Agentic readiness checklist

Ten questions before an agent touches the work

If you can’t answer one, that’s the work to do before the pilot, not after it.

The order everything follows
1 · GovernName owners, classify sensitivity, publish the guardrails.
→
2 · RememberOntology, glossary, skill catalogue, episodic write-back.
→
3 · AutomateAssistant first, drafting agents next, autopilots last.
Start at automate and you’ll pay to come back. Letters on each card (A–F) point to the reference panels at the end.
Frame the work
1

Which band is this work in: simple, complicated or complex? And did you decide, or did a vendor demo?

Each band needs a different method. Most programmes arrived with the band already set by a demo: simple work got agents it didn’t need, and complex work was forced into a fixed swimlane process it couldn’t survive. The band sets the method, the ladder rung and the memory mix.

Ready whenThe band is chosen deliberately, before any tool is picked.
SeeA · Work bandsB · Automation ladder
2

Which route: skills written top-down, or an ecosystem designed to discover them?

Route one defines skills up front. Route two lets agents work out the sequence and promotes what repeats. Readiness for route two is low and thinking is still process-centric, so most organisations will start on route one. What matters is whether route one leaves a path to route two.

Ready whenMemory is written so a planner agent can read it, from the first release.
SeeC · Two routes
3

Where is the work actually done today, and does the system of record know?

In most programmes we’ve seen, the system of record was a gateway where work was filed, not where it was done. The work happened in documents, email, chat and meetings, and was uploaded at the end, if at all. An agent bolted onto the record acts on stale history.

Ready whenThe agent works where the work happens and writes back to the record.
SeeE · Knowledge territories
Memory & knowledge
4

What does the agent remember after a case closes, and who can read it?

Episodic memory, the record of what happened on each case and why, was the single capability that separated tools that learned from tools that stalled. MIT calls this the learning gap.

Ready whenEvery closed case writes an episode, with defined read access. Without it, you have a chatbot with a longer context window, not an agent.
SeeD · Memory types
5

Who owns the ontology, glossary and skill catalogue, and when did they last change?

Together these are the agent’s shared vocabulary and toolkit: how the domain fits together, what the words mean, and which skills exist. The unglamorous governance nobody wanted to fund was what every AI product depended on. A knowledge assistant can’t exist until the glossary does.

Ready whenEach has a named owner and a recent change date. “IT, informally” means nobody.
SeeD · Memory types
Authority & control
6

For each activity, what is the authority rung, and which human decides below rung four?

The planner reads the permission table, not the process model: one row per activity, with its rung, guardrail and sensitivity class. Built properly, it did more for trust than any policy document and surfaced contradictions the policies had hidden.

Ready whenEvery activity has a row, and every activity on rungs 1–3 has a named human decider.
SeeB · Automation ladderF · Skill anatomy
7

Which guardrail fires when the agent acts, under what stable ID, and which episode records it?

Guardrails are red, yellow and green rules checked at the moment of acting, not filed in a binder. A guardrail without an ID can’t be versioned, and a decision without a recorded rule can’t be defended.

Ready whenGuardrails have stable IDs, and every firing is logged in the episode.
SeeF · Skill anatomy
8

Govern → remember → automate: what must be true before you automate, and who signs it off?

Automate in order: a knowledge assistant first, because now it has something to retrieve; drafting agents second, because now they have rules; delegates and autopilots last. Every plan that started at automate came back to the first two steps at higher cost, usually after the pilot had been declared a success.

Ready whenThe gate has a named owner and explicit criteria. No owner, no gate.
See↑ The order, top of page
Skills & people
9

Who owns each skill, and what evidence promotes a pattern into one?

On route one, a skill is written, versioned and owned before it runs. On route two, it’s earned: a pattern that works three times, with the episodes to prove it, gets promoted. Either way, the catalogue should grow toward the skills people actually use.

Ready whenEvery skill has an owner and evidence. Otherwise it’s a prompt somebody pasted into a wiki.
SeeC · Two routesF · Skill anatomy
10

Are your people already using AI unofficially, and does your plan assume they aren’t?

MIT found workers at 90%+ of companies use personal AI tools, but only 40% of companies have bought an official LLM subscription. We saw the same on every programme: the people are ready, the organisation isn’t.

Ready whenThe plan focuses on governance, memory and skills catching up with how people already work, not on driving adoption.

Reference

The frameworks behind the questions
A

Three work bands

Each band asks a different design question, sits on different rungs and leans on a different memory mix.

Supports Q1
Simple
What is the process?

Known sequence, stable inputs. Design the workflow, let the system run it, and govern the exceptions.

Ladder
Rungs 2–4
Memory
Procedural carries it; episodic is the audit trail.
Typical work
Annual assessment cycles, routing, approvals under thresholds
Complicated
What is the skill, and who may act?

Knowable with expertise. Define skills with guardrails and a rung; keep accountability human where the stakes say so.

Ladder
Rungs 2–3
Memory
Procedural skill catalogue; episodic precedent; semantic glossary.
Typical work
Case management, due diligence, regulatory review
Complex
What is the goal, the roles and the shared memory?

The process emerges. Define agent roles around a shared blackboard; the plan is composed at run time.

Ladder
Rungs 3–5, gated
Memory
Working memory on the blackboard; semantic ontology; episodic write-back.
Typical work
Cross-unit synergies, portfolio risk, multi-framework analysis
Bands follow the Cynefin framework (Snowden & Boone). Its fourth domain, chaotic, calls for acting first and making sense afterwards.
B

The automation ladder

Five rungs of human–machine balance. Match the rung to the stakes, the domain and the audit need. Most activities land on rungs 2–4.

Supports Q1 · Q6
Human-led 1–2Balanced 3Machine-led 4–5
1Fully human
Human 100% no process
Cynefin fit
Chaotic · Complex
2Manual with process
Human 100% defined steps
Cynefin fit
Simple · Complicated
3Human-led automation
Human ~75% machine drafts, human decides
Cynefin fit
Complicated · Complex
4Machine-led automation
Human ~25% policy + exceptions
Cynefin fit
Complicated · Complex
5Fully autonomous, gated
Human 0% in loop board sets the mandate
Cynefin fit
Complex · Chaotic
The gate that matters most sits at rung 3: the system recommends, a person decides, every time. Rungs 1–3 have a named human decider; from rung 4, the machine acts within policy and humans audit and handle exceptions.
C

Two routes to skills

Supports Q2 · Q9
Route one · top-down
Skills defined before they run
  1. 1Architecture and process engineering define the capabilities
  2. 2Blueprints and the permission table set the rules
  3. 3Skills are written, versioned and owned up front
  4. 4The planner composes within the defined set
Suits simple and complicated work. Where most organisations start.
Route two · ecosystem
Skills earned from evidence
  1. 1Design the roles, the shared blackboard and the ontology
  2. 2Let the planner work out the sequence, case by case
  3. 3Learn which plans work, through episodic memory
  4. 4Promote what repeats into skills (procedural memory)
Suits complex, high-value work.
They converge. Route one builds up episodes and starts promoting patterns; route two needs the guardrails route one writes. Same memory, different starting point.
D

Four memory types

What an agent needs to remember, and where each kind lives.

Supports Q4 · Q5
Working

the live case state: goal, facts, open questions, what each role has done

→
Episodic

what happened on this case: decisions, overrides, who changed what and why

→
Semantic

the ontology, glossary and knowledge graph: the vocabulary plans are written in

→
Procedural

the skills each role carries, versioned and promoted on evidence

How a case flows: agents write to working memory → the closed case becomes an episode → curated episodes feed semantic memory → repeated patterns are promoted into procedural skills.
E

Five knowledge territories

Where an organisation’s knowledge actually lives, and who holds it today.

Supports Q3
Working

Mail, calendars, chat, meetings, working files

Held byPeople, in collaboration tools. Rich and current, but invisible to any system of record.

Structured

Transactions, master records, workflow state

Held byIT-run applications. Hard to ask a historic question without a report.

Unstructured

Policies, procedures, reports, minutes

Held byDocument libraries and drives. Findable if you already know where to look.

Public

News, filings, statistics, regulation

Held byNobody. It changes faster than any organisation can curate.

Human

Expertise, judgement, the why behind every override

Held byIndividuals only. The one territory with no system; it leaves with every resignation.

An agent that ignores a territory leaves a person doing middleware work. Episodic memory is the only way the human territory ever gets written down.
F

Anatomy of a skill

An open standard defines the skill itself; governance adds the rules that run at the moment of acting.

Supports Q6 · Q7 · Q9
Standard core · Agent Skills open specification
1
Name and descriptionwhat it does and when to use it, which is how the agent picks it
2
Instructionsthe steps, examples and edge cases, loaded when the skill is used
3
Resources and toolsscripts, reference docs, templates and pre-approved tools, loaded only when needed
Governance layer · added by DesignChain, carried in the skill’s metadata
4
Guardrailsred, yellow and green rules with stable IDs, checked before and after acting
5
Authority runghow far it may go before a human decides
6
Memory in and outthe episodes and semantics it retrieves, and what it writes back
One row of a permission table (illustrative)
ActivityRungGuardrailsSensitivityMemory inMemory outDecides
Prepare the intake summary2 of 5GR-03, DP-12RestrictedCase facts, prior episodes on similar mattersDraft summary, episode entryCase officer
Skills and the permission table do different jobs. The skill says how to do the work and which rules to check; the table says who may act at each step.