What counts as operational data, and why it is worth money

Most companies think of their data as the customer list, the product database or the analytics dashboard. The data AI labs pay for is something else: the everyday record of how your team actually got work done. If your company ran on chat, email, tickets and shared documents for a few years, you are probably sitting on it already.

Operational data is what accumulates as a side effect of people doing their jobs. Nobody sets out to create it. A support lead answers a ticket, an engineer opens a pull request that references that ticket, a product manager updates the roadmap doc, and someone posts in a channel that the fix is live.

Each of those actions leaves a record in a different tool. Taken together, they show a problem being noticed, discussed, assigned, worked on and resolved. That sequence, with all its detours and back-and-forth, is the part that matters.

This is different from the data most companies think of as an asset. A product database describes the state of things. Operational data describes the decisions and handoffs that produced that state.

In practice, operational data lives in the tools your team opened every morning. The categories we look at most often:

• Team chat, such as Slack or Microsoft Teams channels
• Email, especially threads between colleagues, vendors and customers
• Docs and wikis, including specs, runbooks, meeting notes and planning pages
• Tickets and projects in tools like Jira, Linear or Asana
• Code repositories and pull requests, with their review comments
• CRM records, including deal notes and activity history
• Customer support conversations and how they were resolved
• Finance workflows, such as approvals, invoicing and month-end close

Not every company has all of these, and that is fine. A company with three of them, well connected, can be more useful than one with eight that barely reference each other.

Some data is excluded by default because it is sensitive or simply not what buyers need. HR files, payroll, medical information and legally privileged folders are left out. Direct messages are excluded unless the company chooses to include them and has the right to do so.

Credentials, API keys, card numbers and government ID numbers are removed outright during anonymization, wherever they turn up. Archives tend to hold more of these than people expect, pasted into a chat thread or buried in a config file.

Then there is data that is real but carries less weight. A folder of exported PDFs with no surrounding conversation, or a marketing site's traffic reports, tells a model little about how work gets done. It is not worthless, but it is not where the value sits.

AI labs are building agents meant to do work inside companies: triage a support queue, take a ticket to a merged pull request, chase a late invoice, prepare a handoff. To train and evaluate those agents, they need examples of that work as it actually happens.

The public web is full of finished outcomes. It has documentation, articles and answers to questions. What it mostly lacks is the messy middle: the clarifying question in a thread, the ticket that got reopened twice, the email that changed the plan. Agents tend to struggle in exactly that middle, so that is where good training and evaluation data is scarcest.

Labs use licensed operational data in a few ways. It can become post-training material, the basis for reinforcement learning environments that simulate a company, or private evaluation sets that test whether an agent can handle realistic work it has never seen. In each case, the realism comes from the fact that real people did real work in those tools.

The data is anonymized before anything is delivered. Every person, company, address, email, phone number and account becomes a consistent token, so PERSON_014 is the same person in Slack, email and Jira. The structure of the work survives; the identities do not. The licence is for training and evaluation only, and buyers may never market to, profile, contact or re-identify anyone in the data.

The price of a licence is driven less by raw volume than by how well the data hangs together. A few factors matter most:

• Connection depth: whether the same project can be traced across several tools, from the chat where it started to the ticket, the code and the customer email
• Years of history: longer timelines show projects starting, stalling, changing hands and finishing
• Headcount: more people means more roles, handoffs and points of view
• Number of systems: each additional tool adds another angle on the same work

This is why we say connected data beats big data. Ten years of email on its own is less useful than three years of chat, tickets and code that refer to each other.

Typical licences range from about $20K to $400K and up. Where a given company lands depends on those factors, and there is no honest way to give a firm number before the data has been counted. If you want a rough sense first, the value estimator asks the same kinds of questions.

Start with a simple inventory. Which tools did your team use, for how many years, and with how many people? Did tickets link to pull requests, and did chat threads link to documents? Did customer emails turn into support cases that turned into fixes? Those answers will tell you most of what you need to know about whether your operational data is worth pursuing.

Imagine a 40-person agency that ran on Slack, Asana and Google Workspace for six years. It may never have thought of its project channels and client threads as an asset. To an AI lab building agents for project work, that archive is a detailed record of how a real team planned, delivered and fixed things over time.

A 15-minute call is enough to tell whether a full evaluation makes sense. If it does, the evaluation is read-only and takes under an hour of your time. Book an intro call to find out where your company stands.