Library

Agent

Software that can plan and take multi-step actions toward a goal, not only answer a question.

Why CEOs care. An agent can change a record, move money, or write to a customer. Residual risk is operational, not editorial.

Control plane

The layer that decides which agents may act, on which systems, with what logs and kill-switch.

Why CEOs care. This is how a board sees residual risk on software that acts. Without it, agent spend is unsupervised privileged access.

Residual risk

The risk that remains after controls — the number the board actually carries.

Why CEOs care. Perfect safety is not on offer. The artefact is an accepted residual, with an owner, not a promise of zero incidents.

Dual P&L

One page that shows AI cost takeout and AI revenue lift together, with residual risk noted.

Why CEOs care. Spend is visible. Combined proof is not. Boards are starting to ask for both on the same page.

RAG

Retrieval-augmented generation — the model is grounded in your documents at answer time.

Why CEOs care. It is the usual pattern for ‘chat with our knowledge’. It reduces some hallucination; it does not create a system of record.

Hallucination

A fluent answer that is not true, or not in the sources the system was given.

Why CEOs care. Fluent error is more dangerous than an empty answer. High-stakes workflows need evaluation and a human gate, not a disclaimer.

RPA

Robotic process automation — scripted clicks on existing systems. Not an agent.

Why CEOs care. Many ‘AI’ business cases are still RPA with a new label. RPA is brittle; agents are unsupervised if you let them be.

Sovereign stack

Models, cloud, and data that a jurisdiction can inspect, host, or refuse to let leave.

Why CEOs care. In APAC this is a market-access question: China, and increasingly parts of ASEAN and the Gulf, will not share a single US-cloud stack.

Foundation model

A general model trained at scale, then adapted. The raw material, not the product.

Why CEOs care. You rarely compete on training the base model. You compete on data, workflow, and control of how it is used.

Fine-tune

Further training of a model on your data so it behaves more like your domain.

Why CEOs care. Expensive, easy to over-claim, and often beaten by good retrieval plus evaluation. Ask what residual risk it actually removes.

Evaluation (eval)

A measured test of whether the system does the job — not a demo.

Why CEOs care. Without evals, ‘it works’ is a vendor story. Evals are how residual risk is measured over time.

Model card

A short document of what a model is for, what it was trained on, and how it was tested.

Why CEOs care. Becoming the nutrition label in procurement. Absence is itself a residual-risk signal.

Red team

An adversarial test that tries to make the system fail, leak, or disobey policy.

Why CEOs care. The adult version of ‘we tried it’. Buyers and AISI-style regimes expect evidence, not a slide.

Grounding

Forcing the answer to stay attached to retrieved or authorised sources.

Why CEOs care. Grounding reduces some hallucination. It does not make the model a database, and it fails if the corpus is wrong.

Context window

How much text (and now other media) the model can see at once.

Why CEOs care. A bigger window is not a memory strategy. Cost and error rates still rise with dumped documents.

Tokens

The units models read and write. The usual billable atom of inference.

Why CEOs care. For agents, tokens per completed workflow is the TCO line, not tokens in the abstract.

Inference

Running a trained model to produce an output. This is most of the bill after the pilot.

Why CEOs care. Training headlines are rare. Inference is the operating cost. Agents multiply it.

Latency

How long the system takes to answer or to finish a step.

Why CEOs care. Customer and trading workflows die on latency. A cheaper model that is late is not cheaper.

SLO

Service-level objective — the reliability bar you actually run to.

Why CEOs care. AI features that miss the SLO of the host product (banking app, claims, trading) will be turned off, whatever the demo.

Human in the loop

A person must approve before the system acts — or after, depending on the design.

Why CEOs care. A slogan unless you specify which actions, which people, and what happens when they are slow or wrong.

Kill-switch

A way to stop an agent or a class of agents immediately, without a release cycle.

Why CEOs care. If you cannot stop it, you do not control it. Boards already understand this sentence from cyber and trading.

Shadow AI

Staff using public models with company data outside the official stack.

Why CEOs care. The residual-risk you are not shown. Ban-only strategies usually fail; a sanctioned path plus monitoring is the adult version.

Model risk

The banking discipline for models that affect money, capital, or customers — now colliding with generative AI.

Why CEOs care. Regulated groups already have a model-risk function. Generative and agentic systems often sit outside it. That gap is the story.

TCO (AI)

Total cost of ownership: licences, inference, integration, evaluation, control plane, and people.

Why CEOs care. Seat prices are the smallest line. Agents make inference and integration the real bill.

Vendor mix

The deliberate use of more than one model and cloud, including a China or sovereign option where needed.

Why CEOs care. Single-stack convenience is a concentration risk. APAC groups often need two stacks whether they want them or not.

Licence to operate

The informal permission from regulators, buyers, insurers, and the public to keep using AI in the business.

Why CEOs care. Lost faster than it is won. Policy is now a product constraint because counterparties will not wait for your home statute.

Mixture of experts (MoE)

A model architecture that routes each token to a subset of specialists, rather than activating every parameter.

Why CEOs care. It changes the TCO story: more total parameters, less compute per token — if the serving stack can actually route.

Distillation

Training a smaller model to imitate a larger one.

Why CEOs care. The usual path from a frontier demo to a cost the line will accept.

Small language model (SLM)

A compact model, often specialised, that can run closer to the data.

Why CEOs care. On-prem, sovereign, and edge cases in APAC often need an SLM, not another frontier API.

Multimodal

A system that reads or writes more than text — images, audio, video, documents.

Why CEOs care. Claims, KYC, manufacturing, and clinical workflows are multimodal whether the vendor slide says so or not.

Embeddings

Numeric representations of text or images that make similarity searchable.

Why CEOs care. This is the index behind ‘chat with our knowledge’. It is also a copy of your corpus in another form.

Vector database

A store built to retrieve embeddings by similarity.

Why CEOs care. Brand of database is secondary. Access control and freshness are the CEO questions.

MCP (Model Context Protocol)

A standard way for models and agents to connect to tools and data sources.

Why CEOs care. A common plug is useful. An open plug without a control plane is a new attack surface.

Tool use

The model calling an external function — mail, ERP, browser, payments — as a step in its plan.

Why CEOs care. This is the line between answering and acting. Residual risk sits on the tool, not the paragraph.

Orchestration

The layer that sequences model calls, tools, and human steps into a workflow.

Why CEOs care. This is where TCO, latency, and residual risk actually live — not in the model card.

Model routing

Sending each request to the cheapest or best-fit model that can do the job.

Why CEOs care. The practical hedge against vendor mix and inference TCO.

Guardrails

Filters and policies that constrain what a model may take in or put out.

Why CEOs care. Necessary, not sufficient. Guardrails without tool-gating will not stop an agent that can act.

Prompt injection

Untrusted text that instructs the model to ignore you and use its tools.

Why CEOs care. Any agent that reads email, tickets, or the web can be instructed by that content.

Observability (AI)

The ability to see traces, costs, evals, and incidents across model and agent runs.

Why CEOs care. If you cannot see it, you cannot carry residual risk. The board will ask after the first incident.

Data residency

Where data, embeddings, and logs are allowed to live — and who can compel access.

Why CEOs care. In APAC this is often the binding constraint on a single regional product.

Data localisation

A legal requirement that certain data not leave a jurisdiction.

Why CEOs care. China, Indonesia, Vietnam, India and others already force product SKUs, not language packs.

Synthetic data

Data generated by a model to train or test another system.

Why CEOs care. Useful when real data is scarce or sensitive. Dangerous when it quietly replaces ground truth.

RLHF

Reinforcement learning from human feedback — how many frontier models are steered toward preferred answers.

Why CEOs care. It explains tone and refusal. It does not create a system of record or a control plane.

Copilot vs agent

A copilot drafts; an agent acts. Do not fund one with the other’s risk language.

Why CEOs care. Seat rollouts are copilots. Production workflows that change records are agents.

Complementary investment

The process, data, and people spend without which model spend does not return.

Why CEOs care. This is why Trailblazers out-earn Followers at the same model price.

Switching cost

What it would take to leave a model, cloud, or copilot — data, prompts, evals, habit.

Why CEOs care. Concentration is residual risk. Vendor mix is the hedge.

PwC Global CEO Survey — value, workforce, confidence

How peer chief executives describe AI value versus what they can show.

Trailblazer versus Follower

BCG’s split is now a capital-allocation lens, not a culture slogan.

The agentic era, at CEO altitude

Software that acts, not software that answers. Trust becomes architecture.

Why evaluation is a management system

Demos do not Measure. Residual risk is unmeasured without evals.

Work redesign, not ‘assistance’

Value shows up when a role is redesigned. Assistance language hides that.

Licence to operate is now a product input

Buyers, insurers, and the public will not wait for your home statute.

Generative is not the credit-score model

Same word ‘model’, different residual risk and a different control set.

Your data, their weights

The durable advantage is usually the workflow and the corpus, not the foundation model.

What an enterprise control plane actually is

Identity, tool-gating, data scope, logs, evaluation, kill-switch. Not a chatbot admin page.

Agents as principals

Give acting software the same identity discipline you give a contractor with system access.

Retrieval is an access-control problem

RAG without document-level permissions is a leak with a nice UI.

The evaluation harness

A standing test set and online monitor. Not a quarterly slide.

Two-stack reality in Asia Pacific

US-cloud plus a China or sovereign path is often the honest architecture.

Inference cost as a control

The control plane should cap spend the way it caps tools.

Transformation envelope versus line P&L

Say which one you are running. Mixing them is how returns disappear.

How to write a dual P&L

Spend, cost takeout, revenue lift, confidence, residual risk, owner. One page.

Options, not programmes

Fund the next proof, not the five-year platform, until the control plane exists.

Unit economics the line will own

Cost per completed workflow. Not cost per seat, not cost per token.

Agents are a privileged-access problem

If it can call a tool, treat it like a contractor with production credentials.

Prompt injection, in one paragraph

Untrusted text that instructs the model to ignore you and use its tools.

Data egress and shadow AI

The residual risk you are not shown is usually a paste into a public model.

Designing a kill-switch that exists

Owner on the roster, tested path, class-level stop, a log. Not a vendor ticket.

AI-native versus AI-enabled

One redesigns the work around models. The other bolts a copilot onto last year’s process.

Build, buy, or partner

Three ways to get a workflow into production. Most firms mix them without saying so.

The data flywheel, without the slogan

Usage creates labelled traces; traces make the next model or retrieval set better; better systems get more usage.

Open weights versus closed APIs

A cost, control, and residency choice — not a values debate.

Evaluation as a product

Evals are not a QA phase. They are the product the CEO can inspect.

Compute and energy as a constraint

Tokens are not free. Clusters, power, and cooling now show up in capex conversations.

Training data and intellectual property

Who may train on what — and who inherits the lawsuit.

Point solutions versus a platform

Forty copilots are not an operating model.

The talent market is not the model market

You cannot hire your way to a dual P&L, and you cannot skip complementary skill.

Unbundling of enterprise software

If the interface is a prompt, the suite tax comes under review.

Fallback and graceful degradation

What the customer sees when the model, the region, or the vendor is down.

What ‘value’ is allowed to mean

Time saved, cost out, revenue in, risk avoided — pick the unit before the programme.

Where does AI capital sit, and who can stop it?

Is the budget a transformation envelope or a set of line P&Ls with a kill criterion?

What residual risk do production agents carry today?

Not the policy. The acting workflows: money, records, customers.

Can we show cost takeout and revenue on the same page?

The dual P&L. Activity is not a return.

Which standard do we claim when a buyer or insurer asks?

NIST, IMDA, EU Act, or a private story?

What happens in the first hour of an agent incident?

Owner, kill-switch, customer and regulator path.

Are we a single-stack firm in a two-stack region?

China, and some ASEAN and Gulf buyers, will not share the US-cloud default.

Does the CEO personally own the AI decision?

BCG: 72% now say they do. The board should hear whether that is true here.

How much of the workforce has been upskilled — and on what?

Trailblazers train widely. Tool rollouts without skill are shadow AI.

Where does AI sit in the risk appetite statement?

If it is silent, the firm has no agreed residual.

Which counterparties can already force our documentary pack?

European buyers, insurers, banks, public tenders.

What would we stop funding this quarter?

A board that cannot name a stop has not allocated.

Who is the single owner of each acting workflow?

Not the platform owner. The P&L owner who carries residual risk.

Which two workflows consume this quarter’s capital?

Concentration is the Trailblazer habit.

Does every production agent have an identity and a kill-switch?

Shared service accounts are unsupervised privileged access.

What is cost per completed workflow, not per token?

Inference TCO hides in token dashboards.

Where is shadow AI, and what is the sanctioned path?

Ban-only strategies fail. Staff will paste into the public model.

Which evaluations run before a release — and after?

Demo is not Measure.

How does the China (or sovereign) stack report into the same ELT pack?

A regional control plane that cannot see one P&L is not regional.

Who is paged when an agent acts outside policy?

If the answer is ‘the vendor’, you do not own the residual.

What do we tell customers we automated — and what we did not?

Disclosure is now a product and trust question.

Which roles are we actually redesigning, not ‘assisting’?

Assistance language hides operating-model change.

Is model risk extended to generative and agentic systems?

Or does the MRM policy still assume a credit scorecard?

What is the vendor-exit path for the primary model?

Switching cost is now a strategy item.

Which data classes may never enter a public model?

A short, enforced list beats a 40-page policy.

How many hours did this ELT spend using the tools last month?

Leaders who do not use the stack cannot allocate capital inside it.