AI adoption case study: From Copilot only to agents, pipelines and governance

DevToDollars helped a non technical Toronto company put AI to work: agents where they fit, plain code where they did not, and clear limits on what the AI can reach.

7 min readNames in these screens are placeholders.

A PMO status slide for a body butter in production, marked On track with its delivery date. Formula, stability, artwork, packaging, cost, claims and risk run down the left; latest updates, discussion topics, actions, purchase orders and their dates fill the right; five milestones run along the bottom, three of them done. The operations dashboard open on one retail customer: invoiced sales, units shipped and open orders from NetSuite beside units sold and sales velocity from Power BI, above a chart of weekly velocity from January 2025 to October 2026.
One of the PMO status slides now built from the project tool, on a display, and the operations dashboard DevToDollars built for management, on a laptop.

  • 120+ hours

    of manual work saved every month in demand planning

  • 4 to 2

    customer service staff for the same workload

  • Real time

    project status for management

The results

Results at a glance

Client
A mid sized Toronto company, name withheld
Company
Non technical, several brands
Starting point
Copilot only, no agents
Systems
NetSuite, Power BI, a project tool

Demand planning

120+

hours of manual work saved every month

More than three 40 hour work weeks

Customer service

From 4 to 2

staff for the same workload

Management

Real time

project status, straight from the project tool

Status slides built automatically

What DevToDollars did

  • Discovery and assessment
  • Automated status reports
  • AI agents
  • Data pipelines
  • Team training
  • Access and data governance
Hands working through figures with a pencil, between two open laptops.

The starting point

Core processes ran on exports, Excel and memory

AI meant Copilot, used only as an assistant, never as an agent that carries a task through the company's own systems.

  • Excel by hand

    Thousands of exported rows, worked through by hand in Excel, led to inaccurate results.

  • Departments apart

    Each department had its own processes, data sources and tools.

  • Status from memory

    Project stages in the tool were out of date, and the PMO, the team that tracks projects, built status slides by hand, from memory. Management could not see which projects were late, how late, or why.

Photo: Scott Graham on Unsplash

Most companies use AI. Few have agents doing the work.

Nearly

9 in 10

use AI regularly somewhere in the business

Only

22%

of companies under $1 billion in revenue say they are scaling AI agents

Source: McKinsey, The state of AI in 2026, August 2026.

Most spreadsheets contain errors.

At least

86%

of the spreadsheets checked in field audits had errors

Source: Raymond Panko, Spreadsheet Errors: What We Know. What We Think We Can Do, EuSpRIG, 2000.

Discovery and assessment

First, a map of how every department works

Leadership had measured the broken workflows but had no one to fix them. The DevToDollars team started by asking each department's leaders and senior managers the same four questions.

A hand pinning cards and thread onto a wall map of connected steps. The same four questions, every department
Photo: Alvaro Reyes on Unsplash

  1. Which departments share data with each other?

  2. What is the longest workflow each department needs solved?

  3. Which third party tools does each project rely on?

  4. How does each department do its work today, in detail?

What the map decided

  • Fix first

    Data sources that disagreed had to be fixed before any AI went in.

  • Get ready

    Each department had cultural needs to settle before a rollout.

  • Start here

    The work would start where the impact would be highest.

Why the data came first

Gartner predicts

60%

of AI projects without data ready for AI will be abandoned through 2026

Source: Gartner press release, February 2025.

The solution

No single tool was forced on every team

For each department's longest workflow, DevToDollars proposed a design that fit the problem: AI agents for the PMO and marketing, plain code for demand planning.

  1. AI agent PMO Status reports built from the project tool
  2. AI agents Marketing Campaign updates pulled from the web
  3. Plain code Demand planning Pipelines that feed the forecasting models

PMO reportingAI agent

Status reports straight from the project tool

Status slides are no longer built by hand, from memory. Every major project now gets one straight from the project tool.

  1. Project tool

    Where every project's stages live

  2. AI agent

    Builds the report with ICM

  3. Status slide

    One of two templates, in the brand's colours

  4. Management

    Sees every major project, in real time

A PMO status slide for a body lotion still in development, marked Hold while a brand decision is pending. Formula, stability, packaging, artwork, cost, claims and risk run down the left; latest updates, discussion topics, actions and a table of formula submissions fill the right; five milestones run along the bottom, two of them done.

In development. On hold while the brand team decides. Formula submissions on the right, milestones along the bottom.

A PMO status slide for a body butter in production, marked On track with its delivery date. Formula, stability, artwork, packaging, cost, claims and risk run down the left; latest updates, discussion topics, actions, purchase orders and their dates fill the right; five milestones run along the bottom, three of them done.

In production. On track. Purchase orders replace formula submissions, and the milestones run to the earliest required ship date.

A reference sheet titled Same layout, each brand's own colours: four brand palettes side by side, each colour listed with its hex code and the part of the slide it is used for, such as titles, the left panel, lines and the brand tag.

Each brand's own colours

One layout for every slide, in the colours of the brand it reports on.

How ICM works

ICM, short for Interpretable Context Methodology, is an open method built for step by step work.

  1. A numbered folder for each step
  2. Instructions in plain text anyone can read
  3. Scripts for the steps that need no AI
  4. A person reviews each stage

MarketingAI agents

Agents the team learned to train

The job was research: the latest updates on marketing campaigns, pulled from the web by semi autonomous agents.

Three people around one laptop, pointing at the screen together. Semi autonomous agents, Hermes Agent among them
Photo: John on Unsplash

  1. Search

    Agents pull the latest campaign updates from the web, Hermes Agent among them.

  2. Remember

    Hermes Agent, open source from Nous Research, keeps a memory and refines its own skills.

  3. Train

    DevToDollars coached the team to train the agents, so results match what it looks for.

Demand planningPlain code

Plain code where AI was not needed

Not every automation needed AI. Where the forecasting models had been fed by hand, DevToDollars built pipelines in code instead, and introduced prediction models that forecast more accurately.

120+hours of manual work saved every month

Before

  1. Data exported by hand
  2. Forecasting models fed by hand

After

  1. Exports automated
  2. Pipelines in code feed the models
  3. New prediction models, more accurate forecasts

Why code here

Code runs the same way every time and uses no tokens. Because DevToDollars writes software, it could pick the tool for each job on its merits, saving the company time and token costs.

About

1 in 5

companies say AI running costs, tokens included, limit how much they use AI

Source: McKinsey, The state of AI in 2026, August 2026.

A dashboard for management

The team also built a custom dashboard that gives management a bird's eye view of demand planning: what shipped to each customer, from NetSuite, beside what it sold, from Power BI.

  1. Every retail customer and its invoiced total
  2. Shipped and invoiced, from NetSuite, beside sold, from Power BI
  3. Units sold per store, week by week
  4. Shipped against sold, product by product

Click the dashboard to see it full size. Names are placeholders.

The whole operations dashboard for one retail customer: the customer list, the headline numbers, weekly velocity, velocity by time window, shipped against sold by month, units sold by month, orders by month with open orders marked, and shipped against sold for each product.

Security, data and governance

Where to trust AI, and where not to

Leadership's first worry had been AI near the company's tools and data. Every employee was walked through where it can be trusted, and where it cannot.

  • Access

    DevToDollars mapped the access each department's AI should have to company software, then set it for every user.

  • Responsibility

    Where an agent can change data in company systems, the person running it is responsible for those changes.

  • Data privacy

    The company uses an AI provider's enterprise plan, so its data stays protected.

  • Secrets and keys

    Keys to outside services are stored securely, with guards against outside attacks.

Why the rules matter

Of AI users at work

78%

bring their own AI tools, which the report says puts company data at risk

Source: Microsoft and LinkedIn, 2024 Work Trend Index, May 2024.

Photo: Taylor Vick on Unsplash

Interested in chatting with us?

Get a free assessment of your business and an estimate of your project scope.

Book a free consultation Email us