In a previous column, I closed with
Here's a test you can run in your next partner meeting. Ask the room to list every artificial intelligence tool the firm pays for. You'll get a longer list than you expect. You might have a Copilot in the document platform, something embedded in the tax software, a research assistant, and a meeting notetaker. You likely have two or three pilots somebody started, and nobody ever killed. Now ask, which decision in this firm got measurably better because of any of them?
Chances are, the question will stump everyone in attendance. That's the difference between a technology stack and an intelligence stack.
Your technology stack is organized around systems of record. You have practice management, document management, tax and audit engines, a CRM, a general ledger, a client portal, a payroll feed. You organize your tech stack by function, ensuring you know where the data lives. That diagram has served firms well for 30 years, and every system on it is still needed.
An intelligence stack is organized around cognition. It shows you how decisions actually get made in the firm, and what makes the next one better. In his book "The Organizational Singularity," Salim Ismail describes six layers plus one control plane running across all of them. Read it as a loop, not a pipeline.
- Purpose sets objectives, constraints and delegation boundaries. This is where your firm's purpose stops being wall art and starts governing behavior. Independence rules, risk appetite, client acceptance criteria, and the engagements you won't take. If you never tell an AI agent that your firm doesn't do a certain kind of work, it will cheerfully scope it.
- Sense collects signals like client system feeds, regulatory changes, engagement economics, staffing load and market movement. Its job is compression, turning noise into signal.
- Interpret turns raw signal into usable context: retrieval, memory, provenance, anomaly triage, and scenario framing. Most firms skip this layer, and that's an expensive mistake. Your interpret layer is the senior manager who knows why that client's inventory always spikes in the third quarter and why you shouldn't take the controller's explanation at face value. That knowledge is valuable, and right now it lives in one head, walks out the door at six and retires within the decade.
- Decide generates options, models consequences, ranks tradeoffs and either recommends or commits, always inside the boundaries set in the purpose layer.
- Orchestrate executes through tools, workflows, APIs, people and other agents, inside a bounded permission envelope.
- Learn compares what you expected with what happened, detects drift and updates the playbooks. Every cycle improves the next one. This layer turns spending into an asset, instead of a subscription.
- Govern and assure is a control plane that spans every layer. It incorporates identity, permissions, approval thresholds, policy enforcement, audit trails, kill switches and drift detection. It's never off.
Look at that list, and you'll notice a tool in the drawer usually lives in orchestrate. It executes a task, but it doesn't sense, interpret or learn, and nothing governs it. That is why productivity gains at the individual level are invisible at the firm level.
The stinging diagnostic
Ismail offers a three-part check he calls the Tokenmaxxing Test, and one part of it belongs in every managing partner's head.
It's the geometry check. If you map your agents one to one onto your existing org chart and workflow boundaries, you haven't transformed anything. You automated the org chart. Ismail compares this to the early electrification of factories. Plants bought electric motors and bolted them onto the old central drive shaft. Productivity barely moved for two decades. The boom came only when someone redesigned the factory floor around the motor.
Most firms bolt agents onto the drive shaft. Same review chain, handoffs and approval sequence, but now it's faster in spots. Individual tasks run five times quicker, and the engagement still takes the same number of days because the tasks were never the constraint. The decision layer was.
Harang Ju from Johns Hopkins published work in 2026 finding that roughly
Start with the data, not the tools
The newest thinking in this framework puts something ahead of the agents entirely: the data layer.
In most firms, the practice management system or ERP is the center of gravity. Everything else reaches into it. The AI initiative then becomes a long, fragile integration line stretching from a model to a database that was never designed to be reasoned over.
The inversion is to build an independent, governed, decoupled data layer at the center, and demote every application, including practice management, to one consumer among several. The system of record still holds the truth about state. If the new layer and the general ledger disagree about a balance, the general ledger wins. But meaning now travels with the data rather than being locked inside the application that created it.
This should feel familiar to accountants. We spent our careers insisting that a number isn't usable until you know what it represents, who asserted it, how you can use it and what happens if it's wrong. That discipline, applied to every data object an agent touches, is exactly what the framework calls for. We're not learning a new competency; we're applying an old one to a new substrate.
One related rule deserves emphasis because firms tend to get it wrong: Own your agent logs. The append-only record of every input, output, tool call and result is durable evidence of how an agent reasoned. If that log lives only on a vendor's infrastructure, under the vendor's retention policy, then the vendor owns the agent, and you can't replay, fork, export or migrate it. Whoever owns the log owns the agent. A firm that wouldn't let a third party hold its workpapers should not let one hold its decision trail.
Governance is the part we are already good at
The control plane resolves into four pillars: trusted evaluations, searchable logs with correlation identifiers, granular rollback and a human review queue. Score each one honestly on a one-to-five scale. Don't deploy a new class of agent until every pillar scores at least a three. Most organizations score ones on at least three of the four.
These pillars are not a competing standard. They operationalize what already exists, including the
Jeffrey Sonnenfeld and colleagues at the Yale Chief Executive Leadership Institute made the case that every board needs a
The failure mode to watch for is not dramatic. Agents rarely crash. They slowly degrade, drifting a few percent at a time until someone notices it in a client escalation rather than on a dashboard. Without evaluation thresholds and an override-rate ceiling, you'll find out from the client.
What to build first
You don't need a two-year program. You need a minimum viable intelligence stack, and it can stand up in a week: one event bus, a basic agent registry, central logging and one agent per class. Firms that skip this step regret it within sixty days.
Then pick one workflow. The right candidate has the highest ratio of coordination to judgment, runs at volume, has clear rules, is measurable and is reversible. Run it in parallel with the existing process, define your success criteria before the run starts and watch one number above all others: the human override rate. If it falls over time, you're building an intelligence stack. If it stays flat, you have workflow automation with a chat window on it.
In an episode of
The tools in the drawer are not the problem. The absence of an architecture around them is. Build the stack, govern it and let it learn. That's how the drawer becomes an operating system.
Think, plan, grow!
