Approach
Assess, design, build, operate.
Four phases, in order, with a decision point at the end of each one. You can stop after any phase and keep everything produced up to that point. Durations below are typical ranges, not commitments; the scope you receive after assess carries the dates for your build.
01Phase 01
Assess
Typically one to three weeks
What happens
- We sit with the people who run the process today and watch the actual steps, including the exceptions nobody documented.
- We measure current volume, handling time, error rate, and where work waits on a human.
- We separate the parts a model handles well from the parts that need deterministic code or a person.
- We size the cost of the current process and the cost of running the replacement, including inference.
What you receive
- A written process map with the automation boundary drawn on it
- A scope with a fixed price and a build sequence
- A go or no-go recommendation, including the case for not building
02Phase 02
Design
Typically one to two weeks
What happens
- We choose the model or models per task and document why, including the fallback when the first choice is unavailable.
- We design the data path: what leaves your systems, what is retained, and what never goes to a provider.
- We define the evaluation set before any code exists, using real historical cases with known correct outcomes.
- We specify the human review points, the escalation rules, and what the system does when it is unsure.
What you receive
- A system design document covering architecture, data flow, and model routing
- An evaluation set with pass thresholds you agree to in advance
- Integration specifications for each system we touch
03Phase 03
Build
Typically four to twelve weeks depending on scope
What happens
- We build in short increments against your real data, not synthetic samples.
- Every increment runs through the evaluation set, so quality is a number rather than an opinion.
- We integrate with your systems through their supported interfaces and handle credentials through your own secret management.
- We run the system in parallel with the current process before it takes over, so you can compare outputs side by side.
What you receive
- A working system in your environment with source code and documentation
- Evaluation results against the thresholds set in design
- Runbooks and training for the staff who own the process
04Phase 04
Operate
Ongoing, month to month
What happens
- We monitor output quality continuously against the same evaluation set, so drift shows up as a failing test rather than a customer complaint.
- We track cost per run and per outcome, and tune routing when a cheaper model clears the bar or a task needs a stronger one.
- We handle provider change: deprecated models, new versions, pricing shifts, and rate limit changes.
- We review exception volume with your team and fold recurring exceptions back into the system.
- We publish a monthly report covering volume, quality, cost, and what changed.
What you receive
- Continuous monitoring with alerting on quality and cost thresholds
- A monthly operations report
- A named engineer who knows your system
05Why operate matters
The phase most firms will not sell you
An AI system is not a website. It does not sit still. The provider deprecates the model you built on. A prompt that scored well in March starts missing a document variant that arrived in June. An upstream API adds a field and your extraction silently drops it. None of that announces itself.
A project firm hands over the repository and closes the engagement. Six months later the system is quietly wrong and nobody owns finding out. Operate exists so that the evaluation set keeps running, the cost keeps getting checked, and someone who knows the system is accountable for it staying correct.
If you have an internal team that can carry this, we will train them and step back. Either outcome is fine. Leaving the question unanswered is not.
| Item | Cadence |
|---|---|
| Evaluation set run | Continuous |
| Quality and cost alerting | Continuous |
| Exception review with your team | Monthly |
| Model and provider change handling | As needed |
| Operations report | Monthly |
Start with assess. Decide after that.
The first phase ends with a scope, a price, and an honest recommendation, including the case for not building anything.