Data collection

To work at scale, AI agents need labeled datasets of how your company actually works.

One for every process, every team, every function. 8Flow builds them from the work your people already do, and you own them.

Where it starts · one record

Every record has three parts.

A record is one work item, start to finish: a support ticket, a sales follow-up, an operations task. This one is Maria’s refund. Collect enough of them and you have something an agent can learn from.

  1. 1

    The steps. What she did, in order, with timings.

  2. 2

    The screens. What was in front of her at each step.

  3. 3

    The labels. Why it belongs in the dataset.

And it is yours, not ours.

Every record lands in a private bucket in your cloud, and works with whatever models you choose. The learnings compound for you, not for a model vendor.

Refund · ticket #7431 · Maria

Collected
0:00Ticket openedHelpdesk
1:12Order foundOrders
2:30Policy checkedPolicy
3:45Refund issuedBilling
5:41Reply sentEmail

The order of apps, the time in each, and what was carried between them by hand. This is the part 8Flow sees on its own.

What a dataset is

One selection. Hundreds of records.

Function, team, workflow type, segment. Each selection holds hundreds of records, and changing any one gives you a different dataset.

One of hundreds of datasetsNext selection…
Functionwhere the work livesSupportSalesEngineeringFinanceOperationsHR
Teamwho runs itL1L2BillingEscalations
Workflow typewhat they run, over and overRefundsReset MFAChargebacksCancellations
Segmentthe slice you care aboutOver $200Repeat customerFirst tryEscalated
A datasetSupportL2RefundsOver $200312 records, start to finish

Every nuance · inside one dataset

Work never runs the same way twice. 8Flow maps every way it runs.

Open one dataset and the nuances are the whole story. Every variation is broken out from start to end, with why it happened and what an agent has to learn to handle it. This is the part nobody writes down.

Refunds workflow

2,847 records

Hundreds of paths, grouped into common patterns

Common patterns

2,304the standard path81% of the month, start to end

Why it happened

Order attached, under $200, first refund. Five steps, the way the team designed it.

What the agent learns

What a clean record looks like, end to end. The baseline everything else deviates from.

How agents are graded

An agent’s record counts as clean only if it matches this shape: five steps, no detours.

The honest part · where this goes

Every pitch in the market ends on autonomy. Nobody shows the middle.

People still do the work, and no one ever wrote down how. That is why the unglamorous step is the whole game.

Today

We are here

People do the work. Now it gets collected.

Every work item worth learning from becomes a record you own. The step pays for itself.

Next

The work gets redesigned around agents.

Models read your records and propose a shape the work never had. Where a person still belongs, the records answer.

Eventually

Agents run the routine. People decide.

Nobody is here yet. Jumping straight to it means guessing at how the work is done.