@teamshift/fake-business

fake-business

fake-business is an open-source generator of realistic, fully fictional small-business datasets — CRM, quotes, jobs, invoices, emails, calls and an event timeline — for testing AI agents and integrations.

Every record links into a believable history (lead → deal → quote → job → invoice → payment), every message references real record numbers, dates and amounts, and realistic mess such as an unanswered lead or an invoice attached to the wrong job is injected on purpose and labeled, so you can score an agent against ground truth.

npx @teamshift/fake-business --industry home-services --seed 42 > business.json

• Coherent, not random. One business, one timeline. Records reference each other by id, every timestamp is causally ordered, and staff activity follows the company's business hours and timezone.
• Deterministic. Same (industry, seed, size, asOf, months, messiness) → byte-identical output, on any machine, forever. No Faker, no network, zero runtime dependencies.
• Labeled mess. 11 kinds of operational problems, each listing the records involved, a one-sentence explanation and the money at risk.
• Safe to publish. Every email is under the reserved .example domain, every phone number is in the fictional 555-0100–555-0199 range, and every company, person and town name is invented.
• Exports everywhere. JSON, NDJSON event stream, CSV, SQLite/Postgres SQL, plus HubSpot- and QuickBooks-shaped import CSVs.

Quickstart

# Human summary of a dental clinic
npx @teamshift/fake-business --industry dental-clinic --seed 7 --summary

# Full dataset as JSON (stdout)
npx @teamshift/fake-business --industry marketing-agency --seed 1 > agency.json

# A SQLite database in one line
npx @teamshift/fake-business --format sql-sqlite --out business.sql && sqlite3 business.db < business.sql

# A perfectly clean dataset (no injected anomalies)
npx @teamshift/fake-business --messiness 0 > clean.json

--summary output (real, --industry home-services --seed 42 --size small):

Copper Kettle Plumbing & Air — home-services, seed 42
2025-10-01 → 2026-09-30 · America/Chicago · Birchford, MN

Records
  employees                        4
  customers                       15
  contacts                        20
  leads                           14
  deals                           12
  quotes                           7
  jobs                            19
  invoices                        13
  payments                        15
  messages                        87
  calls                           34
  tasks                           32
  events                         303
  anomalies                       11

Money
  invoiced                  $19,010.00
  payments received         $18,861.00
  outstanding (A/R)            $298.00
  open pipeline              $2,507.00
  deals won / lost / open     7 / 3 / 2

Schedule
  jobs in progress                 0
  upcoming, next 14 days           6
  upcoming, later                  0

Anomalies (11, $5,343.00 at risk)
  duplicate-contact                1
  quote-not-followed-up            1
  missed-call-no-callback          1
  stale-deal                       1
  overdue-invoice                  1
  invoice-wrong-job                1
  reschedule-not-propagated        1
  missing-contact-info             1
  conflicting-status               1
  unmatched-payment                1
  lead-never-contacted             1

Sample output

An excerpt from npx @teamshift/fake-business --industry home-services --seed 42 (unchanged except for trimming):

{
  "meta": { "generator": "@teamshift/fake-business", "schemaVersion": "1.0.0", "industry": "home-services",
            "seed": 42, "startDate": "2025-10-01", "asOf": "2026-09-30", "synthetic": true },
  "company": { "name": "Copper Kettle Plumbing & Air", "timezone": "America/Chicago",
               "phone": "[phone]", "website": "https://www.copperkettleplumbingair.example" },
  "quotes": [{
    "id": "quo_0ejehqw", "dealId": "deal_0xsewsn", "customerId": "cus_0p0hrlu",
    "number": "Q-1084", "status": "accepted",
    "lineItems": [
      { "sku": "AC-INST-3T", "description": "3-ton 16 SEER2 AC condenser and coil, installed", "quantity": 1, "unitPriceCents": 785000 },
      { "sku": "PERMIT", "description": "Municipal mechanical/plumbing permit", "quantity": 1, "unitPriceCents": 15000 },
      { "sku": "HAUL-AWAY", "description": "Removal and disposal of old equipment", "quantity": 1, "unitPriceCents": 12500 },
      { "sku": "DUCT-SEAL", "description": "Duct sealing (per run)", "quantity": 4, "unitPriceCents": 9500 }
    ],
    "totalCents": 850500, "sentAt": "2025-10-20T13:24:00Z", "expiresOn": "2025-11-19",
    "followUps": ["2025-10-22T15:22:00Z"]
  }],
  "events": [
    { "id": "evt_0t3m6pn", "type": "QuoteAccepted", "at": "2025-10-24T13:40:00Z",
      "subjectId": "quo_0ejehqw", "actorId": "con_1heuwtr", "data": { "dealId": "deal_0xsewsn" } }
  ],
  "anomalies": [{
    "id": "anm_0g90hu5", "kind": "quote-not-followed-up",
    "recordIds": ["quo_12b4hrj", "deal_1cid8yb", "con_1nn89l4"],
    "description": "Estimate Q-1089 for Bianca Lewis ($2,170.00) was sent 2026-03-31 and never followed up; it expired 2026-04-30 with no decision recorded.",
    "amountAtRiskCents": 217000
  }]
}

The email that sent that quote, from the same dataset:

Subject: Your estimate Q-1084 from Copper Kettle Plumbing & Air

Hi Jose,

As promised, here is estimate Q-1084 for AC system replacement.

- 3-ton 16 SEER2 AC condenser and coil, installed x1: $7,850.00
- Municipal mechanical/plumbing permit x1: $150.00
- Removal and disposal of old equipment x1: $125.00
- Duct sealing (per run) x4: $380.00

Total: $8,505.00. Valid through Nov 19, 2025.

What's inside

Collection	What it holds
company	Name, timezone, contact details, payment terms and plain-language policies an agent should respect
employees	Owner, office staff, technicians / hygienists / designers, with roles and hire dates
customers	Households (home services, dental) or businesses (agency), with owner, source, status and tags
contacts	People at each customer, plus prospects who never became customers
leads	Inbound requests in the customer's own words, response time, owner and resulting deal
deals	Pipeline stage, amount, next action, close date and lost reason
quotes	Estimates / treatment plans / proposals with line items, expiry and follow-up timestamps
jobs	Service visits / appointments / project meetings with assignees, schedule and completion; about two weeks of upcoming bookings sit after asOf
invoices	Line items, totals, due dates and amounts paid
payments	Card, ACH, check and cash payments with references
messages	Email and SMS threads whose bodies cite real quote/invoice numbers, dates and amounts
calls	Inbound and outbound calls, missed calls and voicemail transcripts
tasks	Follow-ups, visit prep and recall tasks with due dates and status
events	Append-only timeline of every state change (LeadCreated, QuoteSent, JobRescheduled, PaymentReceived, …)
anomalies	Ground truth for the injected mess (see below)

Three industries ship today, each with its own services, price book, roles, policies and customer language:

--industry	Business	Customers	Flavor
home-services	HVAC & plumbing contractor	households	service calls, equipment installs, maintenance-plan tune-ups
dental-clinic	General dental practice	patients	new-patient visits, 6-month hygiene recalls, treatment plans, insurance-style notes (no real PHI)
marketing-agency	Digital marketing agency	businesses	proposals, website and brand projects with deposits, monthly retainers

Sizes: small ≈ 15 customers, medium ≈ 40, large ≈ 150, over --months of history (default 12) ending at --as-of (default 2026-09-30).

The anomalies

With the default 12 months of history, every kind below appears at least once at every size (small included) for any --messiness above 0 and up to 3; much shorter --months can be too young for some kinds (a deal cannot sit stale for 40 days in a one-month history); --messiness 0 produces none, and higher values produce more. Each anomaly lists its record ids (primary record first), a one-sentence description and amountAtRiskCents where money is involved. The test suite checks that every labeled condition actually holds in the data.

Kind	What is wrong in the data
duplicate-contact	A second contact with a nickname or typo of an existing contact's name and a different email, but the same phone number
quote-not-followed-up	A quote was sent, has no follow-ups and no decision, and is aging or already expired
missed-call-no-callback	A missed call or voicemail from a new caller that nobody returned by call or message
stale-deal	An open deal untouched for 30+ days whose next step is overdue
overdue-invoice	An invoice past its due date with nothing paid (an InvoiceOverdue event is in the timeline)
invoice-wrong-job	An invoice whose jobId points at another customer's job; the correct job id is included
reschedule-not-propagated	A job was moved (JobRescheduled), but the assignee's task and the customer's confirmation still show the old time
missing-contact-info	A customer contact with no email, a malformed email or no phone number
conflicting-status	Records that contradict each other: a deal marked lost although its quote was accepted, an invoice marked paid with no payments, or a completed and paid job marked canceled
unmatched-payment	A payment not linked to any invoice, with the invoice it most likely pays still showing a balance
lead-never-contacted	An inbound lead with no response at all, sometimes with no owner

Programmatic API

import { generate, toSQL, INDUSTRIES, type Dataset } from "@teamshift/fake-business";

const data: Dataset = generate({
  industry: "dental-clinic", // "home-services" | "dental-clinic" | "marketing-agency"
  seed: 7,
  size: "medium",            // "small" | "medium" | "large"
  asOf: "2026-09-30",        // simulated "today"
  months: 12,                // history length
  messiness: 1,              // 0 = clean
});

// Which invoices are overdue right now?
const overdue = data.invoices.filter(
  (i) => (i.status === "open" || i.status === "partially-paid") && i.dueOn < data.meta.asOf,
);

// Score an agent: did it find the ground-truth problems?
const truth = new Set(data.anomalies.map((a) => a.recordIds[0]));

const sql = toSQL(data, "postgres");

Exports: generate, INDUSTRIES, SCHEMA_VERSION, every schema type (Dataset, Lead, Invoice, Anomaly, …) and the exporters below. Runnable scripts live in examples/ (node examples/overdue-invoices.mjs, export-sqlite.mjs, list-anomalies.mjs).

Export formats

--format	Function	Output
json (default)	toJSON(dataset)	The whole dataset, pretty-printed
ndjson	toNDJSONEvents(dataset)	The event timeline, one JSON event per line
csv	toCSVFiles(dataset)	One RFC 4180 CSV per collection, plus company, policies and line-item tables
sql-sqlite / sql-postgres	toSQL(dataset, dialect)	CREATE TABLE per collection plus INSERTs, in one transaction
hubspot	toHubSpotImportCSVs(dataset)	contacts.csv, companies.csv, deals.csv with columns named after HubSpot's default property labels
quickbooks	toQuickBooksImportCSVs(dataset)	customers.csv and line-level invoices.csv with columns modeled on QuickBooks Online import sheets

The HubSpot and QuickBooks files are shaped to be easy to map in each product's import wizard. They are not official templates; review the column mapping when you import.

How is this different from Faker or Mockaroo?

	fake-business	Faker	Mockaroo
What it is	Generator of one complete fictional business	Library of random value generators	Web app / API for schema-defined mock rows
Records linked across CRM, billing and comms	Yes, out of the box	You write the relationships	Partly, via schema references you define
Causally ordered history and event timeline	Yes	No	No
Messages that cite real record numbers and amounts	Yes	No	No
Labeled anomalies (ground truth for evaluation)	Yes, 11 kinds	No	No
Same seed, same output	Yes; no external data, so output only changes with a new generator release	Seedable; generated values can change between Faker releases	Not a documented feature
Runs offline	Yes, zero runtime dependencies	Yes	No, hosted service
Arbitrary custom schemas	No, a fixed small-business schema	Yes	Yes

Use Faker or Mockaroo when you need arbitrary tables of plausible values. Use fake-business when you need a believable business whose data has to make sense end to end.

FAQ

How do I test an AI agent without real customer data?

Generate a fictional business with npx @teamshift/fake-business --seed 42, load it into the tool your agent uses (a database, CRM sandbox or MCP server), give the agent a task such as "follow up on every stale quote", and compare what it did with anomalies. Because the dataset is deterministic you can rerun the same scenario after every prompt or model change.

Is the data safe to publish?

Yes. Every dataset sets meta.synthetic: true. Emails use the reserved .example top-level domain (RFC 2606), phone numbers are in the 555-0100–555-0199 range set aside for fiction, and company, town and street names are invented. People's names come from a curated list of common US names, so any match to a real person is coincidental. Dental records are invented and contain no real health information.

Can I import it into HubSpot?

Use --format hubspot --out ./hubspot to get contacts.csv, companies.csv and deals.csv with HubSpot-style column names and an External ID column, then map the columns in HubSpot's import tool. Try it in a sandbox or test portal first.

Can I import it into QuickBooks?

--format quickbooks writes customers.csv and an invoice-line CSV with column names modeled on QuickBooks Online import spreadsheets. Import into a sandbox company and check the mapping.

How do I get a SQL database of fake business data?

npx @teamshift/fake-business --format sql-sqlite --out business.sql, then sqlite3 business.db < business.sql. Use --format sql-postgres for PostgreSQL (psql -f business.sql); JSON fields become JSONB.

Is the output really reproducible?

Yes. The generator uses its own seeded PRNG and ships its own word lists, so the same options produce byte-identical JSON on every platform. The default --as-of is a fixed date (2026-09-30) rather than today for the same reason.

How do I evaluate an LLM agent with this data?

Treat anomalies as the answer key: each entry names the primary record to act on, so you can compute precision and recall of the records your agent flags or fixes. The events timeline lets you check that the agent's actions happened in a sensible order.

Can I get clean data with no anomalies?

Yes: --messiness 0. Clean datasets still pass every integrity check (all references resolve, totals match line items, payments match invoices).

Contributing

Issues and pull requests are welcome at teamshift-io/smb-sandbox.

Adding an industry is one file. Copy src/industries/home-services.ts, change the services, price book, roles, policies and customer language, then register it in src/industries/index.ts (and add its id to IndustryId in src/schema.ts). The simulation, anomalies, exporters and tests pick it up automatically. Run pnpm build && pnpm test before opening a PR.

License

Apache-2.0

Built by TeamShift — AI workers for small-business operations.