 0 · Decide
Where to point it
Which priorities matter, which processes carry the value, and what an agent is worth before anyone builds it.
Strategy tooling
Day 1 · Build
Make it and ship it
Engineer the agent, redesign the work around it, and record what it was approved to do on the way out the door.
Build partners
Day 2 · Steward
Know it still works
Detect the drift, measure what it costs, decide what to do about it, and put a name on the answer.
Dai2 Advisory
The engagement
An instrument, a steward, and an opinion.
A methodology on its own is a binder. What a board can act on is a person who read the evidence and put their name to a conclusion.
The instrumentHow we measure
A scored readiness diagnostic, a drift and cost model built on cost per completed task and value against full cost of ownership, and a standard evidence pack. The same discipline on every engagement, so results are comparable quarter to quarter.
The stewardWho decides
A named person who reads what the instrument produces and decides what happens next. Where the decision points are, who is accountable, where the audit trail lives — and when an agent should be retired rather than defended.
The opinionWhat you receive
A quarterly statement your board can rely on: which agents are still doing what they were approved to do, which have drifted, what it is costing, and what we recommend. Signed.
How you engage.
Three steps, and the first one is free. You cannot steward what nobody measured, so the paid work starts by writing down where things actually stand.
Step one
The Check
Five questions, four minutes, scored against the anchors. Tells you whether you have a Day-Two problem and roughly how big it is.
No cost · email for the read-out
Step two
The Baseline
One engagement. We record what every agent was approved to do, what it is actually doing now, and what the difference is worth. This is the starting line everything after it is measured against.
Fixed fee · one engagement
Step three
The Standing Opinion
A named steward reads the evidence every quarter and tells your board what is still true — in writing, with their name on it. A number your CFO can defend, and an answer your risk committee can file. The drift is caught while it is cheap.
Retainer · quarterly opinion
And this is what arrives.
Illustrative, with sample figures. The format is the product.
Quarterly Stewardship Opinion
Q3 2026 · Illustrative
14
Agents under stewardship
11
Performing to approval
3
Materially drifted
$1.8m
Approved annual value at risk
Agent
Approved to do
Cost / task
Value multiple
Finding
Deposit beta monitor
Flag funding drift vs. model
$0.41
6.2×
On approval
KYC file assembler
Pre-build review packets
$1.07
3.4×
On approval
Fee exception triage
Route and justify waivers
$2.88
1.1×
Cost drift
Complaint summarizer
Draft regulator-ready summaries
$0.96
0.7×
Below approval
Collateral revaluation
Nightly re-mark of pledged assets
$0.33
9.8×
On approval
Reviewed and signed
Steward of record · Dai2 Advisory
Human review rate this quarter
38% → 11%
The last figure is the one that ends careers. Reviewers were reading a third of recommendations in July and one in nine by September. Nothing alerted on it.
The method
Five anchors.
What has to be true before anyone can say an AI program is still sound.
Five anchors hold a program to what it was approved to do. Drift is any one of them going slack — with nothing announcing it. The anchors are the method; the baseline is what makes the slack measurable.
What it leaves behind01agent registry · named owner · autonomy tier
01
Know which agent did what
Every agent has a name, an owner, defined permissions and an autonomy tier.
02
Treat instructions as policy
The prompts that drive an agent encode business rules. Version them, review them, own them.
03
Design where humans stay
Decide deliberately where authority remains human — and keep reviewers actually reading.
04
Measure continuously
Drift happens when nothing changes. Track cost per completed task against the value the agent was approved on — not a test at launch.
05
Build the people who run it
The team that operates and oversees agents is not the team that built them.
—
And somebody signs
Anchors are a methodology. A methodology without a steward is a binder.
Who
Three co-founders.
Alan Babbitt
CEO & Co-Founder
An extended run at Oracle selling and delivering the enterprise systems the largest banks and insurers still run on. Later an Executive Partner at New York Life, leading advisor teams whose work grew client assets and built the tax strategies that protect them — and serving as the office's internal AI educator. Founder of Intelligent Systems Advisory. Decades spent learning how decisions this size get made, and how to be useful to the people who live with them.
Richard Mohrmann
Chief Technology Officer & Co-Founder
Leads technology and architecture at Dai2 Advisory, helping regulated enterprises deploy agentic AI safely. Brings 30+ years building mission-critical systems at Morgan Stanley, Merrill Lynch, BNP Paribas and Thomson Reuters, from low-latency trading to market data infrastructure. Currently directs strategy, governance and security architecture for enterprise-scale agentic AI at a global financial institution.
Aimen Ahmed
Chief AI Officer & Co-Founder
Two decades building operating models and governance frameworks across global financial services, including Citi in Hong Kong and New York and Bank of America Merrill Lynch. Chairs cross-functional governance spanning cybersecurity, technology risk, procurement and HR — with decision authority over onboarding, exceptions and remediation — for platforms serving a quarter of a million employees at one of the largest financial institutions in the world. Leads Dai2’s agentic AI: agents that triangulate people, process and product across lines of business, vendors and companies to measure what an AI program is worth.
Read me
Two papers. One for the board, one for the people who have to defend it.
The executive paper names the problem in the language a board uses. The technical companion is the framework underneath it — written for the engineers and risk officers who have to make it hold.
Executive paper · 6 pages · PDF
The Day-Two Problem
Summary
Written for the people who approved the AI initiative, not for the engineers who built it. What most companies miss after they launch AI — and the dividend earned by those who get it right.
Key points to consider
AI does not fail loudly. Value bends down while the cost of running the same work bends up, and nothing alerts on either.
The three failure modes that surface around ninety days after go-live, long before anything looks broken.
Five anchors that have to hold before anyone can honestly say a program is still sound — and a five-question diagnostic against them.
Why this is important
If you cannot show a board what an agent was approved to deliver, you have no way to show them it still does.
Technical companion · 12 pages · PDF
Agentic AI Governance Framework
Summary
The engineering companion to the executive paper. A five-pillar model for deploying agentic AI in regulated environments, written by Richard Mohrmann for the engineers and risk officers who have to make it hold.
Key points to consider
Each pillar laid out the same way: the problem it solves, its core components, and how it aligns to existing standards.
Agent identity and trust, knowledge capital, human oversight, continuous quality, and the workforce pillar most frameworks omit.
A maturity model that says what good looks like at each level, so you can place yourself honestly rather than aspirationally.
Why this is important
Most governance frameworks were not designed for systems that are autonomous by design. This one was, and it was built from deploying them at enterprise scale.
Contact
Start with the conversation, not the engagement.
Bring one AI program you have already shipped. We will tell you what we would want to see to know whether it is still doing its job — and whether that is worth paying anyone for.
Four minutes · right now
Take the check
Five questions against the five anchors. It scores you out of a hundred and tells you plainly whether you have a Day-Two problem. No cost, no call, no one follows up unless you ask.
You have a number and you want to know what it means. Tell us when suits and one of us calls — a steward, not a salesperson. You will get an opinion on the call, not a deck afterwards.
Name one AI program you have already shipped. We will tell you what we would want to see to know whether it is still doing its job — and whether that is worth paying anyone for.