Automate a process, then improve it every version
Continuous improvement with AI means running each automated process as a versioned product: one owner, a process card, a unit cost, and a weekly release that adds reviewed human corrections to the next version.
This playbook gives seven steps, with cases from Foundation Risk Partners and C.H. Robinson.
This playbook belongs to Agents, traceability and evals and Models and routing.
Why now
C.H. Robinson says its Lean AI approach removes waste first, then puts custom agents on the remaining manual steps. Ocean and air quotes that took one to three days now take two to three minutes (26 October 2026). Its annual token cost runs at under $2M (Q3 2026 earnings call, 28 October 2026).
Foundation Risk Partners, a US insurance brokerage with 3,000+ employees and a Partners Group portfolio company, built two agentic systems on a cleaned policy-level dataset with a seven-person team augmented by Version 1. Partners Group says new-client policy processing time fell 94%, close rates doubled and EBITDA margin rose 120 bps, worth USD 10M. The first use case went from concept to production in 14 weeks (August 2026).
CI&T, citing Forrester, calls the gap the process chasm: task copilots bring efficiency gains on an operating model that stays the same, while reinventing a whole end-to-end process changes the business case (August 2026).
94%
shorter new-client policy processing (Foundation Risk Partners)
USD 10M
EBITDA uplift from two agentic systems (Foundation Risk Partners)
2–3 min
per ocean or air quote, from 1–3 days (C.H. Robinson)
Choose one end-to-end process with a unit cost
Pick a process with volume and a natural unit: a quote, a policy, an invoice, a claim, a purchase order. Measure today’s cost and cycle time per unit before anything changes. Every later version is compared with this baseline.
Genpact now prices accounts-payable work per invoice and offers to take a client from $5 to a guaranteed $3 per invoice (October 2026). Citi keeps a named list of 100+ end-to-end processes and reviews progress on each every week (Q3 2026 earnings call).
The step produces one process, its unit, and the current cost and cycle time per unit.
| Change | Example | Approved by |
|---|---|---|
| Prompt or instructions | Clearer extraction rule for a policy field | Process owner |
| Rule | A reviewed agent decision promoted to a deterministic rule | Process owner |
| Tool or data source | New read access to the claims system | Process owner and system owner |
| Model | Swap to a cheaper model after replay | Process owner and engineering lead |
| Human gate or decision right | Remove a manual check after a sustained agreement rate | Executive sponsor |
Remove waste before adding agents
Map the process as it runs, from the event logs where possible. Delete steps nobody uses, merge handoffs, and drop approvals that never reject anything. Every step you remove is a step no agent has to perform or a person has to check.
C.H. Robinson’s Lean AI discipline removes waste with lean methods first, then puts custom agents on the remaining manual steps, workflow by workflow, measured by cost to serve (Q3 2026 earnings call, October 2026).
The step produces a shorter process map with the removed steps listed.
“…we have turned a clean dataset into USD 10 million EBITDA uplift,” said Partners Group on Foundation Risk Partners, August 2026.
Keep rule-based steps in code, without a model
Mark each remaining step as a rule, an agent or a human gate. Use rules for the routine cases, which are usually the majority, and agents for the cases that need interpretation. Put people at the gates where judgment or accountability is required.
Meta’s privacy classification system resolves about 85% of traffic with versioned rules in milliseconds. The LLM handles the remaining 15% at roughly 400 times the compute cost, and a nightly loop distils new rules from reviewed LLM decisions, so the expensive share keeps shrinking (June 2026). Stripe alternates LLM steps with mandatory system gates for lint, tests and commits (Port.io, September 2026).
The step produces a process design that marks every step as rule, agent or human gate.
Write the process card and give it a version
The process card is the single source of truth for the process: owner, version number, inputs, steps, gates, metrics, open issues and change log. Agents work from the current version of the card, and people change it only by publishing a new version.
How often a process can be released limits how often it can improve. Westpac says the time to put a new agent capability into production in its lending flow fell from about six months to four to six weeks (August 2026).
Keep the card short enough to read in five minutes. When the card and the running system disagree, the card is the reference and the system is fixed in the next release. Auditors, new team members and agents all read the same page.
The step produces process card version 1.0, owned by one named person.
Capture every human correction
At each human gate, record what the person changed and why, in a structured field. Corrections made in email or chat are not recorded anywhere the next version can use. Each logged correction goes into the backlog for the next version and into the test set for every version after it.
DoorDash moved its quality signal for the Ask DoorDash assistant from an average of one employee feedback report to 2,000 auto-graded sessions a day. A manual regression test that took more than six hours now runs in about 20 minutes (June 2026).
The step produces a correction log and a growing replay set of real cases.
Release a new version every week through replay
Once a week, group the corrections, find the cause, and change the prompt, rule, tool or model. Replay the new version against the historical cases. Promote it only if quality holds and unit cost does not rise.
DoorDash runs changes through an offline evaluation harness, with simulated users generating test sessions, before they reach customers. The same harness validated a base-model swap that cut latency 35% without quality loss (June 2026).
The step produces a weekly release with a change log entry and replay results.
Track cost per successful unit
Divide the total cost of all attempts, failures included, by the number of units that pass the quality gate (Braintrust, October 2026). A model that is cheaper per token can cost more per finished unit.
Uber tracks cost per session, per 1,000 requests and per code review next to quality scores. With the model held constant, its cost per session fell 52% from the June peak (August 2026). C.H. Robinson says its missed-pickup agents automate 95% of checks and save 350+ hours a day (September 2026).
The step produces a monthly unit-economics line for each process in the portfolio.
Do this week
- Pick one end-to-end process and write down its unit, volume, cost and cycle time per unit.
- Name one owner for that process who can approve a release.
- Add a structured “what I changed and why” field at the main human review step.
Measure
| Metric | How to measure | Target direction |
|---|---|---|
| Cost per successful unit | All model, tool and review cost divided by units that pass the quality gate | Down |
| Cycle time per unit | Median time from request to completed unit | Down |
| Human correction rate | Share of units a person edits at the gate | Down |
| Release cadence | Versions promoted per month after replay | Up |
| Share handled by rules | Units resolved by deterministic rules without an agent call | Up |
Questions
What is continuous improvement with AI?
It is running an automated process as a product with versions. Agents and rules do the work, people review at defined gates, and every correction is logged. Each week the owner turns the corrections into a new version, replays it against past cases, and promotes it only if quality holds and cost per unit stays flat or falls.
How is this different from earlier automation projects?
Earlier automation was built once and maintained when it broke. Agents handle the ambiguous cases rules could not, and the weekly loop improves the process from its own corrections.
How often should a process get a new version?
Weekly is a workable default once the replay set exists, because corrections accumulate fast at volume. What limits the pace is how fast a new version can reach production. Westpac cut the time to put a new agent capability into production from about six months to four to six weeks, which is what makes frequent versions practical.
Who should own an automated process?
One named person who runs the business outcome, usually the head of the function, with an engineering lead. They own the process card, approve releases and answer for cost per unit. Changes to human gates or decision rights go one level up to the executive sponsor.
