How we built our research data
Every statistic and example on our industry, workflow and research pages comes from three datasets. This page explains where each one comes from, what we recorded, how we categorized it, and what it can and can't tell you.
Noesis research draws on three datasets. The main one holds 3,362 deduplicated, published case studies of operational automation and AI, each read at the publisher's own page where possible and categorized by industry, workflow, AI type and value driver. Alongside it, 25,262 operations job postings from Canada and the US show what companies hire people to do by hand, and 33,501 industry associations with 67,735 listed events show how much each industry is learning about AI. Case-study results are as reported by the organizations that published them.
The dataset at a glance
Case studies: where they come from
We screened case-study listings on case study database websites, which collect customer stories published by software vendors and service providers, across more than 1,100 software and service categories. For each selected case we followed the listing to the publisher's own web page or PDF and read it in full. 128 cases (4%) could only be read from the directory summary because the original was unavailable; they are marked as summary-only in our records.
Duplicates (the same story listed under several categories or titles) were merged, keeping every listing link. Each case keeps a stable ID (for example CS-1441) and the name of the organization that published it, which you'll see on every example we show, with a link to the publisher's own page where we have one.
Case studies: what qualifies
A case had to name the customer, describe a recognizable operational workflow, describe a technology intervention, and report an identifiable outcome. We included both companies adopting a tool and service firms implementing a solution for a client (162 cases are consulting- or partner-led).
We did not require the word "AI". Rule-based workflow automation, RPA, document capture, scheduling and optimization software count alongside machine learning, generative AI and agents, because the question we care about is how work changed. We excluded website redesigns, branding and advertising, infrastructure migrations with no process change, generic "digital transformation" stories, vague testimonials and anonymous customers.
Case studies: what we recorded
For each case: the customer, industry and size; the department and process; the problem, the trigger and the previous way of working; the technology, the new workflow and what people still do; integrations and time to implement; the buyer's job title; and reported outcomes for time, cost, revenue, productivity, cycle time, errors, headcount, customer experience and risk.
Facts come only from the source. If the source is silent, the field is recorded as not stated rather than estimated. We publish job titles, never personal names.
Case studies: how we categorize them
Industries: 25 industry labels, of which 20 have pages. Workflows: 38 standard workflow labels grouped into 18 workflow families (for example finance back-office, or logistics, dispatch and warehouse). Cases the standard labels didn't fit were assigned to a family using written keyword rules applied to the original workflow description; 50 cases remain uncategorized.
AI type describes the mechanism the source reports (for example workflow automation, document intelligence, classification and routing, forecasting, generative AI). When we say a case uses "AI beyond rule-based automation", we mean at least one AI type other than workflow automation or RPA. Value drivers (labor reduction, cycle-time reduction, capacity increase and so on) are analyst-assigned from the reported outcome.
Case studies: location
Country and region come from what the source states, or from the customer's name or title when it's explicit. Location is never guessed: 720 cases have no determinable location. 191 cases are from organizations in Canada, 51 of them in Ontario.
Case studies: what they can and can't tell you
These are published case studies, so they describe what vendors and service providers choose to publish. Counts show where documented implementations exist. They don't measure how common a problem is across the market. Results are self-reported by the publishing organizations and are not independently verified; some describe projections or customer estimates, and some sources are several years old.
We use the dataset to show patterns: which workflows get redesigned in which industries, what people still do afterwards, which systems are involved and what kinds of results get reported. We always state the number of cases behind a figure.
Job postings: what we collected
We collected 25,646 job postings for operational roles from Indeed and LinkedIn, posted in the 90 days before 7 October 2026 (most between July and October 2026). Canada came first, with Ontario and the Toronto area as the priority, plus a smaller US sample. Searches covered operational job titles in all 18 workflow families, from accounts payable clerks to dispatchers and claims processors, and left out senior leadership titles. Duplicate postings were merged.
384 postings came from a separate search for AI and automation terms. Because that search looked for AI on purpose, those postings are left out of every share we publish, which leaves 25,262 postings (19,304 in Canada).
Job postings: how we read them
Each posting is placed in one of our industries using the job board's industry label, the employer's label on other postings, or keywords, with a review pass on unclear employers. Postings we couldn't place are left out of industry figures. Workflow family comes from the job title.
We then look for specific phrases in the full description: data entry or manual input; following up or chasing for information; reconciling records; preparing reports; processing invoices, forms or other documents; scheduling appointments, jobs or deliveries; giving customers status updates; Excel or spreadsheets; named systems such as an ERP, CRM, TMS, WMS or applicant tracking system; and mentions of AI or automation tools. A figure counts the postings that mention a duty at least once.
We publish a share only when at least 50 postings sit behind it. We publish aggregate counts only, never posting text, employer names or anyone's contact details.
Limits: a posting shows what an employer asks for when hiring, which is a different thing from how people spend their time. Most Indeed postings leave the industry blank, French-language postings from Quebec are underrepresented, and the US figures come from a smaller sample weighted toward recent postings.
Industry associations: what we collected
The association dataset lists 33,501 industry and professional associations: US organizations from the IRS exempt-organization files and Form 990 filings, and Canadian organizations from the Corporations Canada federal not-for-profit dataset, each mapped to one of 20 industries. For every Canadian association and every US association with at least $250,000 in revenue, we found the website where possible and read its public event and education pages.
That produced 67,735 listed events dated from October 2024 onward: conferences, webinars, courses, workshops, meetings and networking events. Events about AI were identified and graded by depth, from a passing mention to a hands-on session.
Limits: members-only calendars and app-only event platforms can't be read, so activity counts are a floor. Associations incorporated only in Ontario are mostly missing, because the Canadian source covers federal corporations. About 2,400 priority associations have no website we could find.
Versions and citation
Snapshots: case studies 7 October 2026 (3,362 cases, method vault-stats-v1); job postings 7 October 2026 (method jobs-stats-v1); associations October 2026 (method associations-stats-v1). When a dataset is refreshed, figures on every page update and this page records the change.
To cite: Noesis AI, review of 3,362 published automation and AI case studies (snapshot October 2026), noesisai.tech/research/methodology/.
You may reuse and adapt the figures and categories from our research under the Creative Commons Attribution 4.0 license (CC BY 4.0), as long as you credit Noesis AI and link to this page. The case studies themselves belong to the organizations that published them.
Want to see where your workflows fit?
Bring one workflow to a free call and we'll compare it against what the case studies show.