How to Build an AI Team: Roles, Org Chart, Approval Workflow

How to Build an AI Team: Roles, Org Chart, Approval Workflow

TL;DR: An AI team mirrors a human org chart in structure but not in cost or overhead. You define roles, assign functions, set approval boundaries, and the AI workforce executes within those guardrails. This guide walks through the five-tier AI org chart, the nine standard IC roles, the five C-level roles launching September 2026, the approval workflow architecture, cost budgets per agent, and the complete onboarding flow for adding new agents to your team.

Building an AI team is not a technology decision. It is an organizational design decision. The questions that matter are the same ones you would ask when building a human team: What functions need to be covered? Who has authority to make which decisions? How does work get assigned and tracked? What does escalation look like? When should the team escalate to the founder?

The AI layer answers these questions with much lower overhead – no recruiting cycles, no ramp time, no benefits administration – but the underlying organizational logic is identical. Companies that treat AI workforce deployment as purely a technology decision, and skip the organizational design thinking, consistently underperform compared to companies that approach it as a team-building exercise with a different hiring mechanism.

This guide gives you both: the specific roles available today, and the organizational thinking framework for deploying them well.

The 5-Tier AI Org Chart

ewpire formalizes a five-tier structure that maps AI agents to organizational layers. Understanding this structure helps you deploy agents at the right level and avoid the common mistake of using IC-level agents for coordination functions they are not designed to perform.

In the current ewpire product, you interact primarily with tier 3 – the specialist IC agents. You are the CEO and C-level, providing goals and strategic direction. The tier 4 custom layer is available in the Business plan today via custom agent configuration.

This five-tier structure is not arbitrary. It mirrors the span-of-control principles from organizational design research. Each IC agent (tier 3) has a clearly defined domain, measurable output, and bounded authority. Each C-level (tier 2) has visibility across its domain and coordinates between IC agents. The CEO (tier 1) has visibility across all domains and translates founder goals into domain-level assignments. This hierarchy prevents coordination failures – the most common reason AI workforce deployments underperform in practice.

The 9 Standard IC Roles: Detailed Breakdown

ewpire's nine specialist agents cover the functions that consume the majority of operational capacity in a growing B2B company. Here is what each one actually does, what tools it uses, and when it pays for itself:

1. Sales Agent

Function: outbound prospecting, cold email sequencing (up to 50 personalized emails per day), follow-up management across a 5-touch sequence, reply classification and warm-lead escalation. Uses four email frameworks – PAS (Problem, Agitate, Solution), BAB (Before, After, Bridge), QVC (Question, Value, Call-to-action), and AIDA (Attention, Interest, Desire, Action) – selecting the optimal framework per prospect segment.

Trigger-event aware: detects funding announcements, new executive hires, and company news to time outreach optimally. Plain text emails only, anti-AI-detection techniques applied, per-lead local timezone windowing (sends only 9 AM–noon in the recipient's timezone).

Pays for itself when: one qualified meeting booked per month converts at your average win rate. At $199/month and a $10,000 average contract value, you need one win per 50 months of outreach – or one win from every 150 emails sent, which is below industry average reply-to-win ratios for targeted outbound.

View the full agent spec at ewpire.com/agents/sales-agent.

2. Support Agent

Function: first-response customer support, FAQ handling, billing inquiry triage, onboarding guidance, escalation management with full context transfer. Confidence scoring ensures it only answers within its knowledge boundary – low-confidence responses flag to human support rather than fabricating answers. Multilingual (communication language detected from customer, responses matched). RAG-powered knowledge base search over your uploaded documentation.

Deployable via single script tag on your website (the ewpire.com support widget visible on this page is the Support Agent in live deployment) and via Telegram, Slack, and other connected messengers for internal team support workflows.

Pays for itself when: 60% of your current support volume is first-response FAQ and status inquiries that the agent handles without escalation. At a $35/hour loaded support cost and 15 minutes average resolution time, 200 tickets/month handled autonomously = $1,750/month in labor savings from a $199/month subscription.

View the full agent spec at ewpire.com/agents/support-agent.

3. Research Agent

Function: competitive intelligence, market landscape analysis, company profiling, sentiment analysis, weak signal detection. Produces 9-field structured briefs covering: company overview, competitive positioning, recent news and trigger events, identified product gaps, pricing signals, growth indicators, risk factors, key personnel and decision-makers, and recommended strategic action.

Sources: web search, news APIs, public company databases, semantic search over indexed content. Produces a complete brief in 4–7 minutes that would require 3–4 hours of human analyst time. Output is a structured document, not a wall of unorganized text.

Pays for itself when: you currently spend any meaningful analyst or founder time on competitive research that feeds into sales, product, or strategic decisions. One substantive competitive analysis per week at $60/hour analyst rate = $240/month in labor savings from a $199/month subscription.

4. Lead Qualification Agent

Function: BANT + MEDDIC hybrid scoring applied to inbound leads and outbound prospects. BANT covers Budget, Authority, Need, Timeline. MEDDIC adds Metrics, Economic Buyer, Decision process, Decision criteria, Identify pain, Champion identification. Smart disqualification logic flags leads that pass surface criteria but fail on structural fit.

Buying signal detection identifies behavioral patterns (content consumption, repeat visits, specific page combinations) that correlate with purchase intent. Outputs: scored prospect record with reasoning, recommended next action, and confidence level. Warm leads flagged to your messenger immediately; cold leads scheduled for nurture sequences.

Pays for itself when: your team is spending more than 2 hours per week reviewing leads that turn out to be unqualified. Two hours saved per week at a $75/hour founder or sales manager cost = $600/month in reclaimed time from a $199/month subscription.

5. Growth Agent

Function: five distinct growth modules covering the full marketing production function.

• SEO Article Generator: Long-form content (1,500–3,500 words) optimized for both traditional search and AI engine visibility (GEO – Generative Engine Optimization). Structured with the entity density, citation patterns, and question-answer formatting that AI search engines index preferentially.
• Landing Page Copy: Three hero section variants per brief with A/B test hypothesis embedded. Headline, subheadline, three value propositions, primary CTA.
• Email Campaigns: Lifecycle sequences, newsletters, and drip campaigns for opted-in lists. CAN-SPAM and GDPR compliant. Anti-AI-detection techniques applied.
• Ad Creative Briefs: Full creative specifications for Google, Meta, LinkedIn, and YouTube. Includes copy angles, image direction, audience targeting rationale, and bid strategy notes.
• Brand Monitoring: Tracks your brand mentions across public sources and flags opportunities for engagement or reputation response.

View the full spec at ewpire.com/agents/growth-agent.

6. HR Screener

Function: resume review, STAR (Situation, Task, Action, Result) scorecard scoring, skills categorization, bias-free processing per EU AI Act requirements for employment AI. Produces structured candidate summaries with percentile rankings on defined skill dimensions, comparative analysis across the candidate pool, and a recommended shortlist with reasoning.

Reduces hiring manager review time by 60–80% for initial screen stages by eliminating the unstructured "read every resume" step and replacing it with a structured comparison of scored summaries. The human hiring manager reviews 5 structured candidate assessments instead of 50 raw resumes.

Compliance documentation for EU AI Act requirements for AI in employment contexts is at ewpire.com/legal/ai-act.

View the full spec at ewpire.com/agents/hr-screener.

7. Tender Agent

Function: 18-field bid/no-bid analysis covering all relevant dimensions of a procurement tender, MEAT (Most Economically Advantageous Tender) scoring per EU procurement rules, consortium strategy analysis (when teaming with partners strengthens the bid), and public tender discovery via SAM.gov (US federal) and UK Find a Tender Service.

For companies participating in government or institutional procurement, this agent pays for itself on a single accurate bid/no-bid decision. The cost of pursuing an unwinnable bid – staff time, external consulting, printing and submission – frequently exceeds $10,000. The cost of missing a winnable bid is the contract value itself. The Tender Agent reduces both failure modes.

View the full spec at ewpire.com/agents/tender-agent.

8. Document Processing Agent

Function: handles seven document types (contracts, RFPs, invoices, compliance forms, NDAs, vendor applications, financial statements) through a six-step pipeline: ingest, classify, extract key fields, detect PII, flag anomalies, produce structured output. The output is a structured data object with extracted fields, flagged concerns, and a summary – not a restatement of the original document.

PII detection flags personal data per GDPR processing obligations. Anomaly detection flags unusual clauses in contracts, missing required fields in forms, and numerical inconsistencies in financial documents.

Pays for itself when: you are spending any time on routine document review that follows consistent extraction logic. Contract review at $200/hour legal rate for one hour per contract, 20 contracts per month = $4,000/month in legal time partially replaceable by a $199/month subscription.

View the full spec at ewpire.com/agents/document-processing.

9. Executive Concierge

Function: 13 intent types including email triage and draft responses, meeting scheduling and coordination, travel research and booking briefs, research requests (non-competitive), reminder and deadline management, lifestyle search, and administrative coordination tasks. The CEO-adjacent function that handles everything that should not reach domain specialists – and everything that should not require the founder's attention but currently does.

The Concierge is the agent most often added as a second subscription alongside a primary function agent. The Sales Agent handles pipeline; the Concierge handles everything else. Founders consistently report the Concierge recovering 5–10 hours per week of time previously consumed by administrative coordination that required their personal attention to handle.

View the full spec at ewpire.com/agents/executive-concierge.

The 5 C-Level Coordination Roles

Five C-level orchestration layers sit between the founder and the IC agents. These layers manage the IC agents within their domain and translate domain-level goals into specific IC agent tasks:

These coordination roles are available across Pro and Business plans. See pricing →

Custom Role Creation

The Business plan supports configuration of custom agents beyond the nine standard IC roles. Custom agents inherit the complete base framework – memory architecture, tool access layer, escalation patterns, compliance guardrails – and are specialized via onboarding configuration specific to the custom function.

The threshold for a viable custom role: the function must be repetitive enough that consistent rules can be defined, volume-heavy enough that AI execution saves meaningful time, and bounded enough that four to eight onboarding questions can capture the relevant constraints and objectives.

Examples of custom roles deployed in ewpire's beta program: investor relations coordinator (manages LP update schedules, quarterly report drafts, data room requests), supplier compliance screener (evaluates new supplier documentation against defined compliance criteria), partner onboarding concierge (guides new channel partners through onboarding milestones), grant application researcher (identifies relevant funding opportunities from public grant databases and produces eligibility assessments).

The Custom plan covers fully bespoke agent development with ewpire's team. View the Custom plan.

The Approval Workflow Architecture

Approval workflow is the operational architecture that determines which decisions agents make autonomously and which require human confirmation. Getting this wrong in either direction is costly: too much autonomy creates uncontrolled risk; too much human-in-the-loop creates a bottleneck that eliminates the efficiency value of AI agents entirely.

The ewpire framework uses three approval tiers, configured per-agent during onboarding and adjustable at any time:

Tier A: Fully Autonomous

The agent acts without asking. Examples: sending a follow-up email within an approved sequence, responding to a support inquiry within documented FAQ scope, classifying a document that matches a known template, adding a prospect to a nurture sequence, searching for leads that match an approved ICP definition.

Tier A decisions share two properties: they are reversible within the rollback window (15 minutes for operational actions), and they operate within explicitly defined guardrails that bound the decision space. The agent cannot exceed the guardrails without escalating.

Tier B: Notify and Proceed

The agent acts and notifies you simultaneously. You can override within a defined window (default 15–30 minutes) if needed. Examples: sending the first email to a new prospect batch, publishing a research brief to a shared channel, flagging a lead as qualified for the sales pipeline, scheduling a follow-up call suggestion.

Tier B decisions are lower-stakes but time-sensitive enough that waiting for approval before acting would degrade effectiveness. The simultaneous notification ensures you maintain visibility without being a bottleneck.

Tier C: Escalate and Wait

The agent presents its analysis and recommended action, and waits for your confirmation before proceeding. Examples: a support refund request above your set threshold (default $100), a tender bid decision in the "uncertain" zone of the bid/no-bid model, a candidate in the "borderline" qualification range, any action that involves external commitment or spending.

Tier C decisions are high-consequence, low-frequency, and justify human decision time. The agent's escalation includes its full reasoning, the relevant context, and the options available. One tap to approve, reject, or redirect. The agent logs your decision and applies it as a pattern to future similar situations – so over time, well-calibrated Tier C decisions generate precedent that can move common sub-cases to Tier B.

The starting configuration recommendation: begin with more Tier C than you think you need. In the first two weeks, actively use escalations to observe the agent's judgment in your specific context. As you build confidence that the agent's reasoning in a given category is sound, promote those sub-cases to Tier B or Tier A. The autonomy should be earned through demonstrated judgment, not assumed at deployment.

Issue Assignment and Delegation

ewpire supports a formal issue system. Each issue is a structured task with an owner (agent or human), due date, success criteria, current status, and linked output artifacts. Agents become assignees that can receive issues, execute against them, produce linked outputs, and report completion with artifact references.

The delegation workflow: you describe a business goal. You then decompose the goal into issues and assign each to the appropriate agent. The Research Agent gets the landscape brief issue. The Sales Agent gets the prospecting campaign issue. The Lead Qualification Agent gets the response scoring issue. Each agent works independently; you monitor aggregate progress, surface blockers, and intervene only when human judgment is needed.

Issue assignment happens directly via messenger commands. You describe a task to the agent, the agent executes and reports results, and you provide feedback or redirect as needed.

Cost Budgets Per Agent

Cost comparisons use fully loaded human hire costs: salary, benefits, recruiting, equipment, training, management overhead. AI cost includes all inference, tool use, and infrastructure. No per-email, per-document, or per-query charges – volume is included within plan limits.

Tools Each Agent Uses: The Full Stack

Agents are not text generators. They use external tools to take actions in the real world. Here is the full tool stack per agent:

• Sales Agent: multi-source web search (Searlo, Exa, SerpAPI, Brave Search), SMTP for email sending via configured domain, calendar API for follow-up scheduling, Google Sheets for prospect tracking output
• Support Agent: RAG search over your documentation (pgvector semantic retrieval), billing API for subscription queries, email for ticket management, Telegram and other messenger APIs for channel delivery
• Research Agent: web search, news APIs, public company databases (Crunchbase public data, LinkedIn public profiles), semantic search over indexed sources, structured output formatter
• Lead Qualification Agent: CRM data access (read-only), web enrichment for firmographic data, scoring model execution, Google Sheets output for pipeline visibility
• Growth Agent: SerpAPI for keyword and SERP research, URL monitoring for brand mention detection, structured content formatter, image brief generator
• HR Screener: document parser (PDF, DOCX), STAR rubric engine, structured comparison table generator, anonymized candidate scoring output
• Tender Agent: SAM.gov API (US federal procurement), UK Find a Tender Service API, document parser, MEAT scoring model, consortium analysis engine
• Document Processing: multi-format document parser (PDF, DOCX, XLSX, images), PII detection model, classification model, structured field extractor, anomaly detection
• Executive Concierge: Google Calendar API, email triage (IMAP read), web search for lifestyle and research queries, reminder scheduler

Onboarding a New Agent: Step by Step

Adding an agent to your team follows a consistent five-step process that takes under 15 minutes total:

1. Select the agent – Browse the agent marketplace at ewpire.com/agents. Each listing describes the function, tools used, typical outputs, and plan requirements. Select the agent that matches your primary current need.
2. Connect your channel – Link your communication channel (Telegram, Slack, Discord, or email) via a single configuration command. The agent will communicate with you through this channel for all escalations, reports, and task outputs.
3. Complete the onboarding interview – Answer four to eight structured questions about your business context. For the Sales Agent: company name and description, target audience definition, product or service value proposition, communication preferences, and any behavioral rules (industries to avoid, countries to exclude, maximum email frequency). For the Support Agent: company context, documentation upload, escalation threshold, tone preferences, and language settings.
4. Configure approval tiers – Define what the agent can do autonomously (Tier A), what it should notify-and-proceed on (Tier B), and what requires your explicit confirmation (Tier C). Start narrower than you think you need and expand as confidence builds.
5. Review first outputs – The agent begins executing immediately. Review the first five outputs in your messenger and refine configuration if needed via /settings. Most configuration issues surface in the first 24 hours and are resolved with a single settings update.

Total setup time: under 15 minutes. First meaningful output: within the first hour. The onboarding walkthrough for v2 is documented at Getting Started with ewpire v2.

Common Build Mistakes and How to Avoid Them

Three patterns consistently produce underperforming AI team deployments:

Starting too broad: Deploying all nine agents simultaneously before establishing a baseline. Without a comparison point, it is impossible to know which agents are performing and which need configuration work. Start with one or two agents, establish quality expectations, then expand.

Under-configuring Tier C: Setting too much to Tier A at launch because it feels efficient. When the agent makes a decision you would have made differently, and you did not see it in advance, you lose the opportunity to correct and calibrate. Start with more Tier C escalations than feel necessary, observe the pattern of decisions, then promote categories to Tier A or B once you have verified the agent's judgment in that category.

Skipping the first-week calibration: Not reviewing the agent's first five to ten outputs carefully. The configuration onboarding gives the agent direction; the first outputs reveal whether that direction is translating correctly. Most configuration issues are obvious in the first week and trivially fixable. Ignoring them for a month means running on miscalibration for a month.

Frequently Asked Questions

How many agents do I need to start?

One. Start with the function where you have the most volume and the clearest need. For most B2B companies, that is Sales or Support. Add agents as you build confidence in the model. The Starter plan is one agent; Pro is three; Business is nine. Add-on agents are $199/month each.

Can different agents work together on the same task?

Yes. You can coordinate agents by taking output from one (Research Agent competitive brief) and feeding it as input to another (Sales Agent ICP refinement). The /transfer command passes a full conversation summary from one agent to another. The @mention syntax lets you query a specific agent one-off without switching your active session.

Do I need to train the agents?

No training in the machine learning sense. You configure them via onboarding questions. The agents come pre-trained on their domain functions; you personalize them to your business context. As they operate, they learn from outcomes automatically.

What if I need a function that is not one of the nine?

The Business plan supports custom agent configuration. Contact support via the FAQ page or the Axi widget to discuss your specific function. ewpire also offers a Custom plan for bespoke agent development for functions outside the nine standard roles.

How do approval workflows change as the team scales?

Start narrow and expand based on observed judgment quality. In the first two weeks, approve more decisions manually and use those approvals to calibrate confidence in the agent's reasoning. By week four, most teams have expanded agent autonomy significantly. As you scale to multiple agents, the C-level coordination layer handles cross-agent approvals so you only see the decisions that genuinely require founder attention.

Build Your AI Team

Nine specialist IC agents available today. Start with the one function that costs you the most time or money per month, validate the model, then expand.

Browse all 9 AI agents →

See pricing and start your trial →