'%3e%3ccircle%20cx='500'%20cy='530'%20r='315'%20fill='url(%23g)'/%3e%3cpath%20d='M707%20254c66-44%20121-103%20151-177-85%2033-147%2079-197%20141%2020%207%2035%2018%2046%2036Z'%20fill='url(%23g)'/%3e%3cpath%20d='M716%20310c73-14%20131-2%20178%2039-73%2022-133%2017-181-13%205-8%206-16%203-26Z'%20fill='url(%23g)'/%3e%3crect%20x='251'%20y='417'%20width='496'%20height='210'%20rx='105'%20fill='%23100A1F'/%3e%3ccircle%20cx='393'%20cy='522'%20r='34'%20fill='%23F9F7FF'/%3e%3ccircle%20cx='605'%20cy='522'%20r='34'%20fill='%23F9F7FF'/%3e%3ccircle%20cx='388'%20cy='520'%20r='10'%20fill='%23100A1F'%20opacity='0.35'/%3e%3ccircle%20cx='600'%20cy='520'%20r='10'%20fill='%23100A1F'%20opacity='0.35'/%3e%3cpath%20d='M265%20680c49%2088%20126%20133%20231%20136%20107%204%20190-38%20249-126'%20fill='none'%20stroke='%23100A1F'%20stroke-width='32'%20stroke-linecap='round'%20opacity='0.28'/%3e%3c/g%3e%3cg%20fill='none'%20stroke='%23A967FF'%20stroke-width='22'%20stroke-linecap='round'%20stroke-linejoin='round'%3e%3cpath%20d='M167%20338h-56v-62'/%3e%3ccircle%20cx='111'%20cy='276'%20r='22'%20fill='%230E0B1C'%20stroke='%23A967FF'/%3e%3cpath%20d='M842%20417h65v-70'/%3e%3ccircle%20cx='907'%20cy='347'%20r='22'%20fill='%230E0B1C'%20stroke='%23A967FF'/%3e%3cpath%20d='M817%20662h80v73'/%3e%3ccircle%20cx='897'%20cy='735'%20r='22'%20fill='%230E0B1C'%20stroke='%23A967FF'/%3e%3cpath%20d='M172%20649h-71v71'/%3e%3ccircle%20cx='101'%20cy='720'%20r='22'%20fill='%230E0B1C'%20stroke='%23A967FF'/%3e%3c/g%3e%3c/svg%3e)
AI Ninjasby HaloNinjasKnowledge
The answer exists in three systems and nobody can find it in time.
Retrieval over the documents, records and prior cases you already have, surfaced in context with sources attached.
AI Beyond Halo
Not every AI problem lives in a service desk.
Agents, integrations, MCP tooling, retrieval systems and bespoke applications: the engineering work we do when the problem is not Halo-shaped. Built by people who run AI in live operations, with the evaluation and monitoring needed to trust what it produces.
A briefing, not a pitch. We will tell you where AI is worth the effort, and where it is not.
Work trace
REQ-4471 · illustrative“The delivery window has passed. Where is order 4471?”
Email from a customer, 09:12. Not a ticket. Not in Halo.
- Intake
Classified as a delivery exception. Priority raised: customer-facing, service level at risk.
- Context
Retrieved and attached:
- Held at depot since 08:40 carrier API, MCP
- Two pallets, next-day service ERP, integration
- Credit due if more than 24h late contracts, retrieval
- Routing
Sent to Dispatch with the context attached, not a link to go and find it.
- Draft
Reply drafted. Rebooking proposed for tomorrow, 08:00 to 12:00.
- Held
Awaiting human review. The agent may propose a rebooking. It may not book one.
- Resolved
Reviewed by Dispatch. Rebooked. Reply sent. Credit flagged to Finance.
Every step carries its source, so the answer can be checked.
Three problems that arrive without a ticket.
None of them is a service desk problem. Each becomes an AI problem only once the plumbing is right.
Tooling
An agent needs to act in your systems, not just talk about them.
Tool and MCP server development, with an authority model that decides what the agent may do and when it hands back.
Integration
The data that would make AI useful sits behind an old interface.
Identity, records, data and observability wired together so AI works with live operational data rather than beside it.
What we build.
Six ways to engage us on work that sits outside the Halo lifecycle. Most projects start with one and grow into two or three.
Workflow AI automation
AI applied to the operational workflows that actually run.
Automated routing and classification, enrichment before work reaches the next team, and the checks that keep a human in the loop where it matters.
Knowledge and retrieval
Your existing knowledge, grounded and usable in context.
Retrieval over runbooks, prior tickets and internal documents, surfaced at the moment someone needs it, with sources attached so answers can be checked.
Agents and MCP tooling
Agents that reason, use tools, act within limits and hand back to people.
Agent design, tool and MCP server development, evaluation harnesses, and the authority model that decides what an agent is allowed to do.
Enterprise AI integration
AI connected to the systems your organisation already runs on.
Identity, ITSM, data and observability wired together so AI works with real operational data rather than sitting beside it as a separate tool.
Bespoke AI applications
Prototypes and production applications built around a specific problem.
From a working prototype in weeks to a supported application, including the evaluation and monitoring needed to trust what it produces.
Managed AI operations
AI run as a production discipline, governed and measured.
Oversight of behaviour, prompt and model changes, data boundaries and operational metrics.
On Halo, this is delivered as AI Service Assurance inside the lifecycle.We primarily work with OpenAI, Anthropic and Microsoft, selecting services to fit the requirement rather than leading with a model.
What AI looks like across a work lifecycle.
The detail varies by environment, but the shape of a well-applied AI workflow tends to look like this. AI runs alongside your people, with a human in the loop at every stage.
StageToday, typicallyWith AI applied operationally
01
Intake
Work waits in a queue until a person reads it. Priority is whatever the subject line says.
Classification, priority and intent are detected as work arrives. It lands already classified, prioritised and contextualised.A person still owns the queue.
02
Context
People search a knowledge base that is often out of date, then go and look in two other systems.
Relevant knowledge, similar past cases and asset context are retrieved automatically, from live sources, with the source attached.Answers can be checked.
03
Routing
Work bounces between teams until someone recognises it as theirs.
Sent to the right team or workflow based on content and operational rules, with the context travelling with it.Rules are yours, and visible.
04
Resolution
Recurring issues are solved manually each time they appear.
A draft response and suggested next steps are surfaced for review. Recurring patterns are handled by workflows that people review.Actions within limits; approval where it matters.
Illustrative. Actual implementation is scoped to each environment.
How an engagement runs.
Briefing
A conversation about what you are trying to build. We tell you where AI is worth the effort and where it is not.
Prototype
A working version, used on real cases rather than demoed on slides. If it does not earn its keep, it stops here.
Production
Hardened and supported, with the evaluation and monitoring needed to trust what it produces.
Managed operations
Behaviour, prompt and model changes, data boundaries and metrics, run as a production discipline.
How we build.
The same discipline we run on Halo, applied anywhere.
Not model-led
We primarily work with OpenAI, Anthropic and Microsoft, selecting services to fit the requirement rather than leading with a model.
Evaluated before trusted
Every agent ships with an evaluation harness, and every application with the monitoring needed to trust what it produces. Behaviour is measured, not assumed.
Limits, then hand-back
An authority model decides what an agent may do, what needs approval, and when it hands back to a person. Review sits where the risk is, not everywhere.
Client data can be processed wherever the organisation requires, including on premises, in a private cloud, or in regional and national data centres.
Client data is never used to train models.
Questions we get asked about the engineering.
Do we need to be on Halo?
No. This is the platform-agnostic half of what we do. If you do run Halo and the problem is a service that needs running, start with the lifecycle instead.
How do we start?
With a briefing. Then, usually, a prototype in weeks, used on real cases. Production and managed operations follow if the prototype earns them.
Which models and clouds do you use?
We primarily work with OpenAI, Anthropic and Microsoft, selecting services to fit the requirement rather than leading with a model. We will say so if a workflow does not need a model at all.
How do you keep people in the loop?
An authority model sets what the agent may do and where it must ask. Review checkpoints sit where the risk is, and every step carries its source, so a person can check an answer rather than trust it.
Can you work with our older systems?
Usually that is the job. The useful data tends to sit behind an old interface, and an agent that cannot reach it is a chatbot. Integration is where most of the value is unlocked.
What is managed AI operations?
Ongoing oversight of behaviour, prompt and model changes, data boundaries and operational metrics, so an AI application keeps working after the project ends. For AI services on Halo, it is delivered as AI Service Assurance inside the lifecycle.
Tell us what you are trying to build.
Bring the problem, the systems it lives in and what you have already tried. We will tell you where AI is worth the effort, and where it is not.
Running AI on Halo?
Start with the lifecycle.
Take the five-minute scorecard or see how AI services are managed in Halo.