Method · Post 4 of 6
Going fully agentic, one value stream at a time
Company-wide AI programmes stall because they start from tools. Start from the work: one stream from request to outcome, agents on it with the people who run it, then the next.
Most companies' AI programmes look the same eighteen months in. Every department has a chatbot. Licences have been bought. A few people are dramatically more productive, and no process has changed. The org chart got AI; the work did not.
The reason is where they started. Tools are bought per department, and the work crosses departments. A claim moves from intake to triage to an adjuster to payments to the ledger. Make the adjuster twice as fast and the claim still waits at every hand-off.
Start from a value stream
A value stream is the work from a customer's request to its outcome: claim to payout, order to cash, issue to resolution. It has a start, an end, a measure the customer feels (days to payout, first-contact resolution), and a sequence of steps that each have an owner, a system and a decision.
Going agentic one stream at a time means changing a whole flow of work, end to end, and measuring the outcome the customer sees.
The method
- Pick the stream. The one with the most hand-offs and the clearest outcome measure. Not the most visible one, and not the easiest demo.
- Map it in a week. Walk the stream with the people who run it. For every step: who does it, in which system, what they decide. Mark each step with one of the four states below.
- Build with the team. A business applications engineer builds the first agentic app with the front-line team on Kilter: a working app in two to four days, refined with them daily.
- Put it on rails. Before an agent acts on anything that matters (money, customers, records), Airbender ARC decides whether it may, and records the decision as evidence.
- Measure, then move. Measure the stream's outcome. Move the next step. When the stream runs, pick the next one.
| State | Who acts | Who decides |
|---|---|---|
| Manual | A person | A person |
| Assisted | A person, using AI | A person |
| Agentic | Agents | A person approves |
| On rails | Agents | Rails decide; people handle exceptions |
Stream two is cheaper than stream one
The first stream pays for the platform: identity for people and agents, the data layer, the policies, the portal departments reach their apps through. The second stream reuses all of it. Each stream makes the next one faster, and none of them adds a vendor.
What it looks like after a year
A board of every value stream in the company, step by step, showing what agents do and what rails govern. It is the transformation plan, the progress report and the risk register in one view, and it stays true, because it is drawn from what is running.
Leave the org chart alone. Change the work, one stream at a time.
Next step