learn-ai-project-management-with-phoebe / Leader session 2 of 6
Learn AI + Project Management with Phoebe · Leader track · Session 2 of 6

The governed setup: what goes in, what never does, and who could notice

Session a1 drew the line between labour and judgment. This one draws the harder line: the boundary around the workspace itself. What documents your PMs may load, what must never cross, what a connector can actually reach, and why letting it write back to a ticket is a different decision from letting it read. You leave with a one-page policy and a permission review you can run on Thursday.

🟢 Leader track Heads of delivery · PMO leads Policy + permissions 45 min
0-3 · Where we are 3-22 · The boundary 22-42 · Policy + permission review 42-45 · Q&A
Part 0

Where we are

In a1 you split your function's week and named three artifacts worth moving. You also, if the homework went the way it usually does, discovered that some of your PMs had already started - on whatever account was handy, with whatever was on screen. That is not a discipline problem. It is what happens when there is pressure from above, no written boundary in the middle, and a keen person at the bottom. Tonight you write the boundary. It has four parts: where the work lives, what may go in, what may never go in, and what the tools are allowed to reach on their own. Northwind runs underneath again - six milestones, M3 slipping from a 10 August gate to a 24 August forecast, a vendor called Veridian going quiet, and a sponsor who will ask why nobody flagged it sooner.

Live - presented in session Self-study - read after class ★ Try it now prompt Official sources covered
★ What you walk out with today A drafted one-page AI-in-delivery policy with a green list, a red list, a named signer and a logging rule; a permission review you can run on any connector in ten minutes; and a clear-eyed view of blast radius - what a tool can see, what it can change, and who would notice if it did.
Part 1 · covers PMI's governance principle

Anatomy of a governed setup 7 min live

A governed setup is five things, and only one of them is a tool. One workspace per programme so context does not leak between clients. A context pack that makes real answers possible. Connectors, scoped and deliberate. A written policy short enough to be read. And a name at the bottom of each artifact type. Miss any one and you do not have governance, you have a licence.

INSIDE THE BOUNDARY NEVER CROSSES One workspace per programme, not per person The context pack charter, plan, RAID, owners, glossary Connectors read-only, scoped, logged Policy + a name green list, red list, who signs it The red list personal data - staff, customers, candidates client-confidential terms, anything under NDA unreleased financials and forecasts HR or performance material about named people Your privacy regime sets the edges. Legal decides it, the PMO does not. If a PM has to guess which side something sits on, the boundary is not written down.
🔍 Click to zoom - the four parts inside the boundary, and the material that never crosses it
LiveThe five parts, and what breaks without each4 min
PartWhat it isWhat breaks without it
One workspace per programmeA persistent, named home for Northwind and nothing elseContext bleeds between programmes and clients. The fastest way to put one client's material in front of another.
The context packCharter, milestone plan with gate criteria, RAID, owner list, glossary, last three or four status reportsGeneric project-management advice. Nobody in your function needs a definition of critical path.
ConnectorsScoped links to the delivery tools so the pack partly refreshes itselfA pack that decays quietly and produces confident answers about last month's programme.
A written policyOne page: green list, red list, who signs, what gets loggedEvery PM invents their own boundary, and you find out which one during an audit.
Named accountabilityOne human name per artifact type, with authority to hold it backEleven approvals in nine minutes and nobody who can say who read it.

PMI's 2026 standard puts this squarely inside your existing governance rather than beside it - it asks for "clear structures, roles, authority, and guardrails", which is a description of a RACI and a policy, not a new committee. If your organisation already has a data classification scheme, an information security policy and a delegation of authority, you are extending three documents you already have. That is a far easier paper to get approved than a new AI framework, and it is also the correct answer.

LiveWhy one workspace per programme, not one per person2 min

Per-person workspaces feel natural and are the wrong unit. The programme is what has a charter, a risk register, a sponsor and a boundary. The person moves teams.

  • Handover survives, and the boundary is inspectable. When Marcus L. goes on leave or a PM rotates, the context stays with the programme instead of walking out in somebody's chat history. And you can ask "what is in the Northwind workspace" and get an answer, which you cannot meaningfully do across nine personal accounts.
  • Client separation is structural, not behavioural. Nobody has to remember to keep two clients apart if the workspaces were never shared. And refresh becomes a habit with an owner: five minutes on a Monday replacing the plan, the RAID and the latest report. Assign it, or it will not happen.
Real world

The pack that was three weeks old. A Northwind PM asked which milestone was most at risk and got a clean, confident answer naming M2 and the landing zone. Entirely reasonable, and entirely wrong - M2 had signed off, and the live problem was M3, where the DQ pass rate had stuck at 96.4% against a 99% gate and ticket NWD-412 was still failing. Nothing had misbehaved. The pack had just not been refreshed since before the ingest work started. Stale context does not produce obvious nonsense, it produces last month's answer in this month's confident voice, which is much harder to catch.

Six documents, not thirty Your PMs build the pack in session b1. Your interest as a leader is that they build a small one: a bloated pack is not safer, it is noisier - more places for stale material to hide, more surface to review, more chance something from the red list slips in unnoticed. Charter, milestone plan with gate criteria, RAID, owner list, glossary, last three or four reports. If a PM cannot say why a document is in there, take it out.
Part 2 · the red list and who owns it

What never goes in 6 min live

The green list is easy and your PMs will guess it right. The red list is where people get it wrong, usually not out of carelessness but because the material arrived attached to something legitimate: a scope discussion with the vendor's commercial terms in the appendix, a resourcing note with a performance comment in it, a board pack with next quarter's numbers on slide nine.

LiveThe five categories3 min
  • Personal data. Staff, customers, candidates, contractors. Names in an owner list are usually fine and are the point of a RACI - a spreadsheet of customer records is not, and neither is a CV.
  • Client-confidential terms. Commercial terms, pricing, anything your client would be surprised to learn had left their document.
  • Anything under NDA. Including the vendor material that arrives in a delivery thread. Priya N. forwards Veridian's SLA schedule to explain why VN-2291 is still open; the explanation is useful, the schedule is not yours to load.
  • Unreleased financials. Forecasts, board numbers, anything with a market-sensitive date on it. The two contract data engineers at roughly £38k for four weeks is a delivery number and fine. Next quarter's unpublished revenue is not.
  • HR and performance material about named people. This one catches good PMs out. "Marcus is the single point of failure on the CDC job" is a legitimate delivery risk and belongs in the RAID. "Marcus has been struggling since his review" is not a delivery artifact and does not go anywhere near a workspace.
The test that works under pressure Would you be comfortable if this document were read aloud, in full, at a meeting with the client, your regulator and the person named in it all present? It is blunt, it takes two seconds, and it catches the appendix problem better than any classification scheme.
LivePrivacy regimes, and the review that is not yours3 min

Wherever you operate, you have a privacy regime - GDPR, PDPA, a sector rule, or several at once if you run cross-border programmes. This course stays deliberately generic about which, because the delivery answer is the same in all of them: personal data has a lawful basis, a purpose and a retention story, and "we pasted it into a tool to draft a status report" is not one of them.

Two things worth saying plainly to your PMO, because unclear ownership here produces either paralysis or recklessness and both are expensive:

  • The legal and procurement review of vendor terms is not a PMO job. Data processing terms, training-data commitments, sub-processors, retention, jurisdiction, indemnities - these belong to legal and procurement. Your job is to raise it as a gate, in writing, with a date, and then stop. A delivery lead reading a vendor's terms and forming a view is a risk to the organisation and to their own credibility.
  • Your job is the delivery boundary. Which artifacts, which documents, which people, which tools, who signs. That is entirely yours, and nobody else is going to write it for you.
Real world

The appendix nobody read. A Northwind PM loaded a scope-change pack to draft an impact note for D-14, the sponsor decision to descope real-time POS to nightly batch until M5. Sensible artifact, sensible use. Page eleven of the pack was Veridian's rate card, under NDA. Nothing bad happened, and that is the point - nobody would ever have known, which is exactly why the control has to be a habit at the point of loading rather than a discovery later. The fix was one line in the policy: documents go in whole or not at all, and if you have not read the whole thing, you have not checked it.

Handling grey cases There will be grey cases weekly at first, then rarely. Handle them the same way every time: the PM writes the case down in one line, you answer in one line, and the answer goes into the next version of the policy. Within a quarter your one-pager has eight or nine hard-won lines on it and the questions mostly stop. A policy that grows from real cases is trusted; one that arrives fully formed from a template is worked around.
Part 3 · connectors, permissions, blast radius

Connectors and blast radius 7 min live

A connector links the workspace to your delivery tools - the ticket system, the docs, the sheets - so the context pack partly refreshes itself. It is genuinely useful and it is where permissions stop being theoretical. The rule that matters is short: a connector inherits the permissions of whoever authorised it. If your most senior PMO admin sets one up, it can see what that admin can see, which is usually a great deal more than the programme.

Read access sees the ticket, the sheet, the thread - and nothing more. Blast radius: whatever its authoriser could already see. Start here. Still scope it. Write access can change a ticket, close a task, comment - in your name. A separate decision from read, taken separately, signed off. Default off. Never anything that sends externally, approves spend, or reaches HR and personal data. No pilot, no exception, no "just this once". Not a connector question. An audit trail must show who authorised it, what it could reach, what it changed, and when. Connector features move fast. Re-check what yours can reach and change this quarter. A connector inherits the permissions of whoever authorised it. Choose that person.
🔍 Click to zoom - read, write and never, plus what your audit trail has to be able to show
LiveRead is not the same decision as write3 min

Most organisations approve a connector once and never revisit the read-versus-write question, which is how a tool that was signed off to summarise tickets ends up able to close them.

  • Read has a visibility problem: write has an integrity problem. The read worst case is that something confidential gets summarised into a place it should not be - bad, recoverable, and mitigated by scoping the connector to the programme's projects rather than the whole instance. The write worst case is a changed record with your name on it that nobody remembers making. Worse, harder to unpick, and it corrupts the artifact your delivery data is drawn from. A ticket closed in error does not just lose a task, it makes next week's report wrong in a way nobody can trace.
  • Approve them separately, in writing. Different worst case, different decision, different signature - if read and write arrive as one tick-box, split them before you sign. And scope beats trust: "only this programme's projects" is a better control than "only careful people use it", because scope survives a bad week and trust does not.
LiveWhat an audit trail actually has to show2 min

Four questions. If you cannot answer all four for any connector in your function, you have a finding, and better it is yours than an auditor's.

  • Who authorised it, when, and what it can reach. A person, not a team inbox. Then the actual scope - which projects, which spaces, which files - not the tool's marketing description of it.
  • What it can change, if anything, and who would notice. Read-only or write, and if write, on what. Then: if it changed something tomorrow, would a human see it, and how soon? That second half is the question people forget, and it is the one that turns a permission into a real control.

Keep this vendor-neutral in your policy. Name the capability, not the product, because the product will rename the feature next quarter. Connector behaviour, defaults and permission granularity move fast enough that anything you write down about a specific tool should carry a re-check date.

Real world

The connector nobody remembered approving. During a routine review a PMO found a ticket-system connector still live from a pilot four months earlier. It had been authorised by a portfolio admin, so it could read every project in the instance, not just the two in the pilot. Nothing had gone wrong. Nobody could say who owned it, when it was last used, or who would have noticed if it had been. The review took eleven minutes and produced three actions: re-authorise under the programme service account, scope it to Northwind's projects, and put a named owner and a review date against it. Eleven minutes, and it was the most valuable thing that PMO did that week.

Where the standards sit The Model Context Protocol is the open spec behind most connector work: it defines how a tool exposes what it can do and what it can reach. Worth knowing it is a published standard rather than a vendor feature, because you can ask a supplier a precise question about scope and auth posture instead of accepting a diagram. What it does not do is decide your permissions. The spec is the plumbing; the authority model is yours, and it should look like the one you already run for any other integration into your delivery tooling. The requester being enthusiastic does not make it a different category.
Demo 1 of 2

Write the one-page AI-in-delivery policy ★ 12 min · everyone drafts

This is the a1 one-pager hardened: green list, red list, who signs, what gets logged. One page. If it runs to three, nobody reads it and you are back to nine private boundaries.

Green list: the documents your PMs may load, named specifically. Charter, plan, RAID, owner list, glossary, past reports. Anything else needs a reason, and the reason gets written down.

Red list: the five categories from Part 2, in your own words, with one example each from your actual programmes. Examples are what make a red list usable at nine on a Monday. Then name the signer per artifact type, and the person who owns the workspace itself - including the weekly refresh.

Write the logging rule (which artifacts were AI-drafted, who reviewed, who signed - reviewed monthly at the PMO, not annually by an auditor) and the escalation line: what a PM does when something is unclear, and how fast they get an answer. If the answer takes two weeks, they will decide for themselves.

Send it to legal and procurement with the vendor-terms question attached, and say plainly that the terms review is theirs. Do not wait for them to finish before you publish your delivery boundary.

★ The policy skeleton - one page, your name at the bottomAI IN DELIVERY - POLICY v1.0, [date], review [date + 3 months] Owner: [your name, head of delivery]. Escalation: me, same day. WHERE THE WORK LIVES One workspace per programme, owned by the programme PM, never a personal account. Refreshed every Monday: plan, RAID, latest status report. GREEN LIST - may be loaded Charter or mandate · milestone plan with gate criteria · RAID register · owner list glossary · last 3-4 status reports · decision log. Anything else: ask, in one line. RED LIST - never loaded, no exceptions - Personal data about staff, customers, candidates or contractors - Client-confidential commercial terms - Anything under NDA, including vendor material forwarded into a delivery thread - Unreleased financials, forecasts or board numbers - HR or performance material about a named person Documents go in whole or not at all. If you have not read all of it, you have not checked it. CONNECTORS Read: scoped to this programme's projects, authorised by a named person. Write: separate approval, separate signature, default off. Never: sending externally, approving spend, anything touching HR or personal data. WHO SIGNS Status pack: the programme PM, weekly. Escalations and sponsor material: me. Anything a client, regulator or auditor reads: me, after I have read it. WHAT WE LOG Which artifacts were AI-drafted, who reviewed, who signed, which connectors are live and who authorised them. Reviewed at the monthly PMO. NOT OURS Legal and procurement own the review of vendor terms, data processing and retention. Raised by me on [date]. We do not form our own view on those terms.
The version that leaves every PM guessingTeams may use approved AI tools for project documentation, provided no sensitive or confidential information is shared and appropriate human review is applied. Connectors to internal systems should be configured in line with existing security policy. Queries to the AI working group.
Why the second one fails Read it as a PM with a status pack due at four. Is the RAID register sensitive? Is Veridian's SLA schedule confidential, or is it just a vendor document? Does "appropriate human review" mean read it or approve it? Is a read-only ticket connector "in line with existing security policy" or does someone have to say so? Every clause is unfalsifiable, so the PM decides alone and you inherit whatever they chose. A policy that cannot be broken cannot be followed either.
Demo 2 of 2

Run a permission review on one connector ★ 10 min · pick a real one

Pick one connector that is live in your organisation right now - any tool, any team, it does not have to be a delivery one. Four questions, ten minutes. Most people find something on the first attempt, which is the honest reason this exercise is in the course.

Who authorised it, and what can it see? Find the human name and the date - if the answer is a team inbox, an unowned service account or "it came with the pilot", that is finding one. Then establish what it can reach, not what it is used for. Whose permissions did it inherit? If an admin set it up, assume instance-wide until proven otherwise.

What can it change, and who would notice? Read-only or write, and if write, was that ever approved as a separate decision or did it arrive bundled with read? Then: if it changed a record tomorrow, who sees it and how soon? "The next audit" is the same as nobody.

Write the three actions that come out of it. Usually: re-authorise under the right account, narrow the scope, name an owner with a review date - and put that date in a calendar, not a document.

QuestionA healthy answerA finding
Who authorised it, and what can it see?A named person, dated, still in role, scoped to this programme's projects and spaces"It came with the pilot", an unowned service account, or everything the admin who set it up could see
What can it change, and who would notice?Nothing - read-only, approved as read-only; and if it did change something, a named human sees it within a dayWrite, bundled into the same approval as read, and nobody watching - or "it would show up in the audit log" with no one reading it
Take the finding upward, not sideways If the review turns up something uncomfortable, report it yourself, the same week, with the three actions already attached. A head of delivery who finds their own connector problem and fixes it is credible. One whose problem is found by an auditor spends the next year explaining it, and every AI conversation in that PMO gets harder.
Homework

Before session a3 ◐ 30-45 min total

Source material

Official sources covered

Taught from PMI's public standards, the open connector spec, and vendor documentation for the tooling. Legal and procurement review of AI vendor terms is deliberately out of scope and stays with legal and procurement. This session covers:

PMI - Standard for AI in Portfolio, Program and Project Management (2026), governance principleParts 1-3 · clear structures, roles, authority and guardrails inside existing governance
Model Context Protocol - connector documentation and specPart 3 · what a connector is, what it can reach, auth posture. Re-check current behaviour.
Anthropic docs - workspaces, projects and connectorsPart 1 · where the work lives. Depth in PM sessions b1 and b9.
Check yourself

Three questions before you go 🎯 ◐ 90 seconds

1 · A PMO admin sets up a read-only connector to the ticket system so drafts can cite live tickets. What is the blast radius?

A connector inherits the permissions of whoever authorised it. Intent is not scope. Choose the authorising identity deliberately, scope it to the programme's projects, and treat read and write as separate approvals.

2 · A PM asks whether they may load the vendor's SLA schedule to help draft an impact note. What is the answer?

NDA material stays out even when it arrives attached to something legitimate, and documents go in whole or not at all. Redacting a name does not change the obligation. Raise the terms question with legal in writing, with a date, and stop there.

3 · Your connector review answers three of the four questions but nobody can say who would notice if the tool changed a record. What does that mean?

A log nobody reads is not detection. "Who would notice, and how soon" is the question people skip and the one that decides whether write access is genuinely governed. If the honest answer is the next audit, the answer is nobody.

Leader session 2 cheat sheet · pin this

Five partsOne workspace per programme, a context pack, scoped connectors, a written policy, a named signer. Four of the five are not tools.
Per programme, not per personHandover survives, the boundary is inspectable, client separation is structural.
The red listPersonal data, client-confidential terms, NDA material, unreleased financials, HR and performance material.
Whole or not at allThe appendix is where it goes wrong. If you have not read all of it, you have not checked it.
Not your reviewVendor terms, data processing and retention belong to legal and procurement. Raise it in writing, with a date, then stop.
Inherited permissionsA connector sees what its authoriser sees. Scope it to the programme, choose the identity deliberately.
Read is not writeRead has a visibility problem, write has an integrity problem. Separate decision, separate signature, write off by default.
Four audit questionsWho authorised it, what can it see, what can it change, who would notice. Re-check quarterly - features move fast.