Imagine a fictional building-materials distributor selling paint, primer, and dry mixes. Orders arrive through WhatsApp, deals live in HubSpot, and prices and stock live in its ERP. Reps assemble invoices by hand. The director spends evenings combining sales spreadsheets.

AI business process automation starts on the director’s computer. An agent helps investigate sales, the team adopts it for everyday tasks, and the director builds a small internal tool. Engineering becomes the next step when that tool enters daily company operations.

The distributor’s path in this article
  1. The director and staff on their own
    Agent report

    The director and an ERP export on one computer

  2. AI in each team

    Sales, purchasing, and finance take small tasks

  3. Internal tool

    A quote calculator built with vibe coding

  4. With developers
    Production system

    ERP invoices, CRM deals, checks, and access control

The example that runs through this article. The first three steps need no developers

Here is a typical customer conversation:

Fictional conversation. The rep still looks up the order, checks stock, and prepares the invoice

The first AI report

Revenue looks steady, but the director suspects some regular buyers have gone quiet. They install Codex or Claude Code, open a working folder, and add an anonymized sales export. It contains account codes, dates, products, reps, and amounts.

An agent can read the files, run calculations, and save the output. The director describes the task in plain language. They want a finished report to open, check, and share with sales.

Find accounts with no September ordersWorking foldersales.xlsxERP exportAIAgentCodex · Claude Codereport.xlsx3 accountswith source rowsreadsmatchescalculatessaves
The director describes the task in words. The agent opens the files, runs the numbers, and saves the report in the same folder

The first task is specific:

Find accounts that ordered in July and August but placed no orders in September. Show monthly totals and the rep. Save a separate report. Reference the source rows for each finding. If data is missing, do not invent it.

The agent matches accounts and produces this report:

Accounts with no ordersSeptember · sales.xlsx
3 regular buyers placed no orders

They ordered in July and August. September has no orders for these accounts.

Total revenue barely moved Revenue by month, $
40,010
40,040
40,500 0
JulyAugustSeptember
  • A-014, A-027, A-041
  • Other accounts
AccountJulyAugustSeptember
A-0141,8002,1000
A-0279601,0400
A-0411,2501,3000
Teaching report, all amounts are fictional. A list to check and follow up, not proof of lost customers

Other customers’ purchases concealed the missing accounts in the revenue total. The director now has a concrete follow-up list. They verify the rows and hand them to reps. One buyer stocked up early, another may have received no reply, and another may have switched suppliers. People investigate the reason; assembling the evidence is already done.

The company and figures are fictional. Your export may reveal rising demand for a product, recurring payment delays, or accounts that buy one item but never its usual companion product.

From one report to a daily routine

A week later, the director adds stock and unpaid invoices. They ask for one morning report: accounts to call, products to reorder, and payments to check. They save the calculation definitions beside the files so the next report uses the same method.

salesstockunpaid invoicesAIcalculation notesMorning reportaccounts to call3items to reorder5payments to check2the director still refreshes the exports
Three exports and saved calculation notes produce one morning action list. Figures are fictional

The result is a file and an action list. The director still updates exports and starts each run. Direct ERP access and scheduled execution can follow later; installing an agent does not connect business systems automatically.

The team joins in

Reps want help with replies, purchasing wants stock analysis, and finance wants document comparisons. Each starts with a small task whose result they can check.

Each team starts with its own task
Sales
Customer question and price list AI Draft reply Checked by: sales rep
Purchasing
Stock and sales AI What to reorder Checked by: buyer
Finance
Order and invoice AI Differences between them Checked by: accountant
Shared ground rules
  • One place for the current price list
  • Trials use anonymized data
  • Every output has a named checker
AI prepares the draft. The decision and the check stay with the employee

They agree where the current price list lives, which data can go to a service, and who checks outputs. Trial tasks use anonymized copies rather than raw customer databases.

Corporate AI training helps turn experiments into repeatable work using the team’s own tasks and checking criteria. The director can start with individual AI training. Our guide to business processes you can automate with AI covers other departments.

Vibe coding: an internal tool

A rep asks: “Could I just select a product and quantity and get a quote?” The director has seen the agent work with files. Now they ask it to build a small application.

Vibe coding means describing behavior, letting AI write code, trying the result, and requesting changes. For a first internal calculator, the director needs to understand the price list and recognize correct calculations. They do not need to learn a programming language before trying it.

How the tool takes shape
Check
1 · You Describe what the page should do
2 · AI Writes the code and runs the page
3 · You Try it on ordinary orders
4 · You Ask for a fix or a feature
Calculator versions
  1. v1 Price-list quote, draft, and “needs confirmation”
  2. v2 Discount for regular accounts
  3. v3 Price-list date
AI writes the code. A person sets the task and checks the result on real examples

They provide the price list and ask:

Build a quote page using only the attached price list. Let me choose a product, set a quantity, and save a draft. Show “needs confirmation” for a missing product. Leave customer sending and ERP writes out of this version.

A small tool takes shape. Try this teaching version:

Quote builder
Teaching price list: $120
Total
$1,440
Working teaching prototype: pick a product, change the quantity, and save a draft. Nothing is sent to a CRM or ERP

The director checks a normal order, a missing product, and an empty quantity, then gives the page to one rep. The rep asks for account discounts and a price-list date. The agent adds them; the director checks calculations again.

The same approach can produce a file comparison tool, a sales dashboard, or a search interface for internal instructions. Anthropic describes a marketing workflow that processes advertising CSVs and creates new variations, and a legal team’s prototype for routing requests to the right lawyer.

The distributor has a modest calculator. It solves a real task, and employees have helped improve it. It may remain useful for a long time.

The tool grows

All reps now use the calculator. The director wants prices updated from the ERP, quotes saved against deals, and reports produced without manual exports. Requirements grow with adoption.

The tool grew. So did the requirements
A report for yourself
check rows in the ERP
Drafts in each team
data rules, a named checker
Calculator for one rep
checked calculations
Calculator for all reps
one price list, no duplicates
Writes to ERP and CRM
access, monitoring, tests
From here, bring in developers
What breaks and what fixes it
Reps see different prices
One current price list
Two clicks, two invoices
Duplicate protection
An update broke discounts
Checks before release
Stopped during a busy day
Monitoring and recovery
More users and costlier mistakes mean more engineering. This is where a prototype becomes work for a development team

Asking AI to fix a problem sometimes works. Understanding why it returns, or what else changed, is harder. One employee sees a working page while another gets a loading screen. Two reps open the same deal. An old price list stays cached.

The director now needs a development team. The prototype already shows what staff use and where they lose time. You can get help before a serious failure: review a tool before expanding its access to money, customer data, or business-system writes.

We join as developers

We start with what the director has built. We review the calculator, check calculations, and keep useful decisions. Then we fix failures, connect the ERP and CRM, and add employee access and an activity history. A complete rewrite is not always necessary.

For the distributor, the next step turns a quote into an ERP invoice draft. Prices and stock come from the business system. A rep approves customer sending. A repeated click cannot create a second invoice, and an ERP outage leaves the request queued. This extends the tool people already use.

repeat click: no second invoiceCalculatorquoteQueueERPinvoice draftHubSpotdeal updatedSales repapproves sendingERP offline: request waits
A quote becomes an invoice draft. A repeated click creates no duplicate, and an ERP outage leaves the request queued

We check changes against real examples: a normal order, a substitution, a discount, a cancellation, and an incomplete address. Those cases must still work after updates. See why AI projects need evals for the checking approach and the 30-day AI pilot guide for a gradual rollout.

Checks before each update 5/5 scenarios
  1. Normal order → invoice draft
  2. Substitution → price and stock from the ERP
  3. Discount → regular-account discount
  4. Cancellation → no invoice created
  5. Incomplete address → routed to a rep
Ready to release
A set of real orders runs after every change. If one case fails, the update waits

Engineering experience matters for integrations, calculations, and recovery. AI accelerates our work; the team remains accountable. A director who wants a dependable system from the start can skip independent experiments. Both paths fit our AI implementation work. The Magnum HR agent shows one system handling live conversations.

The director started with a question about quiet accounts. The team now has reports, a shared tool, and a clear next step. An agent and existing data made the first improvements possible. Engineering entered when the tool became important to the whole company.

FAQ

RPA repeats predefined actions such as opening a window, moving a record, or clicking a button. AI helps interpret messages and documents whose wording varies. In the distributor's order, AI interprets “same as last time”; prices and availability come from the business system.

Yes. Start with export analysis, draft replies, or a small internal tool in Codex or Claude Code. An employee checks the result. Bring developers in when access controls, unattended operation, integrations, or the consequences of an error require engineering work.

Usually you can keep them. Exports are enough for initial experiments. Ongoing automation needs configured data exchange, and capabilities vary by system and version. Check interfaces, catalog quality, and access permissions before development starts.

Reps approve invoices at the start. Automatic sending can be considered for tested scenarios after measuring errors. Missing addresses, unknown products, and disputed prices should stop sending and route the order to a person.

Use anonymized copies for experiments: account codes instead of names, without phone numbers or payment details. Before connecting live systems, define what data goes where, who can access it, and your storage requirements. Local folder access does not mean an agent processes everything exclusively on your computer.