Internal AI agents for CPG operations

The operating system your CPG brand already wishes it had.

Fieldwork watches distributor, retailer, warehouse, accounting, and inbox data, then surfaces clear answers to your team when inventory, sales, or fulfillment starts to drift.

Request access

See what Fieldwork would flag from your last 24 hours of operations.

Fieldwork agent loop

Last 24 hours

DETECT

DIAGNOSE

ROUTE

UNFI Velocity shift

Hot Honey 12oz down 18% vs trailing average

trendcontextowner

Route to Sales

KeHE Inventory risk

Northeast DC below 2 weeks on hand

trendcontextowner

Route to Supply

Target Recovered

Weekend dip normalized over the last 24 hours

trendcontextowner

Keep watching

Meet your team where they are.

No spreadsheet archaeology. No portal chasing. The agent sends the headline, the diagnosis, and the handoff.

The problem

Your operating truth is scattered across eight portals, three inboxes, and one VLOOKUP nobody trusts.

Every CPG team runs the same loop: log into distributor portals, download CSVs, chase retailer scorecards, reconcile against the warehouse, paste it into a Monday-morning deck. Exceptions surface days late. Decisions get made on stale data.

UNFI

synced 4h ago CSV

KeHE

synced 6h ago CSV

Wegmans

stale 3d CSV

Target

synced 18h CSV

Whole Foods

stale 5d CSV

Albertsons

synced 22h CSV

3PL Vendor

synced 2h CSV

QuickBooks

synced 1h CSV

Today, without Fieldwork Ops manager opens four tabs, exports three reports, builds a pivot, and pings ops in Slack 90 minutes later. By Wednesday, the exception is old news.

With Fieldwork The agent notices Hot Honey 12oz down 18% at a distributor's Northeast DC on Monday at 6:12 AM, attaches the trend and last week's promo context, and posts the diagnosis to the sales lead before the standup.

How it works

A small number of agents, doing the work of a much larger ops team.

Ingest

Scheduled jobs pull distributor, retailer, warehouse, and accounting feeds into a tenant-owned database. No shared tables, no commingling.

Sources

Normalize

Brand catalog, aliases, store hierarchies, DCs, and SKUs are reconciled into a single operating model your team can actually query.

Tenant DB

Reason

Read-only agents run scoped tools against the normalized data. They find the anomaly, attach the context, name the owner.

Agents

Deliver

Answers go where work happens — Slack, email, dashboards — with a citation trail back to the source row and an audit of every tool call.

Channels

Meet the agents

Internal coworkers, scoped by domain — not chatbots scoped by hope.

Every agent runs on a dedicated read-only role with grants only on the tables inside its domain. The prompt explains the rules. The database enforces them.

@wholesale

Distributor desk

Owns the read on distributor inventory, sell-in, velocity, and DC-level supply risk.

Source scope

UNFIKeHE3PL

What people ask it

What's our weeks-on-hand at the Northeast DC for the top 10 SKUs?

Which distributor items dropped more than 15% vs the trailing 13-week average?

Did any of last week's promo SKUs miss the shipment window?

@retail

Retailer desk

Owns the read on store-level sell-through at retailers — voids, dips, recoveries, distribution gaps.

Source scope

WegmansTargetWhole FoodsAlbertsons

What people ask it

Show me retailer stores where my hero SKU went void in the last 14 days.

Which regions are trending below plan this period?

Compare weekly sales vs the same week last year by chain.

@finance · soon

Finance desk

Reads PO status, AR aging, and chargeback exposure straight from accounting — without granting write access to anything.

Source scope

QuickBooksDistributor invoices

What people ask it

What's open AR over 60 days by customer?

Which POs are unfulfilled vs their promised ship date?

Flag distributor chargebacks above the trailing 90-day median.

A real exchange

The agent does the work a junior analyst would, in the time it takes to type the question.

Ask in plain English. The agent picks the right tables, runs read-only SQL against the tenant database, attaches the citation trail, and returns something a human can act on — not a chart that needs more interpretation.

find_sku

Resolves product names, aliases, and distributor item codes to a single catalog identifier.

describe_table

Inspects allowed columns, row counts, and sample values before composing a query.

sql

One read-only SELECT per call. Capped at 1,000 rows. Logged to an audit table you own.

#ops-room · thread with @wholesale

09:42

Ops lead09:42

@wholesale which DCs are below 2 weeks on hand for the top‑10 SKUs?

@wholesale09:42

using describe_table · kehe_inventory

using sql · 1 SELECT · 217 rows

Three DCs are below the 2-week threshold on at least one top-10 SKU:

Distributor — Weeks on hand below 2.0

SKU

WoH

Δ wk

Hot Honey 12oz

Northeast

1.3

−0.6

Spicy Marinade 8oz

Northeast

1.8

−0.4

Cold Brew 32oz

Mountain

1.9

−0.1

Sparkling Yuzu 6pk

Pacific NW

1.4

−0.7

source: distributor inventory · sample data · audit #a1f0

Want me to draft a reorder note for the Supply lead, or compare this to UNFI for the same SKUs?

Ops lead09:43

Now compare against our other distributor for the same SKUs.

Sources

Plug into the systems CPG already runs on.

Fieldwork ships with first-class connectors for the systems most independent brands wrestle with daily. New sources are added per tenant — SPINS, NetSuite, Sprouts, Costco, and many more are in active development.

Distribution

KeHE

● Live

Distribution

UNFI

● Live

Retailers

Wegmans

● Live

Retailers

Target

● Live

Retailers

Whole Foods

● Live

Retailers

Albertsons

● Live

Online

Amazon

● Live

Online

Shopify

● Live

Operations

3PL

● Live

Operations

QBO

● Live

Coming soon

+ many more

SPINS · NetSuite · Sprouts · Costco · …

Built for trust

Read-only by construction. Tenant-isolated by architecture.

Physical tenant isolation

Every brand gets its own database. A shared brand column is never the security boundary. Your data does not coexist with another brand's data.

Read-only at the database

Agent roles have SELECT grants only. No UPDATE, no DELETE, no DDL. The prompt explains the rules; the database enforces them.

Full audit, in your backend

Every tool call, query, latency, and refusal lands in an agent_activity_log table inside your tenant database — reviewable by your team, not ours.

Scoped tool surface

Agents get three tools: find_sku, describe_table, and a one-statement SELECT. 15-second query timeout, 1,000-row cap, single-statement SQL guard.

Your secrets, your env

Tenant credentials sit in your deployment environment, not in Fieldwork code. Brand config holds names, never values.

Source-truth migrations

Every schema change ships as a committed migration with a paper trail. No ad-hoc DDL through a SQL editor — ever.

Request access

The operating system your CPG brand already wishes it had.

Your operating truth is scattered across eight portals, three inboxes, and one VLOOKUP nobody trusts.

A small number of agents, doing the work of a much larger ops team.

Internal coworkers, scoped by domain — not chatbots scoped by hope.

The agent does the work a junior analyst would, in the time it takes to type the question.

Plug into the systems CPG already runs on.

Read-only by construction. Tenant-isolated by architecture.

Physical tenant isolation

Read-only at the database

Full audit, in your backend

Scoped tool surface

Your secrets, your env

Source-truth migrations

The operating system your team already wishes you had.