AI that does the work, and shows its work.
Each station that needs a model runs on the Claude model that fits its job, at the autonomy level you set for each action, inside rules the server enforces. Every run leaves a record you can read, down to the tokens and the cost.
The right Claude model for each station.
Volume work runs on the fastest model, judgement on a stronger one, and quotes and bids on the most capable.
- Pinned per station, and its test set re-run before that ever changes
- At most one narrow call per decision, with a fixed output that code validates
- Every call logged against its run, with the model, tokens and cost
- 20 Claude Haiku 5.5
Fast and economical: reading, sorting, extracting and routine replies, at volume.
- 11 Claude Sonnet 5.5
Judgement: ranking, matching, planning and anything that weighs several records.
- 2 Claude Opus 5.5
The most capable: priced quotes from drawings and scopes, and RFP responses.
Code where it must be exact. AI where it takes judgement.
A language model is brilliant at reading a supplier's messy PDF. It's the wrong tool to add up the invoice, so it never does. Every decision splits the work three ways.
- Code
Exact, every time
Anything with one right answer is plain code, never a model.
- Totals, tax, rates and dates
- Who is allowed to do what
- Interlocks: a non-compliant vendor, a bill over its order
- Posting to the ledger
- AI
Judgement, narrowly
The model answers one narrow question per decision, in a fixed format.
- Reading messy emails, PDFs and photos, in any language
- Classifying and routing
- Matching a bill to the right order
- Drafting replies, quotes and bids
- People
Accountable, always
A person decides what runs on its own, and is called when anything is off.
- Approve at Suggest, check at Review
- Promote an action on evidence
- Take the andon when the line stops
- Switch any agent off
The model reads it. The rule stops it.
Code gathers the facts, the model answers one narrow question, code checks the answer against your records, and the action's level decides whether a person approves.
- 01A supplier's bill arrives as a PDF.
- 02The model reads it into fixed fields: supplier, order, lines, total.
- 03Code matches it to the order and checks the total against it.
- 04It's over the order, so the line stops and Accounts gets it, with the reason.
- Code gathers the facts Code
- AI judges one narrow question AI
- Code checks the answer Code
- The level decides who acts Person
Which model runs which station.
Every station, routed to the model pinned to its job.
-
20 stations Claude Haiku 5.5
Fast and economical: reading, sorting, extracting and routine replies, at volume.
- RFPs & tendersTender finder
- RFPs & tendersBid tracker
- SalesQuote follow-up
- Service DeskTriage
- Service DeskClient updates
- PurchasingWarranty claims
- FinanceCollections
- FinanceSLA reports
- Suppliers & subcontractorsCompliance
- Jobs & DispatchVisit prep
- Sites & EquipmentImport cleanup
- CertificationCloseout
- ProductionShop-floor follow-up
- QualityInspection
- FulfilmentOrder fulfilment
- FulfilmentShipping
- Returns & claimsReturns
- ContractsPlanned maintenance
- ContractsRenewals
- Projects & timeTimesheets
-
11 stations Claude Sonnet 5.5
Judgement: ranking, matching, planning and anything that weighs several records.
- RFPs & tendersBid qualification
- PurchasingMaterials
- FinanceBilling
- PeopleStaff assistant
- Jobs & DispatchDispatch
- Jobs & DispatchChange orders
- RolloutsRollout coordinator
- ProductionProduction planner
- QualityNonconformance
- Returns & claimsFreight claims
- Projects & timeBudget watch
-
2 stations Claude Opus 5.5
The most capable: priced quotes from drawings and scopes, and RFP responses.
- RFPs & tendersResponse drafting
- SalesQuoting
Every run leaves a record you can read.
Who acted, why it started, what it read, which model, what it cost, who decided and what changed. Hover a part to see what it is.
Agent runRUN-50812
- Agent
- Billing · Finance
- Trigger
- Job SJ-2207 closed out
- Read
- Check-in/out, materials, rate card
- Model
- Claude Sonnet 5.5
- Usage
- 6,214 in · 402 out · $0.0165
- Level
- Suggest
- Decided by
- Accounts · approved
- Result
- Invoice INV-20419 posted
Example record · names and numbers invented
- Agent
The station that acted, and the line it sits on. Each agent is its own account with its own permissions.
- Trigger
The handoff that started the run: a record reaching a status, never a person's guess.
- Read
What the agent based its work on, linked to the records.
- Model
The model pinned to this station.
- Usage
Tokens and cost of every model call in the run.
- Level
The action's autonomy level. A ledger post never runs autonomously.
- Decided by
Who approved, changed or refused it, and when.
- Result
What changed in your records because of the run.
Autonomy, earned action by action.
Every action starts where you set it and moves up only when your own work shows it's ready.
- Set per action: Suggest, Review or Autonomous
- Promoted by a person, on evidence from your own work
- Drops back a level on its own if it starts going wrong
- A ledger post never runs autonomously
- Suggest Level 1
- Review Level 2
- Autonomous Level 3
Built to stop safely.
Jidoka is the Toyota principle of automation with a human touch: the machine runs on its own until something is off, then stops and calls a person. Every agent works the same way, inside seven layers of guardrails.
- 01
Input
Incoming email is untrusted. The model gets fixed instructions and must answer in a fixed format.
- 02
Model
Pinned per station, one narrow call per decision, and its test set re-run before it ever changes.
- 03
Validation
Code checks every answer against your records before anything happens.
- 04
Rules
Interlocks enforced on the server. No agent, and no prompt, can get around them.
- 05
Autonomy
Set per action. Money never moves on its own: a ledger post never runs autonomously.
- 06
Runtime
Each agent has its own identity, a daily spend cap, and an off switch of its own.
- 07
Audit
Every run is recorded: what it read, which model, what it cost, and who decided.
Bring one process that slows you down.
We'll walk it through the line with you: what comes in, what each station would do, and which of its actions earn autonomy first.