Goodwork.ai | Enterprise AI for retail analytics

Thesis introduction

Data coverage & scale

Retail specificity & accuracy

Speed & ease of use

Engineered for trust

Built to evolve & scale

Thesis introduction

In a retailer, category managers decide which product to sell, where to sell them, how to price them, how to promote them, and review a product's historical performance. They are the lifeblood of a retailer and the engine that drives profit making activity.

To support category managers, big retailers collect vast amounts of data on products, their performance and the world around them. They build dashboards and employ entire finance teams to answer questions and prepare regular reports. A category manager might have dozens of spreadsheets open, each containing multiple tabs.

Yet, despite all of this data, we think category management is broken.

We surveyed employees across three business units of a large Australian retailer. Despite the massive volume of data available, 80% wanted help with basic reporting, and 60% wanted help with deeper questions such as explaining why something happened, understanding outside market competitors and providing recommendations / opportunities. One person we spoke with said

“We’ve never had more dashboards and data and we’ve never been more confused. 90% of our time is in getting and wrangling the data and 10% of the time is in making decisions”.

Why is there this disconnect? The reality of the tools available for category management today is this:

Key problems we're addressing

Data takes work.

Raw data needs to be cleaned, combined and processed.

Coverage is incomplete.

Explaining performance needs a broad view of the trading environment. Too often, important context—external events, weather and competitor data—is missing.

Spreadsheets are unreliable.

Errors creep in, versions diverge, and numbers go out of date.

Dashboards are rigid.

They answer the questions their authors anticipated, leaving little room to explore new ones.

The harder questions need specialist help.

Understanding cannibalisation, customer switching or even basket size means analysing millions of transactions—far beyond what a category manager can reasonably tackle in a spreadsheet.

But finance is already stretched.

Requests for further analysis join a queue and can take days or weeks.

Analysis is inconsistent.

The quality of the answer depends heavily on who does the work.

Trust breaks down.

Messy data and conflicting answers leave people unsure which numbers to believe.

And eventually, people stop asking.

Getting an answer takes too much effort, so decisions fall back on gut feel.

All of this has consequences. Monday morning meetings become about reporting what rather than figuring out why and what’s next. Supplier negotiations lean on relationships and persuasion, with little ability to test a supplier’s claims or challenge their numbers. Useful questions go unasked because getting an answer would take two or three weeks.

At Goodwork, we solved this by building our own trustworthy, 24/7, retail analyst that everyone can work with. Users ask questions in their own language and get answers in seconds. It brings internal and external data together, follows questions beyond fixed dashboards, and handles analysis that would otherwise require specialist help. It applies consistent retail methods aligned with finance, so everyone works from the same analytical foundation. Above all, it makes asking the next question easier than falling back on gut feel.

We call it Atlas. It knows the retail landscape, connects the dots, and helps you decide where to go next.

Building Atlas means more than solving a chat problem. It means developing a retail specific harness, giving it the data to see the whole picture, and building the analytical tools, libraries to quickly turn that data into answers retail people can trust.

Data coverage and scale

“Retail is notoriously difficult. You need a critical mass of the right data before the tools become useful. It’s like having a great car with no windscreen—the rest might be excellent, but it still isn’t ready to use.” - Category Lead

Explaining performance needs a broad view of the trading environment. Atlas’s data foundation brings together internal and external data, ensures it is clean and connected, and makes it ready for fast analysis at scale.

Internal data is consolidated:

The data foundation starts by bringing together a retailer’s internal data into a unified data cube from over 20 sources covering sales, basket transactions, finance, products, brands, suppliers, price, promotions, stock, ranging, stores, channels, budgets, customer and loyalty data.

Retail rules are built into our data preparation process. Product and store identifiers are unified across datasets pack sizes and units are standardised, and retail calendars and trading days are aligned. Sales, stock and ranging information are then connected, giving Atlas the context to distinguish weak sales from changes in availability. These rules give Atlas a consistent foundation for its analysis.

Swarms of AI agents clean and harmonise the data:

Specialised agents resolve inconsistencies that conventional cleaning rules struggle with. For example, a brand may appear under several names, splitting its performance across apparently separate businesses. Our agents investigate product examples, category context and brand relationships to assess whether labels refer to the same brand.

Calibration:

The final internal data is also calibrated against internal finance reports, with reconciliation checks repeated periodically.

External data is added:

The data cube is enriched with external data on factors that impact retail performance, including weather, public holidays, local events and competitor activity.

Weather is estimated at the store-day level for each store using distance weighted interpolation techniques. Public holidays and local events are sourced from local event calendars and mapped to relevant locations and dates. Our competitor pipelines match products by brand, pack count, volume and vintage, and distinguish direct discounts from offers requiring a minimum quantity or spend. This lets Atlas compare products and promotions across retailers, even when they are described differently.

The pipeline is engineered to be fast-to-refresh and fast-to-query.

Fast-to-refresh: Parallel processing and shared calculations let us prepare over 1.25 billion transaction records across thousands of stores and tens of thousands of products in less than 1 hour. This supports daily refreshes, giving Atlas fresh data each day, and making the latest data available prior to 7am meetings.
Fast-to-query: Precomputed retail comparisons and tables organised around dates, categories and products help Atlas answer questions in seconds rather than hours, while preserving access to individual transactions for deeper investigation.

Takeaway:

Goodwork uses swarms of agents to bring in any external data at scale, for a tenth of the usual cost, and connects it to a gold-standard view built from 30+ internal data types that perfectly replicates finance’s reports.

Retail specificity and accuracy

“Context is not implicit in AI tools. It must be deliberately constructed by combining workflow design, expert rules, structured data, and integrated decision systems”- HFS Research

The Atlas harness embeds retail expertise throughout its analytical process through the playbooks it follows, the tools it uses, the guardrails it employs and the benchmarks that measure its performance. This helps Atlas produce accurate, relevant analysis that retailers can use to make commercial decisions. The playbooks, analytical tools and evaluation criteria are built around the same retail definitions and data foundation, connecting how Atlas reasons, calculates and checks its answers.

Retail playbooks and methodologies.

Atlas’s harness contains a specialist retail encyclopedia: metric definitions, analytical methods and step-by-step playbooks for questions such as promotion effectiveness, range optimisation and customer switching. These prescribe the scope, comparisons, calculations and checks an analysis requires. For example, Atlas’s ranging playbook compares products within relevant peer groups, combines sales and profit productivity with growth and margin, then considers customer loyalty and basket role before recommending a deletion.

Purpose-built analytical tools.

Atlas assembles analysis from controlled building blocks, with defined inputs, calculations and validation rules. Custom built retail tools handle operations such as ranking products, comparing trading periods and analysing basket composition directly against the data foundation. For example, Bamboo, our proprietary data-processing library, extends those results through declared operations such as filtering, joining, grouping and ranking, without executing arbitrary Python. Our Machine Learning Engine lets Atlas orchestrate models, including decision-tree clustering, preparing data, assessing model reliability and translating results into retail insights. GGRAF, our proprietary visualisation library, turns the saved results into charts using specific retail templates that control layout, labels, annotations and number formatting.

The right context at the right time.

Atlas brings the relevant retail definitions, methodologies and guidance into each stage of an analysis. Built-in controls tie analytical tools to the knowledge needed to use them correctly, helping Atlas apply consistent methods while keeping its attention on the question at hand.

Optimised against a retail benchmark.

Atlas has been optimised for accuracy against our proprietary retail benchmark - a set of commercial tasks real retail professionals actually face: diagnosing category declines, assessing promotional uplift, comparing new product launches, identifying delisting candidates and prioritising opportunities across range, price and margin. Answers are graded on task alignment, analytical validity and rigour, completeness, communication and traceability, and decision usefulness. Over three years, we have used these benchmark results to refine and improve Atlas’s performance to a level suitable for commercial decision-making.

Takeaway:

Goodwork has spent three years building a unified, AI-native, retail analytics engin (patent pending). It comprises of 500+ inter-connected retail specific skills, LLM native toolings, calculations, workflow and instructions which collectively encoding the retail reasoning and giving the harness its retail specificity.

Speed and ease of use

Time is a category manager’s most precious resource. Atlas is built to save it, with fast answers, clear explanations and results that are easy to understand and use.

Intelligent routing.

Atlas uses specialist agents designed for different levels of analytical complexity. Custom classifiers assess each question in its conversational context and route it to the most appropriate agent. Straightforward lookups used agents optimised for speed, while diagnostic and interpretive questions are assigned to agents capable of deeper reasoning and more advanced analysis.

Clear explanations.

Atlas is trained to lead with the commercial answer, explain its assumptions and focus attention on the few findings that matter. Its proprietary plotting library, GGRAF, turns analytical results into consulting-grade charts using purpose-built templates for trends, comparisons, bridges, distributions and other retail patterns. These templates control chart selection, layout, labels, annotations and number formatting to maximise information content while minimising cognitive load.

Prebuilt retail workflows.

Repeatable retail tasks, prescribed workflows that combine the right data, analytical tools, calculations and output structure, are encoded directly into Atlas. Atlas can automatically execute the relevant steps and produce consistent commentary for activities such as weekly performance reviews, supplier meetings, promotion analysis and range reviews. This embeds Atlas directly into the retailer’s operating rhythm.

Atlas Canvas.

Atlas allows users to creates and edit persistent documents and dashboards, turning conversations into reports that users can revisit, refine and update, for example, a weekly trading review or supplier briefing.

Software connectivity.

Atlas includes a secure file layer that can ingest Excel, CSV, PDF and other supporting files within the user’s conversation. It can inspect workbook structures, read selected sheets and cell ranges, combine uploaded information with governed retail data, and return results as charts, tables or downloadable Excel and CSV files. This lets users work with the files they already have without manually rebuilding their analysis inside Atlas.

24/7 availability.

Atlas runs as a persistent cloud service connected to prepared analytical data, rather than depending on an analyst to receive and complete each request. Health and readiness checks monitor the service, while conversation state preserves the work already completed. Combined with daily data refreshes, this means users can continue an investigation whenever a question arises and work from the latest available trading data.

Takeaway:

Optimisation happens at an individual question level, using multi-agent routing and predictive optimisation.

Engineered for trust

Retailers need confidence in both the answers Atlas gives and the boundaries it operates within. We build that trust through controlled data access, checks on the underlying numbers, visibility into how answers are produced, and ongoing oversight.

Controlled data access.

Atlas uses a dedicated managed identity with read-only permissions for explicitly approved tables. The data foundation enforces these limits, restricting what Atlas can read and preventing it from modifying or deleting the underlying data.

Separated customer environments.

Each retailer’s analytical data sits within client-specific catalogs and storage locations, with dedicated service identities and explicit access permissions. Development, staging and production use separate configurations and controlled access boundaries.

Traceable analysis.

Atlas retains all questions, answers, tool calls, inputs and results, giving authorised reviewers a record of the work behind an answer. These records are inspected and exported to investigate unexpected results and assess analytical quality.

Controlled oversight and sharing.

Authorised administrators can review conversations to investigate issues and support users. Selected conversations can be shared as snapshots, making useful analysis available to others without exposing subsequent messages.

Checks on the underlying numbers.

Automated validation reconciles ingested sales against supplied control totals and flags unexpected changes in data volumes. These checks help identify incomplete or inconsistent data before teams rely on it for decisions.

Content safety.

Atlas combines customised safety classifies with enforced access controls and strict monitoring to resist and identify attempts to override its instructions, bypass restrictions or access unauthorised data.

Enterprise agreements with AI providers.

Atlas uses leading AI models under enterprise agreements that guarantee data residency and zero data retention, and commit that customer data is never used for model training. Each retailer’s commercial information stays theirs.

Takeaway:

Atlas is built to enterprise-grade security standards from the ground up. It can only read the data it’s approved to see, it can’t change or delete anything, each retailer’s data is kept fully separate, and every answer leaves a complete audit trail. Security is engineered in, not bolted on.

Built to evolve and scale

Retail knowledge should accumulate, and multiple AI agents should share that knowledge. Atlas is designed to learn from continued use and work with specialist agents across the business.

Atlas Dreaming:

Every Atlas conversation captures knowledge that retail data alone misses: the reasoning behind a delisting, a supplier’s commitments or the lessons from a promotion. Atlas Dreaming turns these insights into continual learning in token space, helping Atlas improve through accumulated context. Nightly, swarms of AI agents extract, organise and consolidate this knowledge into two complementary layers of memory:

Personal memory: An indexed OKF second brain holds each user’s preferences, responsibilities, active work and saved analytical recipes. Nightly, specialist agent swarms extract useful context, remove noise, merge duplicates and update existing memories. A category manager returning to a supplier review can pick up with their preferred format, previous findings and open questions already available.
Company brain. The same process connects lessons across authorised conversations with share “context containers” (category plans, range strategies, supplier agreements and promotion reviews). It captures decisions, their rationale and outcomes, linking them back to company strategy while distinguishing agreed policy from individual suggestions. Lessons from range reviews, promotional trials and supplier negotiations become shared knowledge, and enable continual improvement from the organisation.

Interoperability by design.

Within defined access permissions, Atlas is designed to share context and coordinate tasks (e.g., passing category findings to a replenishment agent or drawing on a supplier-management agent’s record of commercial commitments). As the organisation’s AI capabilities grow, its accumulated retail knowledge travels with them.

Takeaway:

Goodwork goes beyond a simple chat memory - the moment it turns on, Atlas learns from every question from every user to form an ever involving organisational memory that encompasses mistakes, learning, goals and context.

The highest performing AI Agent for the Retail sector

Thesis introduction

Key problems we're addressing

Data takes work.

Coverage is incomplete.

Spreadsheets are unreliable.

Dashboards are rigid.

The harder questions need specialist help.

But finance is already stretched.

Analysis is inconsistent.

Trust breaks down.

And eventually, people stop asking.

Data coverage and scale

Internal data is consolidated:

Swarms of AI agents clean and harmonise the data:

Calibration:

External data is added:

The pipeline is engineered to be fast-to-refresh and fast-to-query.

Takeaway:

Retail specificity and accuracy

Retail playbooks and methodologies.

Purpose-built analytical tools.

The right context at the right time.

Optimised against a retail benchmark.

Takeaway:

Speed and ease of use

Intelligent routing.

Clear explanations.

Prebuilt retail workflows.

Atlas Canvas.

Software connectivity.

24/7 availability.

Takeaway:

Engineered for trust

Controlled data access.

Separated customer environments.

Traceable analysis.

Controlled oversight and sharing.

Checks on the underlying numbers.

Content safety.

Enterprise agreements with AI providers.

Takeaway:

Built to evolve and scale

Atlas Dreaming:

Interoperability by design.

Takeaway:

Lorum ipsum dollar espore