Change Management & Training Plan

Technology is the smaller half of this program. ~109 people, most of whom have never used an LLM at work, must end M12 using the platform weekly because it makes their day easier — not because they were told to. This plan is executed as WBS Phase 10 (work-breakdown-structure.md) on the roadmap calendar; adoption is measured in the kpi-framework.md (§4, §8).

1. Communication strategy

Channels and cadence:

Channel	Content	Cadence	Voice
Teams — #ai-transformation	Program news, milestone posts, links to Mission Control	Weekly-ish, event-driven	TPL lead / champions
Teams — #ai-help	Support Q&A (see §10)	Continuous	Champions + IT
Monthly all-hands segment	10-min slot: one win, one number, one next-month preview	Monthly	CEO opens; champion or user tells the story
Mission Control news	Canonical program record — every milestone, honest status	≥1 post/month (WBS 7.5)	TPL lead
Department briefings	What’s coming to your department and when	Before each wave (M6, M9–M11)	Dept head + champion
Email	Used only for: kickoff announcement, wave onboarding logistics, policy publication	Rare, deliberate	CEO / HR

Rules: no channel proliferation; Mission Control is the single source of truth and Teams points to it. Bad news travels the same channels as good news — a slipped milestone gets a post, not silence.

2. CEO-led messaging

Why the CEO fronts it: in a ~109-person company, everyone reads whether the boss actually means it. A transformation announced by consultants is a project; one the CEO opens every all-hands with is the new way the company works. The CEO personally: opens the kickoff, delivers the monthly all-hands segment, visibly uses the assistant, and hands out champion recognition.

Key messages (repeated until people can say them back):

1. “AI takes the boring work, not the jobs.” The target is the repetitive 30% of everyone’s day — retyping, searching, re-formatting, chasing — so the human 70% gets more room.
2. No layoffs from AI. Recommended policy decision for the board (M1): commit publicly that no role is eliminated as a consequence of this program; capacity freed goes to growth, service quality, and reduced overtime. On a ~1.5% margin, the upside is doing more with the team we have — not shrinking it. Without this commitment, expect quiet sabotage; with it, expect curiosity.
3. Your data stays in the building. The models run on our own hardware in Piraeus; nothing confidential goes to any cloud AI.
4. You stay responsible. AI drafts, you decide. Nothing goes to a customer, a vessel, or an auditor unreviewed.
5. This is a 12-month build, not a switch. Pilots first, honest measurement, then rollout — nobody’s process changes overnight.

3. Department AI champions program

• Selection (M4, WBS 10.1): one champion per department (15 total), chosen jointly by dept head + TPL from discovery observations. Criteria: respected by peers, curious, hands-on with the department’s real work — not necessarily the most senior or most technical person.
• Role (~2–3 h/week): first trained, first users; run their department’s KB content sprints (WBS 6.5); deliver first-line help (§10); collect friction and feed it to the guild; co-run their department’s training lab.
• Weekly champions guild (from M4): 45 min, TPL-facilitated → champion-facilitated by M10. Standing agenda: wins to publicize, friction to fix, KB gaps, next week’s focus. Guild is also the program’s early-warning system for adoption problems.
• Recognition: named on Mission Control; champion spotlight in all-hands; CEO acknowledgment at M12 closure; certification badge (§4) awarded first. Recognition is public and cheap — use it constantly.

4. Training roadmap by audience

Audience	Content	Format	When (roadmap months)
All staff (~109)	AI basics, the Prime Products platform, prompting fundamentals, usage policy, verification duty	One 2-h hands-on lab (own real tasks, not toy demos) + self-serve KB path	With their rollout wave: pilots M5–M6, waves M9–M11
Department heads (15)	Use-case design: spotting automatable work, writing a use-case one-pager, reading their dept dashboard	Two 2-h workshops	M6 and M8
Power users (~15–20, nominated by champions)	Agents & automation: building/modifying simple automations under IT guardrails (WBS 9.6)	Workshop series, 3 × 2 h + project	M9–M10
IT (platform ops)	Serving stack, monitoring, model updates, incident runbooks, security ops	Working handover with TPL, runbook-driven (WBS 10.5)	M6–M7, certification M7
Management / executives	AI-assisted decision-making, dashboard literacy, sponsoring use cases, policy ownership	Two 2-h sessions (WBS 10.6)	M8
New hires	AI-first onboarding flow (§5)	Embedded in HR onboarding	From M9 onward

Formats: 2-h hands-on labs (primary — people learn on their own work), lunch-and-learns (voluntary, from M7, one per month, champion-hosted), self-serve learning paths in the Obsidian KB linked from Mission Control’s training hub (WBS 7.4). Every lab ends with each participant having completed one real task of theirs with the assistant.

Load cap: no employee spends more than 4 h/month on training (see §13).

5. Employee onboarding (AI-first)

From M9, HR onboarding includes: platform account on day 1; the 2-h basics lab within the first two weeks; the new hire’s department KB section as their primary “how we work here” reference; their champion as named buddy for AI questions. New-hire proficiency becomes a natural test of KB quality — if a newcomer can’t find it, the KB is missing it (feed gaps to WBS 6.7).

6. AI usage policies

Published as a one-page policy (KB + Mission Control) at platform GA (M6); acknowledged by every user at first login. Core rules:

1. Acceptable use: business purposes; no harassment, no generating content that misrepresents the company; personal experimentation on the platform is fine and encouraged — that’s how skills form.
2. Confidential data stays on-prem. Company, customer, supplier and employee data goes only into the Prime Products platform — never into public/cloud AI tools. (The on-prem platform exists precisely so this rule is easy to follow.)
3. Verification duty: the human using AI output owns it. Check facts, numbers, names, and regulation references before acting on them.
4. No unreviewed client deliverables: nothing AI-generated reaches a customer, supplier, authority, or auditor without human review — no exceptions, including “small” emails with prices or safety-relevant content (PPE specs, CBRN, chart data).
5. Sensitive-category caution: HR/personal data, legal matters, and safety-critical technical claims get elevated scrutiny; when unsure, ask IT/compliance before prompting.
6. Violations handled through normal management channels, proportionately — the goal is norms, not fear.

7. Prompting guidelines (summary)

Full prompt library and per-department patterns live in the Obsidian KB (template library); the short version taught in every lab:

• Give context: who you are, what the output is for, who will read it.
• Show an example of what good looks like (paste a past quote/email/report).
• Constrain the output: language (Greek/English), length, format, tone.
• Iterate — don’t settle for the first draft; ask for changes like you would from a junior colleague.
• Verify anything factual before it leaves your hands (§6.3).

8. AI safety principles

• Human accountability is never delegated to a model (§6.3–6.4).
• Agents get least-privilege access and human-approval steps on any action with external or financial effect (WBS 9.1; security-and-operations.md).
• Every agent has a named owner and review date in the agent registry — no orphan automations.
• Model limitations are taught explicitly (hallucination, Greek-language edge cases, stale knowledge) — trained skepticism is a feature, not a failure of confidence in the program.
• Anyone can flag a concerning AI output in #ai-help; flags are triaged weekly, and disabling a misbehaving agent requires no approval chain.

9. Feedback loops

Loop	Mechanism	Cadence	Feeds
In-tool thumbs	👍/👎 + optional comment on every assistant answer	Continuous	RAG/KB gap triage (WBS 6.7), model eval
Pulse survey	5 questions, 3 min: usefulness, friction, time saved, sentiment, one wish	Monthly from M6	CSAT & time-saved KPIs; steering pack
Champion retro	Guild session dedicated to “what’s not working”	Monthly	Platform backlog, training tweaks
Pilot/wave hypercare log	Issues during onboarding windows	Per wave	Rollout plan adjustments (WBS 11.6)

Every loop closes publicly: “you said → we did” section in the monthly Mission Control post.

10. Internal helpdesk model

Three-tier, deliberately informal at tier 1:

1. Tier 1 — department champion (in person / Teams): usage questions, prompting help, “is this a bug or me”. Resolves ~70% (assumption — to validate in pilots).
2. Tier 2 — Prime Products IT via #ai-help Teams channel: accounts, access, platform errors, integration issues.
3. Tier 3 — TPL escalation: model/stack defects, agent malfunctions, architecture questions. SLA per governance model.

Plus weekly office hours (from M6): 1 h open drop-in, TPL + IT + rotating champion — no ticket needed. #ai-help is also mined monthly for FAQ → KB articles.

11. Adoption KPIs

Adoption is measured, not assumed — full definitions and targets in kpi-framework.md §4 (WAU/MAU, % employees active weekly, queries/user/day, CSAT), §8 (people trained, certification, prompt-skill), and §9 (time saved). Headline commitments: ≥20% of employees active weekly at M6, ≥65% at M12; ≥100 of ~109 trained by M12.

12. Resistance management

Resistance is information. Expected personas and tactics:

Persona	What it looks like	What’s really going on	Tactics
Skeptical veteran (e.g. 20-year sales or ops hand)	“I’ve seen systems come and go; my way works.” Ignores the tool, quietly proud of it.	Expertise feels threatened; past IT projects burned them.	Don’t argue — recruit. Ask them to test the assistant on their hardest cases and report where it fails (they enjoy this, and the failure reports are genuinely useful). Show their expertise being encoded into the KB with attribution. Never lead with “it’s easy”.
Overloaded middle manager	”Great idea, no time.” Delegates training away, blocks champion time.	Rational triage — their month is already full; this looks like extra work with someone else’s payoff.	Bring one pre-built win for their bottleneck (e.g. the weekly report drafted for them) before asking for anything. Enforce the ≤4 h/month cap so the ask is credibly small. Their boss (CEO) explicitly makes program participation count as performance, not extracurricular.
Fearful clerk (accounting, admin, data-entry-heavy roles)	Quiet non-use; anxious questions asked privately, not in channels.	Reads “automation of repetitive work” as “automation of me.”	The no-layoffs commitment (§2.2), repeated by the CEO, specifically citing their kind of role. Pair with a champion for 1:1 first sessions — no group exposure. Early use cases that visibly assist them (drafting, checking) rather than replace them; celebrate the first fearful-turned-fluent user loudly.

General rules: resistance is handled by proximity (champion, dept head), not by escalation; the guild reviews adoption laggards monthly (KPI §4 heatmap) and picks tactics per person, not per memo; nobody is publicly shamed for low usage — but department heads own their department’s adoption number.

13. Minimal disruption approach

The CEO mandate: the company must keep serving vessels every day of the program.

Period	Max program load per employee	Nature
M1–M3	≤2 h total	Discovery interview(s) only
M4–M5	≤2 h/month (champions: ~3 h/week by choice)	Nominations, one-pager reviews
M6–M8 (pilots)	Pilot users ≤4 h/month; everyone else ~0	Training + pilot use within normal work
M9–M11 (rollout)	≤4 h/month in your wave month; ≤1 h/month otherwise	Lab + hypercare
M12	≤1 h	Surveys, KPI review sessions (heads only)

Principles: no big-bang cutovers — every AI-assisted process parallel-runs beside the existing method until its KPIs prove out, and the old way remains available during hypercare; rollout waves are sequenced so no department onboards during its seasonal peak (peaks identified in discovery — assumption, to validate); training happens on real work during work hours, never as homework; any department may invoke a 2-week onboarding deferral via steering if operations demand it — used deferrals are logged in the risk register (RSK-16) as an early-warning signal, not treated as failure.