setupaiagents.com

Give your team back 5–15 hours a week. One ChatGPT agent, shipped in a week, for $1,000 flat.

I scope one repeatable workflow, wire the connectors, tune it against 30+ real examples from your data, and hand off the agent with a runbook your team can own. Built solo. No agency layer. No new SaaS. No lock-in.

20-min call. You leave with a recommended workflow and rough spec.

Built solo by Ilia. Your agent runs in your ChatGPT workspace, under your admin.
agent · live
example: lead-outreach

Lead Outreach Agent

Connectors: HubSpot · Gmail

triggernew HubSpot contact · lifecyclestage = MQL
readcompany context, past CRM history, public news
decidefit score · stakeholder map · best opening hook
draftpersonalized first email → rep's Gmail drafts
humanrep reviews, edits, sends
build
$1,000
handoff
Day 7

What a typical build actually looks like

Three example agents, drawn from real production scopes. Specifics redacted; full runbooks shared on the call.

01 · Lead Outreach Agent
HubSpotGmail
Before

SDR opens a new inbound lead, spends 30+ minutes researching the company, the contact, recent news, and past CRM history before drafting a first-touch email.

Agent

Watches HubSpot for new inbound. Enriches contact context, checks for past deal history, drafts a personalized first email in the rep's Gmail drafts folder.

Result

First-draft email is ready before the rep opens their inbox. Rep reviews, personalizes the opening, sends. ~30–45 min/day returned per rep.

02 · Support Triage Agent
SlackZendeskDrive
Before

Every ticket lands unsorted in #support. Senior CX reads each one to classify, route, and decide if a doc-grounded reply is possible — before any reply gets drafted.

Agent

Reads every new message in #support. Classifies intent, routes to the right owner, drafts a first-pass reply grounded in your help docs from Drive.

Result

Median first-response under 5 minutes. The same team handles 2–3× ticket volume without adding headcount. Senior CX reviews drafts; nothing sends without their click.

03 · Invoice Reviewer
QuickBooksDrive
Before

AP staff opens each incoming invoice PDF, extracts line items by hand, compares against historical spend, flags anomalies, then keys clean entries into QuickBooks.

Agent

Watches the AP folder for new PDFs. Extracts line items, matches against vendor history, flags anomalies (vendor, amount, account-coding outliers), pushes a clean draft entry to QuickBooks for human approval.

Result

AP review time cut ~70%. Outlier charges caught pre-payment instead of post-close. Human approval still required on every entry.

proof point
3–5 days

From spec to shipped agent

proof point
30+

Real examples tested before go-live

proof point
5–15 hrs/wk

Typical time returned per agent

Works with your stack
HubSpotSalesforceSlackGmailGoogle DriveNotionLinearBigQuery+ any REST API via Actions

Three reasons your team hasn't shipped an AI agent yet

Most teams underestimate connectors, guardrails, and prompt drift. Those are the parts that make the first live agent feel either effortless or immediately fragile.

01

You have 10 good candidates and no way to pick

Every ops, sales, and support team has a dozen workflows that could be automated. Without someone forcing a choice, 'we should try agents' sits on the quarterly-plan backlog for another six months. I pick the right first agent with you on the intro call — it's usually obvious inside 10 minutes.

02

Your team doesn't have a week to learn a brand-new OpenAI product

Workspace Agents, connectors, admin controls, Agent Mode quirks, credit-based billing — none of it is on this quarter's list. That's why I do the learning. You get the working result and a runbook your team can own.

03

The demo worked. Real data broke it.

An agent built in 3 hours on a Friday looks magical. It falls apart the first Monday 50 real tickets hit it. I build against 30+ real examples from your actual data, find the edge cases before you do, and ship with guardrails — so the Monday after handoff is the one where the agent starts paying for itself.

You're a fit if…

Narrow ICP on purpose. If two or more of these don't apply, the engagement won't pay back — better to know on Day 0.

01fit signal

You use ChatGPT Business or Enterprise

Workspace Agents aren't on Plus or Free. Edu and Teachers also work. If you're not sure which plan you have, I check in 30 seconds on the call.

02fit signal

Your workflow lives in connectable tools

HubSpot, Salesforce, Slack, Gmail, Drive, Notion, Linear, BigQuery, Snowflake, Zendesk, Microsoft 365 — all native connectors. Other tools usually bridge via Actions or Zapier.

03fit signal

You can share 20–30 real examples for tuning

Not synthetic data. Real tickets, real leads, real invoices, real meetings — the variety is what makes the agent work in production instead of in the demo.

04fit signal

You can name an internal owner on Day 1

Agents without a named owner decay quietly. The owner doesn't have to be technical — just someone who'll notice when the agent starts behaving weird in week 6.

not for
  • ChatGPT Plus or Free plans (wrong product)
  • Microsoft Copilot Studio / Google Vertex / Bedrock (different stack, different consultant)
  • Regulated workflows (HIPAA, PCI, audit working papers, privileged legal)
  • Non-English / non-US-timezone work
IC
operator · not an agency
Built by Ilia Cherepanov
AI operator. Solo. One client at a time.

I scope, build, test, and hand off the agent myself. Your intro call goes to me. Your spec is written by me. Your agent is built by me. No account managers, no juniors, no “let me get back to you.”

How I work →

Six agents I can ship for you this week.

Each is a scope I've priced and built before. Your version runs on your data, in your workspace, with your admin's permissions.

Not sure which? Take the 2-min quiz →
01 · GTM

Lead Outreach Agent

Every new lead gets researched and a personalized first email drafted — before your rep opens their inbox. 30–45 min saved per rep per day.

inNew contact in HubSpot
doEnrich via web + LinkedIn, cross-reference past deals, draft a personalized email
outDraft in rep's inbox, tagged in HubSpot
HubSpotGmailWeb
$1,000 · 3–5 daysspec →
02 · CX

Support Triage Agent

Every support message classified, routed, and drafted — grounded in your docs. Median first-response drops under 5 minutes; reps handle 2–3× the volume.

inNew message in Slack #support
doClassify severity + product area, check recent changelog, draft a reply
outThreaded reply with draft, routed to the right owner
SlackLinearNotion
$1,000 · 3–5 daysspec →
03 · RevOps

Weekly Metrics Reporter

Monday narrative of what moved, why, and what to do — pulled straight from your warehouse. 2–3 analyst hours saved every week.

inMonday 8am schedule
doQuery the warehouse, compare week-over-week, write a plain-English summary
outSlack post in #leadership with chart links
BigQueryGoogle SheetsSlack
$1,000 · 3–5 daysspec →
04 · Finance

Invoice & Expense Reviewer

AP team reviews the 5% of invoices that actually need attention. The agent handles the other 95% — duplicates, policy checks, and QBO entry.

inNew file in /Finance/Invoices on Drive
doExtract line items, compare to policy + prior invoices, flag anomalies
outException list in Sheets, clean entries pushed to QBO
Google DriveGoogle SheetsQuickBooks
$1,000 · 3–5 daysspec →
05 · GTM

Meeting Prep Agent

A one-page brief lands in every external meeting invite 30 min before start. Your reps stop opening LinkedIn on the way to the call.

inUpcoming external meeting on Calendar
doResearch attendees, pull CRM history, summarize last touchpoints
outBrief doc linked in the calendar invite
Google CalendarHubSpotWeb
$1,000 · 3–5 daysspec →
06 · GTM

RFP & Security Questionnaire Drafter

New RFP → full first draft grounded in your approved-answer library, in under an hour. Saves your SEs 8–16 hours per questionnaire.

inNew RFP uploaded to Drive
doMatch questions to your approved answer library, draft the rest, flag gaps
outDraft document with confidence scores and gap list
Google DriveNotionGoogle Docs
$1,000 · 3–5 daysspec →
process · under two weeks

How it works

One operator, one scoped workflow, one clean handoff. That’s how the timeline stays short.

01

Day 0 — 20-min scoping call

You describe one workflow your team does on repeat. I tell you if it's an agent-shaped problem. If it's not, I say so and point you elsewhere. Free, no obligation.

02

Day 1 — Written spec in your inbox

Trigger, data sources, decisions, outputs, success criteria — one page. You approve, redirect, or kill it. No invoice until we agree on what we're building.

03

Days 2–5 — I build

Inside your ChatGPT workspace, under your admin. I wire the connectors, write the prompt, and tune against 30+ real examples from your data. No consultants in your Slack.

04

Day 7 — Handoff

Loom walkthrough, written runbook, named owner on your team. I walk away with no access. The agent is yours and keeps running without me.

Security model

ChatGPT Business and Enterprise teams have a real security bar to clear. Here's exactly how I work inside it.

01 · control

I build inside your workspace, under your admin

Not a sandbox. Not my account. The agent lives in your ChatGPT workspace from day one, on your plan, with your admin's permissions. No extra SaaS to procurement.

02 · control

Least-privilege connector access

I only request the connector scopes the scoped workflow needs. If the agent reads from HubSpot Contacts, it doesn't get Companies. If it reads Drive folder X, it doesn't get all of Drive.

03 · control

The Day-1 spec names every data path

Before any code: a written spec listing data sources, exact fields, write destinations, approval gates, and guardrails. Your admin signs off in writing before the build starts.

04 · control

Human-in-the-loop on external writes

Any action that sends to a customer, posts externally, or writes to your system of record requires human approval for the first 4–6 weeks. Loosen the gate only after observed reliability.

05 · control

Clean handoff. I revoke. You keep everything

On Day 7 my access is removed. The agent, the prompt, the connector configuration, the documentation, and the runbook are all yours. No lock-in. No recurring access.

06 · control

Standard NDAs signed

Mutual NDA before the spec call if your legal team requires one. SOC 2 posture inherits from your ChatGPT Business or Enterprise plan; agent-specific controls are explicit, not implicit.

Pricing

One flat fee. One week. If the agent doesn't meet spec on Day 7, you don't pay the final invoice — and you still keep everything I built.

01 · Starter
$1,000
per agent, one-time

Best for: your first working agent. One scoped workflow, end-to-end, with a named owner and a real runbook. Most teams recoup the fee in the first month.

  • 20-min scoping call + written spec within 24h
  • 1 agent built end-to-end in 3–5 days
  • 1 data connector wired and scoped
  • Tuned against 30+ real examples from your data
  • 1–2 revision rounds after review
  • Handoff doc + Loom walkthrough
  • 30 days of email support
  • You own the agent. I walk away with no access.
Book intro call
02 · Retainermost popular
$2,500
per month

Best for: teams who want an agent portfolio, not a one-off. Ship 5 agents a month and tune the existing ones — without adding a new hire or retainer renegotiation.

  • Up to 5 agents built or tuned per month
  • Ongoing support for every agent already shipped
  • Async Slack or email within 1 business day
  • Monthly agent-performance review
  • Overages at $1,000 / agent
  • 30-day cancellation — you keep every agent
Book intro call
03 · Retainer Pro
$4,500
per month

Best for: multi-team orgs standardizing agents. 10 builds or tunes a month, quarterly portfolio review, and an admin-control audit so nothing slips through the cracks.

  • Up to 10 agents built or tuned per month
  • Same-business-day response
  • Quarterly agent-portfolio review
  • Admin-control + permissions audit
  • Overages at $1,000 / agent
  • 30-day cancellation — you keep every agent
Book intro call

all prices USD · 10% off on annual prepay · openai credit costs billed separately

Four commitments that decide whether this is worth your $1,000.

New operator, new OpenAI product, no 10-year case study library yet. So the way I build trust is by naming what I promise — in writing — before you wire anything.

01 · Day-7 guarantee

If the agent doesn't meet spec, you don't pay the final invoice.

The Day-1 spec is your acceptance criteria. Miss it on Day 7 and I either fix it to spec on my time, or you keep the deposit as credit against a future build. You never pay $1,000 for something that doesn't work.

02 · Full ownership

The agent lives in your workspace, under your admin.

I build inside your ChatGPT Business or Enterprise workspace. On Day 7 I walk away with no access. The prompt, the connectors, the runbook, the Loom walkthrough — all yours. No SaaS dependency, no retainer trap.

03 · One operator, full stack

The person scoping is the person building is the person handing off.

No account managers, no juniors, no 'let me get back to you.' The intro call goes to me. The spec is written by me. The agent is built by me. The handoff is run by me. You get the same brain for all 7 days.

04 · Narrow by design

I only take work I can ship well.

US-timezone teams on ChatGPT Business or Enterprise. One scoped workflow at a time. If your use case is outside that — regulated data, non-English workflows, Microsoft Copilot Studio — I say so on Day 0 and point you elsewhere. You won't pay for a dead end.

Who builds this

I'm Ilia. I spent the last 3 years shipping AI features at a venture-backed startup — which means I've seen a lot of 'agents' look great in a demo and fall apart in production. I work solo, one client at a time, on scopes I can finish in a week. Your intro call goes to me. Your agent gets built by me. Your handoff comes from me. No account managers, no juniors, no 'let me get back to you.'

More about how I work →
How the engagement feels
01

Solo by design

The person scoping the agent is the person wiring the connectors and tuning the prompt.

02

Opinionated delivery

I keep the scope narrow enough to ship quickly, then expand only after the first workflow works.

03

Handoff over lock-in

You keep the agent, the docs, and the operational context when I step back.

Questions

resources · read more

Everything worth reading before you ship an agent.

88 pages of pragmatic content. Skim the paths that map to your role and stack; the rest is here when you need it.

weekly dispatch · 2-min read

What shipped in Workspace Agents this week — and one operator tip.

No filler. Unsubscribe any time.

20-min call · no deck

Tell me about one workflow.

20 minutes on a call. I'll tell you which workflow is your best first agent, what it would look like running, what it would save, and what it would cost — before you commit to anything. If I can't ship it well, I'll say so and point you where to go.

Best use of the call
01

What team this is for and which tools hold the workflow today.

02

Where the process breaks or burns the most operator time.

03

Whether you are on ChatGPT Business, Enterprise, or still deciding.

If it’s a fit, I send a written spec within 24 hours. If it isn’t, I tell you what I’d do instead.