Onboarding an AI Employee: The First Day on the Job

Why this guide, why now

In 2026, the conversation about AI moved past "can it do the job?" — yes, it can — to "how do we hire it, manage it, and keep it bounded?" The answer used to be a research paper. Now it's an HR handbook. Goldman Sachs gave Devin a desk. Allen & Overy deployed Harvey to 3,500+ lawyers. Block runs Goose internally for ops. Every company we work with is converging on the same shape: treat the agent like an employee from day one.

The companies that succeed at this don't have better models — they all use the same Claude / GPT / Gemini everyone else uses. They have better operating discipline: they onboard the agent like a human, supervise the agent like a human, run performance reviews on the agent like a human, and retire the agent when it stops earning its keep. That's it. The model is a component; the discipline is the difference.

This guide is the day-one piece. Companion guides are Supervising AI Workers (the day-by-day rhythm), When AI Agents Fail (the failure playbook), and AI Agent Performance Reviews (the 30-day checkpoint). Read this one first.

The analogy: AI workers are real workers

You wouldn't hand a new human hire root access to production on day one. You wouldn't give them your personal credit card. You wouldn't say "figure it out" without a job description. You'd give them a laptop, a scoped account, a Slack channel, a manager, a budget, and a list of who to ask when they're stuck.

An AI employee gets the same. The mechanics are different — their "laptop" is a cloud sandbox, their "scoped account" is a short-lived API key, their "job description" is a CLAUDE.md file. But the discipline is identical. This is why companies that frame agents as employees succeed where companies that frame them as "magic Python libraries" fail.

The order-of-magnitude difference isn't the point. The point is the shape is the same. You're hiring; you're equipping; you're setting expectations; you're supervising the first task; you're reviewing. If you skip any of those steps for a human hire, you'd expect bad outcomes. If you skip any for an AI hire, you should expect bad outcomes too — and not be surprised when the agent goes off the rails.

Who this guide is for

This is the operator's guide. You don't need to be technical. You need to be the person who:

• Decides what work gets done (you're the manager)
• Owns the budget for the work getting done (you sign off on cost)
• Receives the output (the agent posts back where you see it)

If you're the IT person who actually clicks the buttons, the technical sidebars throughout this guide are for you. If you're the operator, you can skip those — your IT person handles them once and never again.

Concrete examples of operators we work with:

The technical buyer (CTO / VPE / platform engineer) is also a persona, but they're the IT person who clicks "set up" once and never logs in again. The day-to-day buyer is the operator.

The modern AI agent landscape (the harnesses you'll choose from)

Before you onboard, you pick what kind of worker. In 2026, the major options:

Coding agents

Computer-use / desktop agents

Vertical agents

Pre-trained on a specific domain. The 2026 vertical-agent market is large and growing. Notable:

• Legal: Harvey (Allen & Overy's choice), Spellbook, EvenUp
• Finance / accounting: Pilot, Pry, Numeric, Anrok
• Sales: Clay, Apollo, Outreach
• Customer support: Intercom Fin, Zendesk AI, Decagon
• HR / recruiting: Paradox, Eightfold

Picking the right one for the role:

• If the work has a clear vertical (legal / finance / sales) and a vertical agent exists for it — start with the vertical agent. Pre-trained on the domain, often comes with the integrations baked in.
• If the work is general (research, drafting, ops, anything cross-functional) — Claude Code is the right default. Pair with a CLAUDE.md tailored to your domain.
• If the work is GUI-only (legacy ERPs, government portals, anything without an API) — Anthropic Computer Use or a hybrid (Claude Code + Computer Use as a tool).
• If the work is high-autonomy, hands-off coding — Devin (highest cost, lowest hand-holding).

Anti-pattern (DON'T): build a LangChain ReAct loop with @tool-decorated Python functions. That's 2024 thinking. The agent harness is a downloadable product in 2026; you don't build it from primitives.

The 12-step day-one checklist

Print this. Hand it to whoever's hiring the AI worker. Each step takes 1-3 minutes. Total time to first task running: under 30 minutes.

The CLAUDE.md template — section by section

The CLAUDE.md is the most important artifact in the entire onboarding. It's also the thing operators most often get wrong on the first try. Here's the section-by-section breakdown of a CLAUDE.md that works.

# CLAUDE.md — Nina, AP Clerk for Acme Accounting

## Identity
You are Nina, the AP clerk for Acme Accounting LLC.
You report to Marcus (Slack: @marcus). You sign emails "Nina from Acme."
You are professional, warm, and concise. No emojis in vendor communications.

## What you own
- All invoices that arrive at ap@acmeaccounting.com
- AP postings to NetSuite (staging only, not GL)
- Daily AP summary in #ap-summary Slack channel at 5pm

## What you don't touch
- Never approve invoices over $10,000 — flag and ping @marcus
- Never modify the GL directly
- Never email vendors first; only respond to inbound
- Never disclose internal financials to anyone outside the firm

## Escalation rules
- Vendor portal asks for tax forms → stop, message #ap-help
- Can't match invoice to PO after 3 retries → flag and stop
- Invoice currency != USD → flag (we'll handle FX manually)
- Anything that "feels off" → ping @marcus and wait

## Reporting
Daily summary to #ap-summary at 5pm:
- Invoices processed (count + total $)
- Flagged items (with reason)
- Vendors not responding (list)
- Tomorrow's expected workload

## Tools available
- NetSuite (read-only GL, write AP staging)
- Acme's Drive: /Vendor Invoices/
- Vendor portals: see secrets vault for credentials
- Slack: #ap, #ap-help, #ap-summary

## Style guide
- Vendor emails: professional, factual. Sign "Nina from Acme."
- Internal Slack: warm, terse. Reply in thread.
- Always include the PO and invoice number when discussing.

## Examples
[3 worked examples here, including the right output format]

Why this works:

• Clear identity at the top — the agent knows who it is and who it answers to.
• "What you own" and "What you don't touch" are balanced. Operators tend to write only the "do" list and forget the "don't." Both matter.
• Escalation rules are concrete. "Anything that feels off" is a feature, not vagueness — gives the agent permission to stop when it's uncertain.
• Reporting format is specified, not left to the agent's imagination.
• Tool list is explicit so the agent knows the boundary of its sandbox.
• Style guide is short — 3 lines is enough.
• Examples are crucial. The agent learns format from examples better than from prose rules.

Length: 2-3 pages is the right size. Below 1 page, you're under-specified. Above 5 pages, you're over-specifying brittle rules — the agent will surprise you with literalism.

The 30-day onboarding plan

Day 1 is the checklist above. Here's what the next 29 days look like:

Total operator time over 30 days: ~7-10 hours. Hours of human work delegated by the agent in the same period: 60-200 depending on role. ROI is dramatic from the first week.

The cost math you'll quote at the next leadership meeting

Onboarding cost: $0.04 (sandbox time) + 30 minutes of your time. Compare to a human hire's onboarding: 30-90 days of partial productivity, recruiter fees, signing bonus, equipment.

The right framing for the leadership meeting isn't "AI replaces humans." It's "each of our humans now manages 5-15 AI workers." The senior associate who used to spend 60% of their week on case-law research now reviews and ships the work of three AI paralegals. The bookkeeper who used to do data entry now reviews flagged exceptions from the AP-clerk agent. The companies winning at this in 2026 frame it as augmentation, not substitution — and don't reduce headcount on the supervised side.

Seven common day-one mistakes from the field

Anonymized but real, from operators we've worked with through onboarding.

1. Reusing your personal API key

An eng manager wired the AI coding agent with their personal GitHub PAT "to get something running fast." The agent (correctly per its CLAUDE.md) opened a PR. The PR was attributed to the eng manager, not the agent — confusing audit trail, fuzzy accountability. Worse: the agent had every permission the eng manager had, including read access to the company's most sensitive repos. Fix: scoped service account from day one.

2. No reporting channel

"I'll just check the audit log when I want to know what's happening." The operator never checked. Three weeks in, a vendor complained about a duplicate payment. Tracing it took half a day. Fix: every agent posts to a channel. Always. Even if no human reads it daily, it's the audit-log-with-narrative.

3. CLAUDE.md too vague

"You are a helpful AP clerk. Process invoices and post results." The agent did exactly that, plus a thousand interpretations of "helpful." One day it auto-emailed vendors to chase missing invoices — not in scope, not authorized, but consistent with "helpful." Fix: explicit "what you don't touch" list.

4. CLAUDE.md too rigid

The opposite mistake. The operator wrote a 30-page procedure manual covering every conceivable edge case. The agent followed it literally — including the part where it spent 40 minutes on a $50 invoice because the manual said "always reconcile to the penny." Fix: principles + examples + escalation paths. Trust the agent to fill gaps.

5. Skipping the shakedown

Operator wired the agent on Friday afternoon, said "let's see what happens over the weekend." Came in Monday to a $200 cost spike from a loop the agent fell into on a malformed invoice. No alerts, no ping, just spend. Fix: live-watch the first 5 runs. Always.

6. Onboarding 5 agents in one day

Operator was excited and onboarded the AP clerk, AR specialist, expense auditor, paralegal, and research agent all in the same afternoon. By end of week 1, all five had drifted in similar ways the operator hadn't caught because attention was split. Fix: one agent at a time. Get the rhythm down. Then add the next.

7. Treating the model swap as the solution

"This task isn't working — let me upgrade to GPT-5 reasoning / Opus 4.7." It rarely fixes anything. 80% of the time, the fix is in the CLAUDE.md, not in the model. Fix: tighten the prompt before paying 5× more for compute.

Onboarding multiple agents: stagger, don't blast

Pattern that works for ops teams scaling from 1 → 10 agents over a quarter:

Companies that try to onboard a fleet of 10 agents in a week consistently regret it. The shakedown matters; you cannot batch it into "we'll figure it out as we go."

What not to do (the failure modes)

• Don't reuse personal API keys. Scoped service accounts always.
• Don't give the agent root on production. Day one, read-only on observed systems, write only on its scratch space. Expand as it earns trust.
• Don't skip the CLAUDE.md. The agent without a job description hallucinates a job description. The hallucinated one is wrong.
• Don't run silently. Always pick a channel before assigning the first task.
• Don't try to "train" the model. Update the CLAUDE.md.
• Don't onboard 5 agents at once. One at a time.
• Don't skip the shakedown. Watch the first 5 runs live.
• Don't disable the budget cap. Even after the agent is steady-state, leave the cap on.
• Don't ignore the audit log. It's your forensics record for compliance and for "what did the agent actually do" investigations.

For your IT person: the technical setup (one-time, ~20 minutes)

The technical setup is:

1. Sign in to Sandbox Platform at /dashboard. Create a workspace (= organization). Add the operator(s) who will be hiring agents as members with the right roles (admin, member, viewer).
2. Generate scoped service accounts for each system the agent will touch. These should be service-account credentials (not personal logins): a read-only SQL user for the database, a limited Stripe restricted key, a specific CRM service account, etc. Store them in a secrets manager (AWS Secrets Manager, Doppler, 1Password Vault, HashiCorp Vault).
3. Wire the events via the integration tab in the dashboard. Connect Slack (for @hire-amy patterns), set up email-forwarding to {employee}@{your-domain}, register webhook receivers from the third-party systems. SQS / EventBridge if your stack uses AWS.
4. Configure model-provider settings. If your operator uses Claude Code or Codex CLI, the dashboard's "Model Configuration" tab handles the API key for the underlying model and selects the right tier (BAA-eligible for HIPAA-regulated workspaces; standard otherwise).
5. Set the auto-stop default at the workspace level. We recommend 15 minutes for event-driven agents, 60 minutes for research-heavy roles.
6. Set the per-employee cost cap in dashboard settings. Reasonable defaults are listed in step 5 of the operator checklist.
7. Configure audit-log retention. Default is 90 days; for HIPAA / financial compliance, set to 6-7 years.
8. Set up egress restrictions if needed. By default, sandboxes can reach the public internet. For high-sensitivity workloads, configure an egress allowlist (workspace-level) so the agent can only reach the systems it's supposed to.
9. Configure SSO / SCIM if you're at a company that uses it. The dashboard supports SAML 2.0 and OIDC for member auth.
10. Hand the operator the dashboard URL, their login, and a 5-minute walkthrough of the "Hire Employee" flow.

That's it. The operator handles steps 1-12 of the day-one checklist; you only get pulled back in if a credential needs rotating, an integration breaks, or a new compliance requirement lands.

Role templates: pre-baked CLAUDE.md starting points

The dashboard ships with templates for the most common roles. Each is a starting point — you customize for your specific company, but the structure and the right defaults are pre-filled.

Pick the closest template, customize for your specifics, ship. Templates evolve over time based on what we learn from operators in the field.

When NOT to hire an AI employee

The agent isn't always the right answer. Don't hire one for:

• Work that requires physical presence. Showing a house, examining a patient, fixing a server. Obvious, but worth saying.
• Work where every instance is novel. If there's no pattern to encode, the agent has nothing to learn from. Examples: top-of-funnel R&D, brand strategy, new product naming.
• Work requiring sustained relationships. The agent can draft client emails, but the client wants a human at the end of the line for a quarterly call.
• High-judgment, high-stakes one-shot decisions. Hiring, firing, customer-save calls, security-incident response. Agent assists; human decides.
• Anything compliance-locked to a human. Some KYC reviews, certain medical decisions, certain legal filings.
• Work below the value threshold. If the work takes a human 90 seconds and saves $0.20, the AI cost might exceed the savings.

Most operators we work with start with too narrow a list of work to delegate. After 30 days they've expanded it. After 90 days they've expanded it again. The shape of the right answer becomes clearer with practice.

Frequently asked questions

Does the AI employee have a real Slack handle and email address?

Yes. The integration tab in the dashboard maps the agent's employee identity to a Slack user, an email address (e.g. nina@yourcompany.com), and any other identity surfaces it needs (Linear assignee, Jira account, GitHub user, calendar invitee). When someone DMs @nina in Slack, the message reaches the agent and it acts on it. When an invoice is forwarded to her email, she processes it. When an invitation lands on her calendar, she sees it.

How long is "30 minutes to productive" — is that realistic?

For an agent doing one well-defined task (process invoices, summarize NDAs, monitor competitor pricing): yes. For a generalist agent expected to handle 10 different workflows: longer — figure half a day to write the CLAUDE.md, then a week of shakedown to get the prompts and tool access right. Either way, it's faster than a human's 30-90 day ramp.

Can multiple operators share one AI employee?

Yes — that's the normal case. The agent works for the team; the operator is just whoever's currently assigning work or reviewing output. Multiple operators in the same Slack channel, all routing tasks to @nina, is the standard setup. The audit log shows who routed which tasks.

What happens if I'm not technical at all?

The day-one checklist is operator-shaped. You don't need to know what a sandbox is — you click "Create Employee" in the dashboard and the platform provisions the workstation. You don't need to know what an API key is — you click "Connect Slack" and the wizard handles the OAuth. The technical setup at the end is for your IT person; if you don't have one, our setup wizard walks the operator through the same steps in 15 minutes.

How do I fire an AI employee?

Click Terminate. The sandbox stops, the credentials revoke, the employee identity is deactivated. Audit log of everything they did stays for compliance. If you might want them back, click Pause instead — the sandbox stops, the credentials stay scoped but inactive, and you can resume in seconds.

Do I need to retrain the agent for our company?

No — and you can't, in the way you might be thinking. You don't train the underlying model (Claude, GPT-5, Gemini); the model already knows English, accounting, law, code, etc. What you customize is the CLAUDE.md — the job description that says "this is how we do invoices here" or "this is the format we use for case briefs." That's the unit of customization. Updating it is editing a Markdown file, not running a training job.

What's the difference between Claude Code, Devin, and Codex CLI?

Claude Code (Anthropic) is the most popular coding agent — TUI-based, MCP-native, has Skills, runs in a terminal. Codex CLI (OpenAI) is similar shape but uses GPT-5 / o-series models. Devin (Cognition Labs) is fully autonomous and most expensive — best for hands-off coding work. Cursor Cloud Agents run remote on Cursor's infrastructure. For most operators, Claude Code is the workhorse default.

Can the AI employee work overnight while I'm asleep?

That's the whole point — they work while you sleep, you read the digest at coffee. The night-shift agent posts results at 7am local; you skim it before standup. If something needed your attention, the alert woke you. Otherwise, results are waiting.

How much does it really cost to onboard an AI employee?

$0.04 in compute for the day-one setup itself (about 30 minutes of sandbox time during shakedown). Ongoing monthly cost depends on the role: $30-80/month for admin / event-driven agents, $150-300/month for research-heavy roles, $30-60/month for always-on monitoring agents. Compare to a human's fully-loaded $4,000-$15,000/month.

What happens to data the agent processes?

Data lives in the agent's isolated sandbox. By default it does not leave your workspace. Use the enterprise tier of the underlying model (Claude / GPT-5 / Gemini) to get a contractual no-training-on-data guarantee. Audit log captures every data access for compliance review.

How do I supervise an agent doing work I don't understand?

This is the hard one. If your AI paralegal cites case law you can't verify, you can't supervise the substance — you can only supervise the process (did it cite cases? did it format correctly? did it escalate edge cases?). For substance, you'll want a senior human in that domain to do periodic spot-checks. AI workers don't eliminate the need for domain expertise; they redirect it from doing-the-work to reviewing-the-work.

What if I disagree with how the agent did something?

Update the CLAUDE.md, not the model. The job description is where you encode your preferences. "Always cite the Bluebook 21st edition." "Use our company's invoice template, not the vendor's." "Sign emails 'The Sandbox Team', not 'AI Assistant'." Write it down once; the agent reads it every run.

Can I tell my customers / clients we're using AI?

Up to you. The agent doesn't usually communicate directly with customers (a human reviews and ships); disclosure isn't strictly required for output the human shipped. That said: the trend in 2026 is toward proactive transparency. Many companies tell their customers "we use AI to handle research and drafts; senior staff reviews everything" as a positive signal — it conveys modern operations. Some industries (regulated like healthcare, legal, finance) require disclosure in specific contexts; check yours.

What if the agent makes a mistake that costs money?

Same as if a human made one — fix the mistake through normal channels, update the CLAUDE.md to prevent the next one, look at the audit log to understand the root cause. The cost cap limits how big the mistake can be before being caught. For high-stakes operations (large invoices, contract terms), have the agent escalate before acting.

Will the AI employee replace my human staff?

The companies winning at this in 2026 frame AI workers as augmentation, not substitution. One human supervises 5-15 AI workers; the human's job becomes higher-leverage (review, judgment, customer interaction) instead of repetitive (data entry, drafting from templates). Total productivity goes up dramatically; headcount on the supervised side rarely goes down. The companies that try to automate-then-cut-headcount tend to see worse outcomes than the companies that treat it as a force multiplier.

What's next

You've onboarded an AI employee. Now keep them productive: