1. Home
  2. Resources
  3. Case studies

Case studies: AI agents and ERP systems our team built

Here are four projects our team built, each set out the same way. You get the system, the job, what the software reads and writes, how it connects, who approves and where it stands. Three run on ERPNext, an open-source ERP. The fourth reads crash news for a US law firm. None has a measured result yet, so none is quoted.

4 case studies3 on ERPNext and Frappe1 on a dedicated US server
The four case studies on this page
Multi-agent RFQ and RFP system ERPNext
RFP agent and project management ERPNext
Full ERP for a food processor ERPNext
Crash news agent for a law firm Own server

Each one has its own labelled block below.

What this page holds

Four case studies of work our team built. Three blocks name the client and one, for a law firm, does not. Each answers the same seven questions: which system, which job, what goes in, what comes out, how it connects, who approves, and where it stands today.

An ERP is the software a company runs its orders, stock and accounts in. A CRM holds its customers and deals. An AI agent is a program that uses an AI model to do one set job, such as reading an RFP. No block quotes hours or money saved, because nobody has measured them.

Case study: FactoryJet's RFQ and RFP agents, ERPNext workflows and Frappe CRM

Our team built three things for FactoryJet, all in open-source software from Frappe. The first is a system of AI agents for RFQs and RFPs. An RFQ is a request for a quote, and an RFP is a request for a proposal. The second is a set of AI workflows inside its ERPNext. The third is a Frappe CRM set up around the way FactoryJet sells.

Scroll the table sideways to see every column.

QuestionIn this project
System and versionERPNext and Frappe CRM, installed side by side on FactoryJet's own Frappe site.
The jobTurn a customer's RFQ into a quote. Then send RFQs to suppliers and read what they send back. The customer side was built first, the supplier side second.
What goes inThe customer's request, with its drawings and CAD files (the design files that parts are made from). Then each supplier's reply to an RFQ.
What comes outQuotes for customers and RFQs for suppliers, saved in ERPNext and Frappe CRM. The ERP workflows produce bookkeeping entries, invoices, sales orders and quotations.
How it connectsSeparate agents handle intake, item matching, pricing, supplier RFQs and reading supplier replies. All of them run in ERPNext and Frappe CRM, where the team already works.
Who approvesA person approves every quote before it goes out to a customer.
Where it standsFactoryJet runs ERPNext and Frappe CRM with the AI workflows for bookkeeping, invoices, sales orders and quotations. The RFQ and RFP system was built for both sides.

A CRM shaped around the sales process

Frappe CRM is open source, so FactoryJet pays no per-seat fee for it. We set it up around FactoryJet's own steps for selling. The software fits the team's process, so the team does not have to change how it works to fit the software.

No hours saved or win rates are given for this system. None were measured.

Pages about this kind of work

Case study: project management, RFQs and an RFP agent for Impulse Branding Solutions

For Impulse Branding Solutions, our team built project management, RFQ automation and an AI agent that handles RFPs. All of it runs on ERPNext and Frappe, the open-source framework ERPNext is built on. This is a separate project from the RFQ system above, with its own job.

Scroll the table sideways to see every column.

QuestionIn this project
System and versionERPNext and Frappe.
The jobFind RFPs that suit Impulse on bid portals. Fill each one in with Impulse's pricing. Submit it once a person has done the final review. Projects are tracked in the same system.
What goes inRFP listings on bid portals (websites where buyers post the work they want bids for), the RFP papers themselves, and Impulse's own pricing.
What comes outA filled-in reply for each RFP that suits Impulse, ready for review. After the review, the reply is submitted on the portal.
How it connectsThe agent runs on ERPNext and Frappe, where Impulse also runs its project management. It goes to the bid portals to find RFPs and to submit replies.
Who approvesA person does the final review of every reply before the agent submits it.
Where it standsBuilt: project management, RFQ automation and the RFP agent, all on ERPNext and Frappe.

The agent finds the RFPs and fills them in. Submitting is the step that sends Impulse's prices to a buyer, so before it submits, the agent waits for a person to sign off.

No bid counts or win rates are given for this agent. None were measured.

Pages about this kind of work

Case study: a full ERPNext rollout for GroFresh Agro, a food processing manufacturer

GroFresh Agro makes processed food. Before this project, it kept its accounts in Tally, an accounting program, and the rest of its numbers in separate spreadsheets. Each team had its own figures. Nothing linked the raw material that came in to the raw material used on the floor, so stock went missing between the two. Managers worked from old reports, and finance matched production yields by hand at the end of each month. Our team replaced both with one ERPNext system.

Scroll the table sideways to see every column.

QuestionIn this project
System and versionERPNext, run in Docker containers on AWS, Amazon's cloud. It replaced Tally and the spreadsheets.
The jobRun the whole plant in one system, from buying raw material to sending out finished goods.
What goes inPurchases, stock moves, production entries and accounts, recorded in ERPNext as the work happens.
What comes outBatch records with lot numbers and expiry dates, yield reports, and live dashboards for production yield and wastage.
How it connectsBuying, stock, production and finance share one set of records. Finance reads the same numbers as the plant, with nothing copied between Tally and a spreadsheet.
Who approvesThere is no AI agent in this project. GroFresh's own teams enter their records in ERPNext.
Where it standsDelivered. Procurement, inventory, production and finance run in one ERPNext system.

How the rollout ran

  1. We mapped each workflow as it ran before setting anything up, so the system follows how the plant works.
  2. Procurement and inventory went live first, because that is where the stock gaps were.
  3. Production planning went live next.
  4. Finance came last. The plant kept running while each part switched over.

What food processing needs from an ERP

A plain setup tracks how much stock you hold. A food plant also needs to know which batch the stock came from, when it expires, and how much finished product each batch of raw material gave. So batch tracking, expiry dates, yield reports and lot tracing sit in the same system as buying and stock. Lot tracing means following a finished product back to the raw material it was made from. At GroFresh, any batch can be traced that way.

We chose ERPNext because it is a full ERP, built in modules, with no license fee to pay.

This case describes the system we delivered. It gives no efficiency figures, because none were measured.

Pages about this kind of work

Case study: a crash detection agent for a US personal injury law firm

A US personal injury law firm wanted to know fast when a truck or another commercial vehicle was in a fatal or life-threatening crash. Those crashes are often first reported by a local TV station, a small paper or a state police press release. No one at a firm can read all of those across 50 states every day. Our team built an AI agent that does the reading.

Scroll the table sideways to see every column.

QuestionIn this project
System and versionNo ERP. A Node.js app with a SQLite database, on a dedicated server in the US. An AI model reads the articles, and plain rules make the decisions.
The jobFind fatal or life-threatening crashes that involve a commercial vehicle and are 14 days old or newer, then email them to the firm.
What goes inA licensed news API (a service that hands news data to other programs) on the firm's own paid plan, the GDELT public news index, 34 local news feeds, the California Highway Patrol incident feed, and state police newsrooms that post crash releases. It checks them every two hours, every day.
What comes outOne email per crash that qualifies. A follow-up goes out only if the victim's hometown is published later, so a crash gets two emails at most. Every lead also shows in a private console.
How it connectsEmails go out through Resend, an email service. The console sits behind Cloudflare Access, which checks each person's email with a one-time code before the page loads. The server has no public door, and its database is backed up every day, encrypted, to separate storage.
Who approvesLawyers at the firm review every lead and decide which ones to follow up.
Where it standsLive on a dedicated US server. It checks every two hours, weekends and holidays included.

How the agent decides

For each story, an AI model pulls out the facts: the vehicle, how bad the crash was, the date, the place, and any victim's name, age and hometown. Then fixed rules decide. A commercial vehicle has to be in the crash itself. Someone has to have died or been critically hurt, for example flown out by medevac. Stories about lawsuits, new laws or news roundups are turned away. At the firm's request, rideshare, app delivery and taxi crashes count too, because those cars carry commercial insurance while working.

What each email says

The subject line follows the firm's own format: the town, the state, and the victim's name and age. The body gives the place, the date, the vehicle type, a short summary and a link to every source. If a source says something about fault, the email shows it, but the agent never uses fault to keep or drop a lead. For a few states where the waiting-period rule has been checked, the email shows the earliest date the firm may contact a family, with the rule behind it. For every other state, it says the rule is unknown. Each email notes that this is not legal advice.

Checks against made-up details

Before any name is saved, the agent checks that the name is in the article text. If it is not, the name is dropped. In real runs, this check has caught names the model made up. When no name has been published yet, the crash goes on a watchlist. The agent reads its sources again once a day and emails the firm when a name appears. The same crash reported by several outlets becomes one record, so the firm is not emailed twice.

No lead counts or case results are given here. None have been measured, and the firm decides which leads to follow up.

Pages about this kind of work

Tell us the job and the system it runs in

Name the ERP or CRM you run and the paperwork your team types in by hand. A person on our team replies by email. The first call is free.

Tell us the job

Questions about these case studies

Why don't these case studies show results, like hours saved?

Because nobody has measured them yet. A number goes on this page only after someone measures it and the client agrees to share it. So each case says what was built, what it reads and writes, and who checks its work. The numbers you do see, such as 50 states or every two hours, describe how a system works, not what it saved.

Why is one client not named?

We keep that client's name off this site. The crash detection agent was built for a US personal injury law firm, and we describe it without the firm's name or where it is based. Every other fact in that case comes from the delivered project.

Three of these run on ERPNext. We use NetSuite. Can you still do this kind of work for us?

Yes. ERPNext is the system these clients run. Our team has taken nine ERPs live, from SAP to QuickBooks, so the system is your choice. The others are Oracle NetSuite, Odoo, Microsoft Dynamics, Xero, Sage and custom ERPs. In the usual design, in any of them, the agent reads the document, applies your rules and saves the result in your system for a person to check.

Does a person check the AI's work in these builds?

Yes, in all three agent builds. A person approves every quote before it goes to a customer. In the RFP agent, a person does the final review before anything is submitted. In the crash detection agent, lawyers review every lead, and the agent never decides who was at fault. The fourth case is an ERP rollout with no agent in it.

Can an AI agent make up facts, like a name that is not in the document?

An AI model can, so the build has to catch it. In the crash detection agent, every victim name the model returns is checked against the article text before it is saved. A name that is not in the article is dropped and never emailed. In real runs, that check has caught names the model made up and thrown them out.

Who owns a system like this once it is built?

The contract for each project sets it.

We would share real RFPs and price lists. Will you sign an NDA first?

Yes. We sign your NDA before you share any data, if you ask. An NDA is a contract that says we keep what you share private. A few past documents, with the records your team made from them, are enough to talk through how an agent would handle the job.

How does a project like these usually start?

With one job and the documents behind it. Tell us the system you run and the paperwork your team types in by hand, through the form on our homepage. A person on our team replies by email. What a build includes is agreed per project. The first call is free.