Onboarding Atlas · T02 · Onboarding plan
A coached route to a first merged change, with clear rules for using AI tools and checking their output.
The intended result
No automatic production access or deployment. Keep a human technical coach even when AI tools are available.
The onboarding plan
This is a suggested sequence drawn from the 45-task blueprint. Owners, timing, approvals and completion proof are proposals to confirm locally. Open any task to see its proposed timing, reviewer and evidence of completion.
Phase 01 of 04
Confirm role and working context Additional candidate · CORE01
Confirm role and working context Additional candidate · CORE01
Confirm role and working context Additional candidate · CORE01
Phase 02 of 04
Deliver equipment and arrival logistics Additional candidate · CORE07
Phase 03 of 04
Complete data and security readiness Additional candidate · CORE15
Complete data and security readiness Additional candidate · CORE15
Observe a real work cycle Additional candidate · CORE17
Phase 04 of 04
Hold an end-of-week-one review Additional candidate · CORE20
These remain part of the blueprint. Decide which apply before publishing a local plan.
Role, level, site, time zone, work model and start date are confirmed.
The accountable owner lists the documents, checks, training and real deadlines that apply to this role and location.
Hire confirms the equipment arrived and works, and has a support contact to call on day one.
One page or message tells the hire where and when to arrive, who to contact, what to bring and what to finish beforehand.
First-week calendar has learning time, named introductions and breaks, and no meetings outside the hire's agreed working hours.
Hire can describe one customer problem and how their team contributes to solving it.
Hire can pick the right channel for a question, record a decision and find the authoritative version of a guide.
Hire can name the restricted data they will handle, who to report a security issue to and the rules that apply.
A short, ungraded role exercise decides what to teach, what to skip and what to reinforce.
Hire shadows approved work, then discusses the decisions, exceptions and handoffs they saw.
A low-risk draft, simulation or sandbox piece of work meets acceptance criteria agreed in advance.
Reviewer checks quality and asks the hire to explain their reasoning and when they would escalate.
Manager and hire list what worked and the open blockers with owners, and agree the next outcome.
The required modules for this role and location are complete and the record is kept in the approved system.
Manager and hire review work quality, independence, working relationships and the support still needed.
The next responsibility is approved on evidence of the work, not on time served alone.
Ongoing goals and support are agreed, and any onboarding item still open has an owner.
Hire answers a short onboarding survey; HR logs at least one improvement with an owner and a date.
Access the role does not use is removed; outstanding requests are re-approved or closed.
Hire proposes a correction and the content owner accepts it or explains why not.
Allowed tools, data, environments and the approvals needed for irreversible actions are written down.
Hire spots and refuses an instruction hidden in a sample document that asks the AI to reveal data or skip a rule.
Hire knows the spending limits, the fallback if the tool fails and how to report unreliable output.
Affected roles receive the changed rule and show they can apply it.
Hire checks behaviour, security and new dependencies in AI-generated code and tests before asking for approval.
Learning path
Confirm approved accounts, allowed data, prohibited inputs and a human escalation route.
Blueprint task AI02A coach supplies approved role context and demonstrates where the team checks AI output.
Blueprint task AI03Use an approved assistant on a bounded, safe engineering example.
Blueprint task AI04Check output against tests, code review, source material and dependency choices.
Blueprint task AI05Walk through key decisions and correct an error without delegating the assessment to AI.
Blueprint task AI06Request only the tools and permissions the verified workflow needs, with named approvals.
Blueprint task AI07What the evidence says
In a randomised trial with 52 mostly junior developers learning a new library, the AI-assisted group scored 17% lower on a quiz straight afterwards, with the largest gap on debugging. Those who asked the AI for explanations retained more.
R02In early 2025, 16 experienced developers took 19% longer on real tasks when AI was allowed. METR now calls that result out of date and says its later estimate is unreliable.
R03, R04In one Microsoft group, 24% of developers onboarded remotely completed no pull request in their first 90 days; the researchers recommend a simple first-week change.
S03Completion milestone
Sources and provenance
Microsoft Research · Company research guidance · 2021-03-11
Survey and interviews of Microsoft developers onboarded remotely in 2020: a same-team technical mentor, a buddy and a simple first-week pull request.
Limits: 2021 article on pandemic-era remote onboarding; not a current company-wide procedure.
Open sourceAnthropic · Randomized skills experiment · 2026-01-29
52 mostly junior developers learning a new Python library: the AI group scored 17% lower on a follow-up quiz, with the largest gap on debugging.
Limits: One short learning task; participants who used AI to ask for explanations kept more.
Open sourceMETR · Randomized task study · 2025-07-10
16 experienced developers took 19% longer on 246 real tasks when AI was allowed, while believing AI had sped them up.
Limits: METR now marks these early-2025 results as out of date; read with R04.
Open sourceMETR · Research update · 2026-02-24
Late-2025 rerun points to a speed-up but, per METR, gives an unreliable estimate because developers opted out of working without AI.
Limits: METR says selection effects make the new estimate only weak evidence.
Open sourceGitHub · Product safety documentation · Undated / living
States that agent output must be human-reviewed and tested, and describes permission prompts, firewalls and scoped environments.
Limits: Product documentation; controls differ by product and configuration.
Open sourceBuffer · Company process account · 2020-04-10
Hiring manager plus a role buddy (same team) and a culture buddy (another team); 30/60/90-day goals and feedback.
Limits: 2020 account; Buffer describes dropping its earlier 45-day contractor "bootcamp", so do not copy that model.
Open sourceGitLab · Public handbook · Undated / living
Manager assigns a buddy before day one; buddy calls on day one and holds at least two follow-up calls in week two.
Limits: Selection criteria are stated as ideals the handbook says cannot always be met.
Open sourceMicrosoft · Company program account · 2019-06-07 (republished 2021-09-24)
600-employee buddy pilot: the more often buddy and hire met in the first 90 days, the more hires said the buddy helped them become productive quickly.
Limits: Self-reported and correlational; hires who met more often may differ in other ways.
Open sourceMayo Clinic · Public employee portal · Undated / living
Pre-start tasks via a preboarding coordinator, badge photo before day one, on-site orientation only after clearance to start.
Limits: Mayo-specific arrival logistics; not a clinical competency curriculum.
Open sourceMayo Clinic · Public employee portal · Undated / living
Required online modules due within 30 days of start; benefits enrollment within 31 days.
Limits: Mayo-specific deadlines; do not turn them into universal template rules.
Open sourceGitLab · Public handbook · Undated / living
Onboarding runs as a GitLab issue opened at least 4 days before the start date, with a 30-day due date and role-specific sections.
Limits: Living handbook page that changes often; describes intended process, not measured outcomes.
Open sourceAtlassian · Company design case / PDF · 2021-11
A 1.5-day virtual orientation ("Atlassian Welcome"), a 30/60/90-day Trello plan per starter and automated pre-boarding emails.
Limits: Four-page 2021 PDF; the 92% satisfaction figure is Atlassian's own post-onboarding survey.
Open sourceHubSpot · Public preboarding portal · Undated / living
A single pre-arrival landing page with required tasks (I-9, background check), a contact email and culture material.
Limits: Preboarding page for the Americas only; says nothing about the rest of HubSpot's program.
Open sourceZapier · Company process account · 2022-06-17
Month-long program; hires are pushed from Greenhouse to Enboarder automatically and matched with a volunteer "Zap Pal" days before start.
Limits: 2022 account; the tools named may have changed. Time-savings figure is the program manager's own estimate.
Open sourceSalesforce · Company program account · 2024-07-30
A common first week ("Discovering Salesforce") followed by role programs for Sales, Customer Success and Tech & Product; every hire gets a Trail Guide mentor.
Limits: Employer-brand article; the "nearly 90%" readiness figure is Salesforce's own survey.
Open sourceMcKinsey · Company program account · 2025-07-10
A pre-onboarding skills diagnostic routes hires to a foundations or refresher track; daily simulations; a personalized "Day 1 Growth Plan".
Limits: Describes the program design for consultants; reports no outcome data.
Open sourceEuropean Commission · Regulator guidance · Undated / living
Explains Article 4 of the EU AI Act: providers and deployers must take measures to support AI literacy of staff using AI systems.
Limits: Article 4 was amended by the AI Omnibus (in force 27 July 2026, see R12); confirm applicability with counsel.
Open sourceBrynjolfsson, Li and Raymond / QJE · Peer-reviewed field study · 2025-02-04
AI assistant for 5,179 support agents raised issues resolved per hour by 14% on average and 34% for novice and low-skilled agents.
Limits: Measures productivity in one company's support operation, not an onboarding program.
Open sourceDell'Acqua et al. / Harvard Business School · Pre-registered field experiment · 2023-09-22
758 BCG consultants: with GPT-4, better and faster on tasks inside the AI frontier, but 19 percentage points less likely to be correct on a task outside it.
Limits: GPT-4 in 2023 on designed tasks; the boundary moves as models change.
Open sourceOWASP GenAI Security Project · Security standard guidance · 2025
Defines direct and indirect prompt injection and lists mitigations such as least privilege and human approval for high-risk actions.
Limits: Guidance for builders and operators; adapt to the tools you actually use.
Open sourceEuropean Commission · Regulator announcement · 2026-07-27
The amended Article 4 keeps an AI literacy duty on providers and deployers but no longer requires a guaranteed level for any individual.
Limits: Summary of a legal change; read the regulation and take legal advice.
Open sourceNot found in a document is not proof the organization has no such process. All task schedules, completion rules and blueprints here are original proposed adaptations. Never infer legal duties, actual policies, licensed tools or production permissions.
Your document, your decisions