04 / 09 · Capability
AI Systems & Assistants
Assistants that know your business, with the guardrails to be trusted.
What this is
Custom assistants trained on your documentation, support bots that deflect real tickets, RAG pipelines over your own document sets, and the governance to put them in front of customers without holding your breath.
Included under this pillar
- Custom assistants trained on your documentation
- Customer support bots
- Internal knowledge-base bots
- Sales assistants and proposal drafting
- RAG pipelines over document sets
- Agent builds with tool access
- MCP server development
- Prompt libraries and prompt engineering
- Documented AI SOPs for your staff
- Model and tool selection consulting
- AI usage policy and governance
- Staff training and enablement workshops
The process
How ai systems actually gets built.
| Step | Phase | What happens | You get | When |
|---|---|---|---|---|
| 01 | Use-Case Audit | Which tasks are genuinely suited to a model, and which are automation wearing a hat. | Scored use-case shortlist | Week 1 |
| 02 | Data & Guardrails | Source documents assembled and cleaned; refusal rules, escalation and scope limits agreed. | Indexed corpus and guardrail spec | Week 2 |
| 03 | Assistant Build | Retrieval, prompts, tool access and the handoff to a human, built and evaluated against a fixed test set. | Working assistant and eval scores | Weeks 3–5 |
| 04 | Red-Team | We attack it — prompt injection, data leakage, confident wrong answers — before your customers do. | Red-team report and fixes | Week 6 |
| 05 | Deploy | Into your channel: site, Slack, helpdesk or internal tool. Logging on from day one. | Live assistant and usage logging | Week 6 |
| 06 | Tune | Real conversations reviewed, gaps closed, eval set grown. Quality is maintained, not assumed. | Monthly eval report and tuning | Ongoing |
What you get
Everything below, every time.
0Deliverables
- 01Scored use-case shortlist
- 02Cleaned and indexed document corpus
- 03Assistant with tool access
- 04Written guardrail and escalation policy
- 05Evaluation set and baseline scores
- 06Red-team report
- 07Conversation logging and analytics
- 08Staff training and AI usage policy
How it is scoped
Scoped by number of assistants and the size of the corpus. Governance documentation is included, never an upsell.
Questions
Asked and answered.
Will it make things up?
Retrieval-grounded and evaluated against a fixed test set, with a written refusal path. We report the score rather than promising perfection.
Where does our data go?
Decided with you at the Data stage and written into the guardrail spec, including retention and which providers see what.
Which model?
Selected per use case on evidence and cost, and re-evaluated as models change. Model choice is a decision we document, not a brand we are loyal to.