Volume X: The AI Operating Layer

What the volume does

Where AI belongs in the business, and who governs it.

Each candidate deployment is read through the four lenses: what it returns against what it costs to run, what it does to how the business is understood by customers and regulators, what future capability it opens or forecloses, and whether this organisation can operate and supervise it.

The fourth lens is decisive in this domain more than any other. Most failed deployments were not bad ideas. They were reasonable ideas placed in organisations with no capacity to supervise them, which the business case never asked about.

The AI layer runs seven instructions. Deploy. Hold. Trial. Replace. Withdraw. Pilot. Reframe. Trial and Pilot together are the landing point for roughly 40 per cent of readings in this domain, which is a fair reflection of how much is genuinely unsettled. The readings sit against a design target of 0.78 from a development cohort of fourteen engagements between 2022 and 2026, the lowest of any volume and stated at that level deliberately.

How a reading works

How a reading of the AI layer works.

You provide the working record: what AI runs in your business, what each piece costs, who uses it, and the contracts and people that keep it alive. Every position gets read, from the chatbot added in a rush to the model nobody has touched since it was built. Most were assembled one urgent decision at a time. The reading treats them as one architecture.

Each position is read four ways: what it earns or saves, how customers and staff experience it, what it opens up later, and what could go wrong, from quiet accuracy decay to the obligations regulators attach to automated decisions. The four readings combine into one view with a stated range. A wide range and a narrow one lead to different calls.

What arrives is one instruction per position: deploy, hold, trial, replace, withdraw, pilot or reframe. Each carries a confidence level, a named owner inside your business, and the date the call gets checked against what actually happened. Expect a fair share of readings in this domain to land on trial or pilot, which is an honest measure of how much is still unsettled.

Six worked situations

What a reading actually produces.

Illustrative composites, drawn from the pattern of engagements rather than from any identifiable client. Figures are indicative.

Use case 1

The deployment that was ready and the organisation that was not

Pilot

Situation: A national law firm, roughly $180 million revenue, planning a firm-wide document review deployment.
They asked: How quickly can we roll this out?
The reading found: The tool read well on the first three lenses. The firm had no review protocol, no named accountable partner and no way to tell a good output from a plausible one at volume.
Instruction: Pilot in one practice group for two quarters. Build the supervision protocol before any wider deployment.
Why it matters: The constraint was never the technology. It was the absence of anyone who could be accountable for what came out of it.

Use case 2

The tool that had become load bearing without a decision

Withdraw

Situation: A mid-tier accounting firm, roughly $90 million revenue, where a customer-facing summarisation tool had spread through three service lines.
They asked: How do we scale this properly?
The reading found: The tool had never been assessed. It was producing client-facing output under professional obligations the firm carries personally, with no logging, no review step and no contractual position on the data.
Instruction: Withdraw from client-facing use immediately. Re-read for internal use once logging and review are in place.
Why it matters: Nobody decided to deploy it. That is the finding. Things that arrive without a decision leave without a plan, usually at the worst moment.

Use case 3

The automation question that was a process question

Reframe

Situation: A state government service agency, roughly 1,400 staff, seeking to automate a high-volume assessment workflow.
They asked: Which part of the workflow should we automate first?
The reading found: The reading stopped before an instruction. The workflow contained four steps that existed to correct errors introduced by a fifth. Automating any of them would have made the error correction faster.
Instruction: Reframe. Remove the source step, then re-read what remains.
Why it matters: Automating a broken process delivers the same result sooner and at greater volume. This is the most common finding in the volume.

Use case 4

The unglamorous use case that actually paid

Deploy

Situation: A logistics operator, roughly $400 million turnover, evaluating several AI proposals including a customer-facing assistant.
They asked: Should we build the customer assistant?
The reading found: The assistant read poorly on cost to run and worse on supervision capacity. A scheduling optimisation nobody had put forward read strongly on all four lenses and needed no new governance structure.
Instruction: Deploy the scheduling application with a named owner. Hold the customer assistant.
Why it matters: The proposal that arrives with a business case is not necessarily the one that pays. It is the one that had a sponsor.

Use case 5

The decision that could not be made yet

Hold

Situation: A superannuation fund, roughly $18 billion under management, considering AI in member-facing advice.
They asked: Do we deploy in advice?
The reading found: The reading declined to resolve. The regulatory position on automated advice was in active movement and any instruction issued would have been overtaken inside the check period.
Instruction: Hold. Re-read in two quarters or on regulatory change, whichever comes first.
Why it matters: A method willing to issue a confident instruction into a moving regulatory position is telling you about its incentives, not about your business.

Use case 6

The model that was state of the art in the year it was built

Replace

Situation: A food and beverage manufacturer, roughly $260 million turnover, ran demand planning on a machine-learning model built in 2019 and maintained by one specialist. The model sat outside the firm's AI conversation entirely, filed under operations.
They asked: Which of three proposed generative AI projects to fund first, and how to pay for the winner without a new budget line.
The reading found: The strongest move was not on the list. The 2019 model's accuracy had drifted for two years, its true cost hid inside one salary, and only one person could explain how it worked. Regulators increasingly expect a business to be able to explain and re-check any model that drives commercial decisions, and this one failed that test quietly. A current, supported tool would do the same job for less.
Instruction: Replace. Retire the 2019 model and move demand planning to a supported tool over two quarters, with the supply chain director as owner and a check at twelve months.
Why it matters: A decision made well in 2019 is not a decision about 2026. Old models decay without announcing it, and the funding for new AI is often sitting inside the old AI.

For the academics

Technology adoption research has been extending the technology acceptance model since 1989 and can now explain, retrospectively and with excellent fit statistics, why people used something they had already used. The governance literature runs on a separate track, borrowing model-risk supervision built for credit scorecards and applying it to systems that generate novel text. Volume X puts deliverability and supervision capacity into the same reading as the return, and carries the lowest design target of any volume because the domain deserves the lowest.

And commercially: it names who is accountable when the model is wrong.

The check

Every instruction has a date.

Twelve months after a reading, the instruction is compared against what happened and the result is recorded, whether or not it is flattering.

Discuss a reading All twelve volumes

The AI Operating Layer.

AI is being deployed by enthusiasm and governed by nobody.