For the first hire · the resident expert

Explain it, then decide it

When you're the person everyone asks "should we use AI for this — and how?", you need answers you can show. These are plain-English, opinionated explainers and decision aids — built for the AI lead who has to teach the rest of the org, not for engineers building pipelines.

One rule keeps this page honest: every tool here has to help you make a decision, or defend it to your team, Risk, Legal, or the Board. If it's just a tutorial you could Google, it doesn't belong.
Start here
Decision tools
Concept guides
Hands-on

The most common question you'll get is "can AI do this — and how?" The honest answer is usually "yes, but choose the lightest approach that works." Here's the decision, with each option explained — including RAG — and what each one means for governance.

Quick advisor — answer three questions

1. Does it need your own / internal knowledge (your documents, policies, product data) to answer correctly?
No — general knowledge is fineYes — it must use our content
2. Does it need to take actions in other systems (create records, send things, move money, update tickets)?
No — it just answers or draftsYes — it has to do things
3. Is it a narrow, repetitive task where you have lots of past examples and need a very consistent style/format?
No / not reallyYes — specialized & repetitive

The four approaches

Prompting

Effort: low · days
Just ask a capable model, with clear instructions and a few examples. No new infrastructure.
Best forDrafting, summarizing, classifying, Q&A on general knowledge.
Avoid whenIt must cite your facts, or hallucinations are costly.
Governance: lowest risk — but no grounding, so it can confidently make things up. Keep a human in the loop for anything consequential.

RAG (retrieval-augmented generation)

Effort: med · weeks
Give the model your documents at answer-time: it retrieves the most relevant passages from your content and answers grounded in them — with citations.
How it worksYour docs are chunked & indexed → the question fetches the closest passages → those are added to the prompt → the model answers from them.
Best forPolicy/knowledge assistants, support, "answer from our manuals," anything needing current internal facts.
Avoid whenYour content is thin, messy, or ungoverned — RAG inherits the data's quality and its access rules.
Governance: this is where data access becomes the risk. The assistant can surface whatever it can retrieve — so retrieval must respect the same permissions as the source. (This is exactly what the agent-access register and least-privilege control in the Operating Model are for.)

Fine-tuning

Effort: high · weeks–months
Train the model further on your examples so it absorbs a specific style, format, or narrow task.
Best forHigh-volume, narrow, consistent tasks where prompting/RAG aren't reliable enough.
Avoid whenYou actually need facts (use RAG) — fine-tuning teaches behavior, not knowledge. It's the most expensive to build and maintain.
Governance: training data becomes part of the model — provenance, consent, and IP of that data are now your accountability, and it's harder to "unlearn."

Agents

Effort: high · highest risk
Wrap a model with tools, memory and permissions so it can take actions, not just produce text. (See tab 2.)
Best forMulti-step workflows where the work is doing, not just answering.
Avoid whenA simpler approach works, or the actions are irreversible without strong controls. Most "we need an agent" turns out to be RAG + a human.
Governance: the highest-stakes choice. Risk scales with what it can touch and how autonomously it acts — govern it as an agent (owner, autonomy level, least-privilege access), not as a model.

The rule of thumb (and the cheapest path to value)

Climb only as far as the problem forces you. Each rung adds cost, risk, and governance overhead.

  1. Start with prompting. Astonishing amount of value lives here.
  2. Add RAG when it needs your own knowledge or must cite real sources.
  3. Fine-tune only when a narrow, high-volume task still isn't consistent enough.
  4. Reach for an agent only when the job genuinely requires acting in systems — and then scope its access tightly.

Build vs. buy

A separate question, and the evidence is blunt: across enterprise studies, buying or partnering succeeds roughly twice as often as internal builds (MIT, 2025). Default to buying the capability and owning the governance, integration, and change — build only where it's a genuine differentiator.

Where this connects

This is the lightweight version of the same call Assess makes for a specific workflow (is AI even the right tool, at what autonomy). Use this to teach and decide quickly; use Assess to produce the governed brief.

Before you green-light an AI use case, you need its risk tier — defensible, not a gut call. Answer five questions; get a tier, the controls it requires, and the most autonomy it should ever have. The same logic the Operating Model uses, in 30 seconds.

Profile the risk

1. If it gets one wrong, the impact is…
Low — cosmetic / minorModerate — rework, cost, annoyanceHigh — financial, safety, legal, reputational
2. Can a wrong action be undone?
EasilyWith effortIrreversible — money sent, message out, data deleted
3. What can it reach (blast radius)?
Read-onlyWrites to internal systemsActs in external / money / safety-critical systems
4. Most sensitive data it touches?
Public / internal · non-sensitiveConfidentialRegulated — PII, PHI, financial, special-category
5. How autonomously is it meant to act?
Suggests only — a human actsActs with approval each timeActs on its own
Where this connects

This is the quick version of the autonomy-by-consequence call Assess makes in full, and it tiers into the same control library you set in the Operating Model. Use it to triage fast; use Assess for the governed brief.

RAG — retrieval-augmented generation — is the most common way to make a general model answer from your knowledge instead of its own. Here's what it is, how it works, and the one thing about it that's a governance problem.

The one-sentence version: give the model the relevant pages from your documents at the moment of the question, so it answers from them — not from whatever it happened to memorize in training.

How to explain it to an exec: "It's an open-book exam. The model isn't remembering your policies — it's looking them up at answer time. Which means whatever it's allowed to look up, it can repeat."

Common confusion — "isn't RAG just pointing it at our wiki?"

Not quite, and the gap is the whole point. Naming a source isn't access: tell a plain model "the handbook is on our wiki" and — with no tool to fetch it — it can't read a word, so it paraphrases a guess. The real choice is how it retrieves. A live-fetch tool ("open the book to the page named and read it") is always current and needs no pipeline — but it grabs whole pages, has to know which one, and slows as content grows. RAG ("a librarian who has read everything, returns the three exact paragraphs that answer the question across all of it — with the source") wins on precision, scale, cost and citations. The trade-off runs the other way on freshness: a live fetch always reads the current page, while a RAG index is only as fresh as its last sync. Rule of thumb: one handbook that fits in the prompt → just hand over the doc; a whole wiki, many sources, citations, or per-person access → RAG.

How it works

1
IndexYour documents are split into searchable chunks and stored.
2
RetrieveA question pulls the closest-matching passages from your content.
3
AugmentThose passages are added to the prompt: "here's the question, and the relevant facts."
4
GenerateThe model answers grounded in them — and can cite which passage it used.
Question"What's our refund window?"
RetrievesReturns Policy §2 — "Items may be returned within 30 days of delivery…"
Answers"30 days from delivery." [Returns Policy §2]

When it's the right call

Reach for RAG when

It needs to answer from your current internal facts — manuals, policies, product data, tickets — and you want answers you can trace to a source.

Don't, when

Your content is thin, stale, or ungoverned (RAG inherits its quality), or you actually need a behaviour/style change — that's fine-tuning, not RAG. RAG adds knowledge, not skills.

For the full "prompt vs RAG vs fine-tune vs agent" decision, use the Approach advisor.

The governance catch — this is the part that matters

RAG can surface anything it can retrieve. If a user asks a question and the system pulls a passage from a document they shouldn't see, RAG has just leaked it. So retrieval must enforce the same permissions as the source — RAG inherits your data's access rules, not only its content. Two non-negotiables: scope retrieval to what the user is allowed to see (least-privilege, the same control you set in the Operating Model), and always show citations so every answer is verifiable.

"AI agent" is the most overloaded term in the building right now — and you'll be asked to make decisions about agents while the word still means five different things. Here's the version you can explain to anyone, with the part that actually matters for risk.

Model vs. agent — the distinction that matters

The model / LLM

The engine

Claude, GPT, Gemini. On its own it reads and writes text. It can't see your systems or do anything in the world — it just produces words.

The agent

The entity that acts

That same model, wired up so it can actually do things — call tools, read and write to systems, remember, and decide its own next step toward a goal.

Agent = model + instructions + tools + memory + data access + autonomy

How to explain it to an exec: "The model is the brain. The agent is the brain plus hands, keys, and a job — and we govern the hands and keys, not just the brain."

Autonomy is a dial, not a switch

The single most important governance decision about an agent is how much it's allowed to do on its own — and you set that by consequence and reversibility, not by how impressive the demo is.

On rails
Only pre-approved responses/actions. No freedom. Safest, least useful.
Fenced in
Narrow, rule-bound actions inside tight limits.
Check first
Proposes; a human approves every action before it happens. The default for anything consequential or irreversible.
Acts, you spot-check
Acts within limits; humans review by exception. Needs strong logging and easy rollback.
Runs itself
Self-monitors, flags only. Earn this — don't start here.

Why access is the real risk

An agent's danger isn't its intelligence — it's its blast radius. The same model is harmless as a read-only summarizer and dangerous with write access to your finance system. So the governing questions are always: what can it reach, read or write, and who approved that?

Excessive agency

An agent given more tools/permissions than the task needs. The fix is least privilege — grant the minimum access, nothing "just in case."

Prompt injection → tool misuse

Malicious text in a document or email tricks the agent into misusing its tools. The more it can do, the worse this gets.

Irreversible actions

Payments, deletions, sends. These need "check first" autonomy and a kill-switch — never "runs itself."

Segregation of duties

An agent that can both initiate and approve the same action breaks a basic control. Split them, as you would for a person.

Where this connects

You approve models once (the register in Operating Model); you govern each agent — its owner, autonomy, and least-privilege access. Assess sets an agent's autonomy by consequence; Monitor tracks what it touches and how it behaves.

Two words used loosely. The model is just an engine. Scaffolding is what's built around one model call to make it reliable; orchestration is what coordinates many steps into a workflow. Most of what looks like "the AI is clever" is good scaffolding and orchestration — not a smarter model.

Scaffolding = the structure around a single model call.   Orchestration = the conductor that runs many calls, tools and steps in order.

How to explain it to an exec: "The model is the engine. Scaffolding is the car built around it — the steering, the dashboard, the seatbelts. Orchestration is the route and the traffic lights for a whole trip with several stops."

Scaffolding — around one call

Make a single answer dependable

The instructions and structure wrapped around one model call so the output is reliable and usable: a clear system prompt, a structured output format (so the result is parseable, not loose prose), the tools it may call, worked examples, breaking a big ask into steps, and self-checking / retries when it gets something wrong. A capable model with strong scaffolding beats a stronger model with none.

Orchestration — across many steps

Coordinate the whole job

The control flow over multiple steps: what runs when, passing each step's output to the next, branching ("if it's a refund, do this"), looping, running calls in parallel, holding state across steps, and handling errors. The "conductor" — frameworks like LangGraph, CrewAI, AutoGen or n8n, and durable engines like Temporal for long-running flows.

The distinction that actually matters: workflow vs agent

Orchestration comes in two shapes, and the difference is a governance decision, not just a technical one.

Workflow · the developer sets the path

The steps are defined in advance — do A, then B, then if X do C. Predictable, testable, governable. The right default for most production work: the same input takes the same route every time.

Agentic · the model sets the path

The model decides which tools and steps to use at runtime to reach a goal. Flexible for open-ended work, but less predictable and harder to govern. The more the model controls the flow, the more it behaves like an agent — and the more it needs the autonomy + least-privilege controls in the agent guide.

The rule of thumb

Climb only as far as the task forces it: a single well-scaffolded call → a fixed workflow of a few steps → a free-roaming agent — last, and tightly scoped. Reliability and governability drop at each rung while cost and failure modes rise, so reach for the next rung only when the one below genuinely can't do the job. (Same discipline as the Approach advisor — most "we need a multi-agent system" turns out to be one good prompt with structured output.)

There's no perfect hallucination detector — anything claiming 100% is overselling. But the rate can be cut hard, most of what slips through can be caught, and the rest gated. The whole approach rests on one idea: it's hard to check "is this true about the world?", but easy to check "does this answer match a source we provided?"

Step 1 — give it a source (grounding).   Step 2 — check the answer against that source. Everything below is a variation on those two moves.

First — what "grounding" actually is

Grounding means the model answers from source text put in front of it at the moment of the question (via RAG, or just pasting the document) — not from the fuzzy memory baked into training. It's the difference between recalling and reading.

Ungrounded — from memory

Q: "What's our refund window?" → "30 days."
Plausible, possibly wrong — and the real problem is there's nothing to check it against. It's a guess recalled from training.

Grounded — from a provided source

Retrieve "Returns Policy §2 — items may be returned within 30 days of delivery," put it in the prompt → "30 days [Returns Policy §2]."
Now the answer is copied from real text and verifiable against it.

The quiet payoff: grounding doesn't only lower hallucination — it turns an unanswerable question ("is this true?") into a checkable one ("does this match the source?"). That's what makes every check below possible.

Then — what the "checks" actually look like

Concrete, in rough order of catch-per-effort:

Citation check
Make the model cite the passage behind each claim, then confirm in code that the cited source exists and actually contains it. Example: the answer cites "[Returns Policy §2]" — a script checks §2 exists and contains that line. If it cited "§9" (which doesn't exist), it's caught.
Faithfulness judge
A second model call with a fixed prompt: "Here is a SOURCE and an ANSWER. List every factual claim in the ANSWER and mark each SUPPORTED or NOT SUPPORTED by the SOURCE only. Return the NOT SUPPORTED ones." Each "not supported" is a caught hallucination. Off-the-shelf in RAGAS, DeepEval, TruLens.
Self-consistency
Ask the same thing 3–5 times. If a fact changes across runs, the model is guessing — flag it. Example: "penalty period?" returns 30 / 30 / 60 days → the disagreement is the red flag. Useful when there's no source to check against.
Structural checks
Deterministic, cheap, high-precision — they catch fabricated specifics. Every cited ID must exist in the database; every quote must appear verbatim in the doc; every figure must be in a sane range; every link must resolve. A made-up case number or invented SKU fails the lookup and is blocked.
Abstention
Instruct it to reply "Not stated in the provided sources" when the source doesn't cover the question — turning a confident guess into an honest gap. The cheapest check of all: don't let it answer from nothing.
Human gate
Sample-review routine output; require human sign-off on the high-stakes — proportional to the tier from the Risk profiler. The catch-all for the residual no automated check caught.
The leader's move — measure it

Pick a faithfulness rate (share of answers fully supported by their source) and track it on a test set and a sample of live traffic, so a prompt or model change shows up as that number moving. The honest stack: grounding does the heavy lifting → citation + faithfulness checks catch most of what slips → structural checks nail the fabricated specifics → a human gates the high-stakes remainder. No magic filter — layers.

When an agent goes into production it stops being a project and becomes a thing that acts on its own every day. Someone has to be accountable for it — the way a manager is accountable for a report. The most common governance hole isn't a missing control; it's an "owner" field with a name in it and nothing behind the name.

An owner is a name. An agent manager is a defined role.   The first answers "whose is it?"; only the second answers "who is on the hook when it does X — and what can they do about it?"
Owner — name only

"Owner: Head of AI Engineering." Comforting on a slide. But when the agent mis-routes a case at 2am, nobody knows who decides, who can pull it, or who answers to the board. The name is decoration.

Agent manager — a real role

A named person with written responsibilities, decision rights, escalation triggers, a review cadence and the authority to throttle or stop the agent. Accountability that survives contact with a bad day.

What the role actually is

Five things, written down before go-live — not inferred after an incident:

Responsibilities
What this person is accountable for: the agent's outcomes, its metrics staying healthy, and its behaviour staying inside scope. Day-to-day ownership, not a quarterly glance.
Decision rights
What they can decide alone vs. what goes up. e.g. they can tighten a prompt or narrow scope; expanding the agent's autonomy or data access goes to the Risk Board.
Escalation triggers
The named conditions that force a human up the chain — error-rate breach, an out-of-scope action, a stakeholder complaint, a security flag. Set in advance so nobody improvises the threshold mid-incident.
Review cadence
A standing rhythm — weekly on a new agent, monthly once stable — where the manager actually looks at what it did, not just whether it's up. The antidote to set-and-forget drift.
Override authority
The explicit power — and obligation — to lower autonomy, pause, or kill the agent without waiting for a committee. If no one can stop it on their own say-so, no one is really managing it.
The leader's move — make it a gate, not a nicety

No agent reaches production without a named agent manager and those five lines filled in — same status as a security sign-off. This is the 70% people-and-process layer made concrete: the cheapest, highest-leverage governance you can add, because every other control — identity, monitoring, thresholds — loses its teeth if no human is actually on the hook to act on it.

Every agent acts inside your systems as some identity. If that identity is a shared human login, or a long-lived service account nobody owns, you've created an ungoverned actor with standing access — one that acts on its own, around the clock, faster than anyone watching. Non-human identity is the discipline of governing the agent's own identity and secrets across its life.

Traditional IAM was built for humans.   Agents break three of its quiet assumptions at once — and that's where the risk lives.
What human IAM assumes

A person logs in, does a few things, logs out. Access is occasional, attended, and tied to someone you can call. Offboarding happens when they leave.

What an agent actually does

Persistent — credentials live as long as the agent. Autonomous — it acts with no human at the keyboard. At scale — one pattern spins up many agents. None of the human assumptions hold.

The controls that close it

Not a new platform — a handful of controls wired to your existing IAM / PAM:

Its own identity
Each agent runs as a dedicated machine account — never a shared mailbox, a developer's login, or a generic "agent_user". One agent, one identity, so every action traces to a specific actor.
Rotation & revocation tied to lifecycle
Secrets rotate on a schedule, and credentials are revoked the moment the agent is decommissioned or implicated in an incident. The classic failure is a long-lived key for an agent retired a year ago.
Attributable audit logging
Every machine action is logged against the agent's identity, not buried under a shared service account — so "what did this agent do, and when?" is answerable for auditors and incident response.
Least privilege, scoped
Pairs with the design-time access spec: the identity holds only the tools and data the agent needs, nothing inherited "just in case". Scope it at the agent guide.
The leader's move — three questions per agent

For every agent you put live, make someone answer: Does it have its own identity? Is there a rotation trigger? Is there a revocation path on decommission or incident? Three questions, every time, catch the failure modes that turn an agent into a standing breach. It's a control domain in your Operating Model and a line in every Assess governance pack — not a separate security product to buy.

The job of a new AI leader isn't to ship a few projects — it's to install the operating system the enterprise keeps running long after you. This is the guided version: score where your OS is today across the eight domains, see the load-bearing gaps, and get a 30/60/90 sequence to install it.

Self-assess → find the gaps → install in sequence.   A framing and structuring aid — you author the strategy, the standards and the talent calls; this makes them visible, measurable and governable.

1 · Where's your AI operating system today?

Eight domains. Score each on evidence, not optimism — Missing · Ad-hoc · Defined (written down, owned) · Operating (running, with proof).

1 · Strategy & roadmap — a prioritized AI roadmap tied to business strategy?
MissingAd-hocDefinedOperating
2 · Governance & risk — risk tiers, controls and approvals, integrated with Legal/Security/Risk?
MissingAd-hocDefinedOperating
3 · Use-case portfolio — one de-duplicated, prioritized view of every AI initiative?
MissingAd-hocDefinedOperating
4 · Delivery operating model — a repeatable stage-gated path from idea to scaled production?
MissingAd-hocDefinedOperating
5 · Technical standards — model/data/MLOps/security standards exist, with named owners and adoption evidence?
MissingAd-hocDefinedOperating
6 · Literacy & enablement — role-based AI literacy, playbooks and a champions network?
MissingAd-hocDefinedOperating
7 · Org design & CoE — the operating model (central / federated / hub-and-spoke) chosen, chartered and staffed?
MissingAd-hocDefinedOperating
8 · Value & board — value defined, measured (value/adoption/quality/risk) and reported honestly to the board?
MissingAd-hocDefinedOperating

2 · Your 30 / 60 / 90 — install the OS

Not a project plan — the install sequence for the system above. Leaders who succeed treat the first 90 days as installing the OS, not delivering AI. 0 / 20 done

Days 0–30 · Listen & inventory
Days 30–60 · Stand up the spine
Days 60–90 · Make it real + first board narrative

3 · Starter artifacts (you author, the suite structures)

The highest-leverage things to have on paper. The tool gives structure and makes them auditable; the content and the calls are yours.

Ambition canvas
Translate business strategy into 3–5 AI themes + a north-star. A framing aid — not a strategy generator. Open the Ambition canvas.
Risk-tiering matrix
Consequence × reversibility → tier + required controls. Start it in the Risk profiler; formalize in the Operating Model.
CoE charter
Central / federated / hub-and-spoke, with interfaces and decision rights between hub and spokes. Choose the shape in the CoE model tool.
Standards-existence checklist
For each standard: does it exist? who owns it? when reviewed? adoption evidence? Governs the existence of standards without you authoring the standards themselves. Score them in the Standards checklist.
The one line to remember

Measure your own success by the quality and adoption of the operating system you leave behind — the one that keeps the enterprise getting better at AI long after you move on. Not the number of pilots you shipped.

The first thing a new AI leader owes the business is a translation: from business strategy into a handful of AI themes anyone can repeat. This is the canvas for it — a framing aid, not a strategy generator. You fill it; the discipline is in what you leave out.

Business objectives → AI themes → one north-star → explicit non-goals.   If a proposed use case doesn’t ladder to an objective on this page, it doesn’t get funded.

The five boxes

Business objectives
The 2–4 strategic priorities AI must serve this year — lifted from the company strategy, not invented by the AI team. If you can’t name them, that’s the first conversation, not the first model.
AI themes
3–5 themes that ladder to those objectives — e.g. “cut cost-to-serve,” “compress cycle time,” “catch risk earlier.” Themes, not projects: each becomes a lane the portfolio fills.
North-star metric
The single business measure that says AI is working — not “models shipped.” One number the board already cares about.
Where we won’t play
Explicit non-goals and no-go zones. The restraint that keeps the programme from sprawling into every shiny demo — and the fastest way to earn trust with Risk and Legal.
Proof horizon
Near-term wins (credibility in 1–2 quarters) and the longer bets — so the story has both momentum and ambition.
The leader’s move

Author this yourself, on one page, and socialize it before you rank a single use case — the themes become the lanes in Prioritize, and the north-star becomes the top line in Monitor. The tool structures it and makes it auditable; the judgment is yours.

One structural choice shapes everything: how central is the AI capability? Get it wrong and the CoE becomes the bottleneck every request queues behind. Answer four questions for a recommended shape — and the charter it needs whatever the answer.

1 · Where does AI delivery capability live today?
Concentrated — a central fewMixedSpread across the business units
2 · Regulatory / risk intensity?
LowModerateHigh
3 · Central team capacity?
Tiny — a handfulSmallSubstantial
4 · AI maturity across the business units?
NascentMixedMature
Whatever the shape — the charter defines these

Mission & mandate (why it exists, what it owns) · what the hub owns vs. the spokes · interfaces & hand-offs between them · decision rights (who decides what, alone vs. escalated) · funding model · and a maturity path (how governance becomes more self-service over time). In regulated firms, governance stays central regardless of the shape. The tool helps you choose the model and structure the charter — you still write the actual content.

You don’t author standards content in a tool — that’s engineering and policy work. But you do govern whether each standard exists, is owned, and is actually adopted. Score the six that matter; the gaps are your standards backlog.

For each: does it exist? who owns it? when was it last reviewed? what’s the adoption evidence?   Score — Missing · Drafted · Owned (named owner, current) · Adopted (evidence it’s used).
Model development — dev, testing, documentation, model cards
MissingDraftedOwnedAdopted
Data & privacy — sourcing, lineage, retention, consent
MissingDraftedOwnedAdopted
MLOps / deployment — versioning, rollback, monitoring
MissingDraftedOwnedAdopted
Security & access — least-privilege, secrets, non-human identity
MissingDraftedOwnedAdopted
Responsible / ethical AI — fairness, transparency, human oversight
MissingDraftedOwnedAdopted
Evaluation & quality — evals, faithfulness, acceptance thresholds
MissingDraftedOwnedAdopted

CISO check — beyond these six: is your model-risk process visibly integrated with enterprise MRM / ERM (yes / partial / no)? That integration is what a regulator or CISO looks for first.

The boundary

This governs the existence and adoption of standards — not their content. Owners in Engineering, Data, Security and Risk write the standards; you make sure each exists, has a name against it, and is actually used. Wire the security line to the non-human identity controls and the risk tiers in the Operating Model.

Production AI runs on a handful of hands-on activities — prompt libraries, reusable skills, evals, guardrails, RAG, fine-tuning, LLMOps, and the operating model that holds them together. Each is below in plain English, with an honest read of how hands-on a leader can be versus what they direct — because most of it involves far less "coding" than the word "engineer" implies.

Do it low/no code — a leader can genuinely build this Direct it understand deeply; engineers implement Decide it mostly a yes/no call; rarely built by hand

1 · Prompt libraries & prompt management

Do it · little/no code

A versioned, shared collection of an org's best prompts — treated as reusable assets, not one-off chat messages.

What it isInstead of everyone re-inventing prompts in a chat window, the ones that work are captured as templates with variables, named, categorised, versioned and shared so the org reuses proven patterns. Often owned by a "prompt + skills" role — but the craft is mostly clear writing and disciplined curation, not engineering.
WhenThe moment more than a couple of people use AI for the same kind of task. Usually the first real "production" asset an org ships.
Good looks likeEvery prompt has a name, owner, version, intended use and an example in/out; the risky ones carry a "don't trust it for X" note; people reuse them without asking the central team.
How it failsA graveyard of prompts nobody trusts because none are tested or owned, or five versions that quietly drift apart. Treat it like a small content library, not a junk drawer.
In practice: ~20 templates in a shared doc, a Git repo, or a prompt tool (PromptLayer, Langfuse) — e.g. "Summarise a client policy → 5 board bullets {doc}, {audience}." Code involved: essentially none — it's writing and templates.

2 · Skills & reusable assets (e.g. Claude skills)

Do it / Direct · markdown + sometimes light scripts

Packaged, reusable capabilities the AI is given once and reused everywhere — often shipped as "skills."

What it isA Skill bundles instructions (and optionally a few helper files) so the model reliably does a repeatable job — e.g. "fill our engagement-risk memo to house format." A skill (e.g. a Claude skill) is mostly a markdown instruction file plus any supporting files: authoring the common ones is closer to writing a great SOP than programming.
WhenWhen a task repeats often and consistent, on-format output is needed without re-prompting each time — the step up from a prompt library toward shippable assets.
Good looks likeA small catalog, each skill with a clear trigger ("use when…"), tested on real cases, owned by a service line, and adopted across teams without central hand-holding.
How it failsOver-engineering a skill a good prompt would handle; or skills that silently break when the underlying process changes and nobody owns them.
In practice: a leader can author an instruction-only skill directly (write it, add a sample, test it). Skills that bundle a script (say, to parse a PDF) need an engineer — that's the "Direct it" half. Code involved: little for instruction-only skills; some for script-backed ones.

3 · Evals & evaluation

Direct it · design-led, light code to run

How a team proves an AI output is actually good — before it ships and continuously after. One of the most common technical gaps in AI programmes.

What it isA repeatable test of quality against criteria the leader defines — faithfulness (did it stick to the source?), format, latency, cost, and a hallucination rate — usually run over a labelled set of real examples and/or an "LLM-as-judge" that scores outputs.
WhenBefore anything reaches users, and continuously after. Remember 73% of failed AI projects had no agreed definition of success — evals are how a programme sets and holds that bar.
Good looks likeA small, representative test set; clear pass/fail thresholds; faithfulness + hallucination tracked; an eval run gated into "can this ship?". The leader owns the criteria; the harness can be lightweight.
How it failsVibes-based shipping ("it demoed well, so we launched"), or evals so heavy they never get run. Start with 20–50 real cases and a clear bar.
In practice: promptfoo, Langfuse, LangSmith, or a spreadsheet + LLM-as-judge. The leader defines "what good means"; an engineer wires a production harness. Code involved: light to run lightweight evals (mostly config); more for a full harness.

4 · Guardrails & hallucination / IP-risk control

Direct it · design-led

The controls that keep an AI inside the lines — what it can say, do, see, and when a human must step in.

What it isA layered set: system-prompt constraints, allow/deny lists, output filtering (PII, off-policy), grounding + citations to cut hallucination, human-in-the-loop gates for consequential actions, and IP controls (don't leak source data; respect licences).
WhenAlways for anything customer- or money-facing; proportional to the risk tier — which the Operating Model already defines.
Good looks likeGuardrails matched to the workflow's risk, not a blanket lockdown; the model cites sources for factual claims; a clear "what it must never do" list; a tested escalation path.
How it failsEither none (a confident wrong answer reaches a customer) or so many the thing is useless. Calibrate to consequence + reversibility.
In practice: much of it is design + config a leader can specify directly (the deny list, the human gate, the "cite sources" rule). The code-level filters (PII detection, output validators; tools like Guardrails AI / NeMo Guardrails) are engineer work. Code involved: design = none; enforcement = some.

5 · RAG done right

Direct it · engineers build, the lead sets the bar

Grounding the model in internal content at answer time (covered conceptually in the Approach advisor — here's what "good" looks like in production).

What it isInternal documents are chunked, embedded and stored in a vector database; a question retrieves the closest passages, which are fed to the model so it answers from them — with citations.
WhenWhenever answers must reflect current internal facts — policies, manuals, tickets. The default for "answer from our knowledge."
Good looks likeRetrieval quality is measured (does it fetch the right passages?), not assumed; chunking + metadata are deliberate; access permissions are inherited from the source; answers cite. The hard part is the retrieval, not the model.
How it failsRAG inherits the data's quality and its access rules — messy/ungoverned content, or retrieval that surfaces things a user shouldn't see. Most disappointing RAG is a retrieval problem, not a model problem.
In practice: vector DBs (Pinecone, Weaviate, pgvector) + embeddings + an orchestration layer — or a configure-path inside a tool already in place. The leader sets the bar (retrieval faithfulness, access rules); engineers build. Code involved: real, on the build side — but increasingly a configure option.

6 · Fine-tuning — when it's actually worth it

Decide it · rarely done by hand

Training a model further on curated examples so it absorbs a style, format or narrow behaviour. Here is the leader's call.

What it isML training: curate a dataset, run training (often LoRA/SFT), evaluate, maintain. It teaches behaviour, not facts — for facts, RAG is the tool.
WhenRarely, and last. Only when a narrow, high-volume task still isn't consistent enough after prompting + RAG, and the volume justifies the cost.
Good looks likeA clear, narrow behaviour that prompting/RAG couldn't deliver; a clean labelled dataset with known provenance; an honest cost/benefit. Mostly, "good" is the discipline to say not yet.
How it failsFine-tuning to fix a knowledge gap (it won't); chasing it for prestige; or owning training-data provenance/IP/consent that wasn't thought through.
In practice: this is ML-engineer / data-scientist work — the call is the decision and the cost/benefit, almost always "try RAG/prompting first." Code involved: significant — but not yours.

7 · LLMOps — running AI in production

Direct it · requirements-led, engineers run it

Everything that keeps a deployed AI reliable, affordable and observable — sometimes called FMOps / LLMOps.

What it isDeployment, monitoring (quality drift, errors), cost + latency tracking, prompt/version management, rollback, and incident response — the discipline that turns a demo into a service.
WhenThe moment something is live and people depend on it — the gap between "it works in a demo" and "it works on a Tuesday at scale."
Good looks likeCost-per-use and latency are visible, alerts fire when quality drifts, every prompt/model change is versioned and reversible, and there's an incident playbook. The lead sets the SLAs and cost ceilings; engineers run the pipes.
How it failsSilent quality drift, a runaway token bill nobody watched, or no way to roll back a bad change. "88% of pilots never reach production" partly because no one owned the ops.
In practice: observability tools (Langfuse, LangSmith, Datadog LLM, Helicone) + CI/CD + cost dashboards. The lead's job: the requirements, and the dashboards that actually get read. Code involved: engineering-heavy — not yours.

8 · Standing up an AI Center of Excellence

Do it · no code (this is the leadership job)

The org design that turns scattered experiments into a production AI capability.

What it isDefine the operating model, the team and tracks — typically a Production track (workflow architects, prompt + skills engineers) and a Coordination track (training, use-case + ROI) — an ambassador network across the business, governance (use-case tiering, risk classification, escalation), and an adoption + ROI cadence.
WhenWhen AI moves from a few enthusiasts to the whole organisation — the point of standing one up.
Good looks likeA lean, high-impact team; reusable assets adopted without hand-holding; an ambassador in every part of the business; governance that's proportionate; and a business case the C-suite can see.
How it failsA CoE that becomes a bottleneck (everything routes through it), governance theater that slows everyone down, or no adoption muscle — the people-and-process 70% that kills most pilots.
In practice: pure leadership — structure, hiring, governance, cadence. The Operating Model app literally produces this artifact. Code involved: none.
The honest headline

Five of the eight a leader can genuinely do hands-on — prompt libraries, skills, eval and guardrail design, and standing up the CoE. The code-heavy three — RAG, LLMOps, fine-tuning — a leader directs and judges rather than hand-codes, and even those increasingly have configure/low-code paths (the "configure before build" point). The job is knowing when each is the right tool, and whether it is any good — not writing the pipeline.