02 · Connected SI
SI on your content, with a person still in charge.
System Intelligence (SI) drafts from your approved content. People set the sources, the boundaries, and what is allowed to ship.
- 01Grounded in approved content
- 02Humans decide what may ship
- 03Guardrails kept current
The gap
Answers with no source, and no one accountable.
A general chat tool invents prices, ignores the catalogue, and has never seen your policies. Staff paste live customer data into tools nobody approved. The draft is fast. The business cannot stand behind it.
Connected SI is useful when the sources are approved, the model is allowed to refuse, and a named person reviews what reaches a customer or a live system.
How the system connects
Assess · Build · Run
- 01AssessMap the current website, software and tools. Written gaps, priorities and a plan to approve.
- 02BuildImplement against scope, milestones and acceptance criteria.
- 03RunOperate, fix and extend the system after go-live.
Outcomes
What changes when this is in place.
- 01
Answers from your content
Assistants retrieve from the catalogue, policies, and records you approve — and say when that material does not cover the question.
- 02
A person still decides
Humans set the outcome, review the draft, and approve what ships. The model does not change a live system on its own.
- 03
Boundaries that hold
Permissions, refusal behaviour, and logging are part of the feature, so sensitive data stays in the tools you chose.
- 04
Quality you can revisit
Evaluation before go-live, and a Run plan for when content, policies, or models change.
Capabilities
What the work covers.
This is Human-Led SI. People decide what to build and what may ship. SI drafts and retrieves. The same people stay responsible after it is live.
- 01
Approved sources
The documents, catalogue fields, and records the assistant may use — named, permissioned, and kept current.
- 02
Retrieval before invention
Answers grounded in those sources, with a refusal when the source does not support a claim.
- 03
Human review
A named role approves customer-facing output and any action that writes back to a system.
- 04
Tool use with limits
Links into CRM, search, or internal tools only for the actions Assess allows, with an audit of what ran.
- 05
Evaluation
A test set of real questions before go-live, including the ones the assistant must decline.
In the programme
Assess, build, then stay.
01
Assess
Find where SI saves time without risking a wrong price, a policy breach, or a write to a live system. Sources and the human approver are named here.
02
Build
Ship grounded features with acceptance tests: correct answers, refusals, and the review step a person actually performs.
03
Run
Monitor quality, refresh sources, and keep guardrails current as the catalogue and the model change. People remain in charge.
What’s included
The delivery, in writing.
- A written source list and permission model
- Assistant or feature grounded in that approved content
- Refusal behaviour when the source does not cover the question
- Human review on anything customer-facing or system-changing
- Evaluation set run before go-live
- Logging of sources used and actions taken
- Documentation you own, plus a Run plan as content changes
Outside this service. A public chatbot on the open web, or sending your content out to train public models.
What keeps working
Useful after the model and the catalogue both change.
Connected SI stays trustworthy when sources, review, and refusal are operated — not frozen on the day of the demo.
Sources stay approved
New policies and catalogue updates enter the source list on purpose. Stale documents are removed.
Review stays human
The approver is a role in the process. Monthly care keeps that review real as volume grows.
Your content stays yours
The design uses your approved sources and controls. It is specified so your data is not used to train public models.
FAQ
Practical questions.
Who is in control of the result?
People are. They choose the outcome, the sources, and what is allowed to reach a customer or a live system. SI speeds the draft and the lookup. A named person approves, edits, or stops it.
Do you fine-tune models?
When Assess shows a need. Retrieval over approved content is often enough, and it is easier to keep current when the catalogue changes.
Is our data used to train public models?
No. The service is designed around your approved sources and controls, so your content is not contributed to public model training.
Who is accountable after launch?
The same team, under monthly care. They refresh sources, watch answer quality, and keep the human review in place.
Discuss Connected SI
Tell us which service the operation needs. We map it into Assess, Build, and Run, and stay responsible after launch.