Three ways we put AI to work. Only one of them needs a model in the loop.
Most engagements begin in one of these three areas. Many companies need two, because agents are only as good as the knowledge they can reach.
Agentic transformation
Agents that handle a workflow end to end across your systems, with approval steps where the risk sits. We start with one process and prove it in production before we widen the scope.
Best for. Claims, order exceptions, onboarding, supplier queries. Many steps, many systems, some judgement.
First phase. 8 to 12 weeks to a live agent on one process, with monitoring and an evaluation set.
You keep. The agent, the eval harness, the runbooks. Nothing licensed.
Second brains
A knowledge layer your team can question directly, and that your agents can call. Every answer shows its source. Permissions follow the groups you already run.
Best for. Policies, contracts, SOPs, project history. The questions two busy people keep answering.
First phase. 6 to 10 weeks to a scored baseline on questions your team actually asks.
You keep. The index, the question set, the accuracy scorecard.
Deterministic automation
Rule-based automation for work that has to be right every time, with no model in the path. Same input, same output, and an audit trail on every run.
Best for. Billing, rebates, reconciliations, compliance checks, regulated reporting.
First phase. 4 to 8 weeks, tested by replaying your real history through the rules.
You keep. Rules your own people can read and change.
Which one do you need?
If people cannot get reliable answers out of your own documents, start with a second brain. If a workflow eats hours and needs judgement across several systems, start with agents. If the work has to be exact and explainable every single time, start with deterministic automation. Most companies find they need a second brain underneath the agents, because an agent without good context is just a confident guess.
Work that must be exact is still done by hand because nobody trusts automation with it.
Every engagement includes
Whichever practice you start in, the same things are in the box. They are what separates software that runs from a demo that impressed people once.
A scope agreed up front
One process, a defined outcome, a fixed first phase. You know what you are buying.
An evaluation set from day one
We measure whether it works rather than asserting it. The test set is yours.
Built on your real data
No sandbox demos. Your systems and your permissions from the first week.
Monitoring and escalation
Visibility of every run and a clear path for the cases a person should see.
Documentation and runbooks
Written decisions and operating guides so the system outlives the project.
A named owner on your side
Someone who will run it afterwards, involved from the start, not handed a result.