
Web data taught AI to talk. AI agents need to learn how work actually gets done: the hand-offs, exceptions and outcomes recorded in business systems. Here is why operational data is the missing ingredient and what makes it valuable.
Ask a chatbot to explain how a plumbing company handles an emergency call and you will get a fluent, plausible answer. Ask an AI agent to actually handle one, routing the technician, updating the customer, rescheduling the 2 p.m. job and invoicing correctly, and the gap between knowing about work and doing work becomes obvious. Closing that gap takes a different kind of data.
By OpenDataPolicies Editorial · October 10, 2026 · 6 min read
The short answer
Web data is mostly finished, public content: articles, forum posts, documentation, books. It teaches a model language, facts and how people describe things. Operational data is the trail of real work inside a business: emails, chat threads, tickets, CRM stages, project tasks, transactions and call logs. It shows sequence, decisions, exceptions and outcomes.
Agents, AI systems that take multi-step actions on behalf of a user, need the second kind. They have to learn what happens next, who gets involved, what goes wrong, and what "done" looks like. That information rarely appears on the public web.
Why the industry is moving from chat to agents
The direction of travel has been stated openly. At NeurIPS in December 2024, OpenAI co-founder Ilya Sutskever said the industry had reached "peak data" and predicted that next-generation models would be "agentic in real ways," according to The Verge.
Those two statements belong together. If the supply of public text is close to exhausted, and the next frontier is systems that do things, then the most valuable new data is data that shows how things get done. A model can read a million articles about project management and still have no idea how a real general contractor handles a late delivery of drywall.
What web data is good at, and where it stops
Web data is not worthless; it is the foundation of every large language model. It is excellent for:
- Grammar, vocabulary and tone across many languages.
- General knowledge and public facts.
- Common explanations, tutorials and how-to articles.
- Public conversation, from forums and social platforms.
But web content has systematic blind spots:
- It is edited. A blog post about customer service describes the ideal process, not the messy one that actually happens.
- It lacks outcomes. You rarely see whether the advice worked, how long it took, or what it cost.
- It lacks sequence. Real work unfolds over days or weeks across many people and tools. A web page is a snapshot.
- It is shrinking as a free resource. The Consent in Crisis study found that within a single year, 28% or more of the most actively maintained, critical sources in the C4 training corpus had fully restricted AI crawling through robots.txt.
The result is a model that can describe work beautifully and still stumble when asked to perform it.
What operational data contains that the web does not
Operational data is the record a business leaves behind just by running. Each data category adds a different layer:
- [Email](/categories/email): how requests arrive, how quotes are negotiated, how disputes are resolved, and how long each step takes.
- [Team chat](/categories/team-chat): real-time hand-offs, escalations and the informal decisions that never make it into a formal system.
- [Support tickets](/categories/support): problem, diagnosis, resolution and satisfaction, often with categories and timestamps already attached. This is the raw material for service automation.
- [CRM](/categories/crm): how deals actually move through stages, which ones stall, and why.
- [Projects and tasks](/categories/projects): how work is planned, assigned, delayed and delivered.
- [Accounting](/categories/accounting): transaction patterns, cycles and categorisation that reveal the financial heartbeat of operations.
- [Industry software](/categories/industry-software): the deepest vertical records, such as repair orders, dispatch jobs, legal matters or construction RFIs.
Taken together, these give a model something it cannot get from the web: ground truth about cause and effect in real operations. A ticket that was escalated twice and resolved by a refund teaches an agent more about customer service than a hundred articles on best practice.
Why industry-specific workflows are worth more
General office workflows, such as scheduling a meeting or replying to an email, are relatively well understood by models already. The scarcest, most valuable data tends to be vertical: the workflows of a specific industry, with its own vocabulary, regulations and exceptions.
Consider a few examples from the industries buyers ask about most:
- Home services and trades: dispatch, estimates, call-backs and parts delays.
- Construction and engineering: RFIs, change orders and schedule slippage.
- Law firms: intake, conflict checks, matter workflows and billing patterns.
- IT services and MSPs: tickets, runbooks and incident resolution.
- Insurance agencies: quoting, servicing and claims-support workflows.
This is the heart of founder Jeff Cline's thesis: "Every industry is a geek away from being uberized." The "geek" building an agent for, say, auto dealership service departments needs examples of how real service departments run. The businesses that hold those examples hold the keys, while that data is still unique.
Isn't operational data full of personal information?
Raw business records do contain names, phone numbers and account details. That is exactly why they must be de-identified before anything is shared, and why buyers do not want the personal details in the first place. An agent needs to learn that a customer called back about a leaking valve three days after a repair, not who the customer was.
We never take HIPAA/PHI data or personal customer data. Everything is de-identified before it leaves your control.
De-identification typically removes or replaces direct identifiers (names, emails, phone numbers, addresses, account numbers) and generalises quasi-identifiers such as exact dates or locations, while keeping the structure of the workflow intact. Healthcare businesses can participate only on the administrative side. Our guide to de-identified data explains the methods.
What makes operational data valuable to a buyer
Not every inbox is a goldmine. Buyers generally look for:
- Volume and history. Years of records show patterns, seasonality and exceptions.
- Structure. Tickets with categories, CRM records with stages and projects with due dates are easier to learn from than free-floating files.
- Outcomes. Data that records how things ended (resolved, won, lost, paid) is more useful than data that only records how they started.
- Uniqueness. A niche vertical or an unusual workflow may be worth more than a larger but generic dataset.
- Clean rights and provenance. You need to be able to show the data is yours to license.
According to founder Jeff Cline, small companies have received over $100,000 for de-identified operational data, bids can arrive within 72 hours when the data is correct and a header sample is provided, multiple buyers can bid so you can take the highest offer, and deals typically close in about 90 days. Results vary; not a guarantee.
Our pillar on operational data for machine learning goes deeper on what buyers want.
A quick self-check: is your data agent-ready?
You do not need a data team to get a rough answer. Ask yourself:
- Do we have at least a few years of records in one or more core systems?
- Do our records show what happened next, such as status changes, replies, resolutions or payments?
- Is our work specialised, with vocabulary or steps an outsider would not know?
- Can we export record counts and column names from each system?
If you answered yes to most of these, your operational history is likely the kind of data model builders are looking for. The data readiness checklist turns these questions into concrete steps, and the header sample how-to shows how to share structure without sharing any records.
The bottom line
Web data built AI that can talk. Operational data will build AI that can work. The public web is largely mined; the workflow records inside millions of businesses are not. That makes this moment unusual: ordinary companies hold a resource that frontier model builders cannot get anywhere else, and the value of that resource is highest before the agents for their industry are trained.
To see where you stand, run your systems through the valuation calculator or go straight to a free bid request. There is no cost to ask, and you can say no to any offer.
Frequently asked questions
What is operational data?
Operational data is the record a business creates while working: emails, chat, support tickets, CRM records, project tasks, accounting transactions, call logs and industry software records. It shows how work is actually done.
Why do AI agents need operational data instead of web data?
Agents perform multi-step tasks, so they need to learn sequences, decisions, exceptions and outcomes. Web content is mostly edited, finished text and rarely shows how real work unfolds inside a business.
Does selling operational data mean sharing customer information?
No. We never take HIPAA/PHI data or personal customer data. Everything is de-identified before it leaves your control.
Which kinds of operational data are most valuable?
Structured records with outcomes and long history, such as support tickets, CRM pipelines and vertical industry software, tend to be most useful. Niche industries can be especially valuable because their workflows are less represented in existing models.
Sources
Updated November 19, 2026.
OpenDataPolicies.org connects businesses with buyers of de-identified operational data. We are not a law firm, broker-dealer or investment adviser, and nothing on this site is legal, tax or financial advice. Statements about amounts, timing and outcomes are Jeff Cline's statements based on past transactions. Results vary; nothing here is a guarantee of any bid, price, buyer or closing. Every deal depends on the data, its quality, its rights and buyer demand at the time.