AI where it earns its place.
Useful AI inside a real process — working from your own information, showing what an answer rests on, and leaving consequential decisions with a person.
Useful when AI looks promising, but nobody can say what it would actually be allowed to do.
Illustrative system example · fictional product and data · not client work
AI-assisted workflowLUMEN · illustrative system, fictional data
You may need this when
Some of this probably sounds familiar.
- Staff search for the same answers in the same documents every week.
- Work that needs judgement is slow because the context is scattered.
- AI has been suggested, but nobody can define what it should be allowed to do.
- A tool was trialled, gave a confident wrong answer, and lost the room.
- You need the reasoning behind an answer, not just the answer.
What we could build
Assistance with a boundary around it.
Scoped to one process where the benefit is measurable and the limits can be stated plainly.
- Answering over your own documentsQuestions answered from your records, with sources available alongside.
- Assisted triageIncoming work classified and routed, with uncertain cases sent to a person.
- Drafting inside a workflowFirst drafts produced in your format, reviewed before they leave.
- Extraction from documentsStructured fields pulled from invoices, forms and reports.
- Evaluation harnessA test set showing whether the system is right often enough to use.
- Human approval gatesDefined points where the system must stop and ask a person.
See an example
Ask a question. See what the answer rests on.
Take the three steps below. In this design the system gathers evidence and proposes an action — then stops, because that is where the boundary was drawn.
Shown as a completed run. The step controls need JavaScript, so they are not offered here.
- Ops leadAsked: why have Northern deliveries slipped this month?
- SystemRead 3 connected sources: delivery log (214 records), carrier reports, depot notes. Claims linked to their source.
- SystemProposed: re-route Northern orders via Depot 2 for 14 days. Held for human approval — the system does not act on this.
OutcomeNothing has been asked yet, and the system has taken no action.
OutcomeWhere an action is consequential it can be placed behind an approval gate, so a person decides.
Illustrative system and fictional data. Not client work.
How far the boundary sits is a decision you make deliberately for each process — not something that arrives with the tool.
What you receive
Something you can judge, not just try.
- A working assistant inside one agreed process
- An evaluation set showing how often it is right, and where it fails
- Written limits: what it may do, and what needs a person
- An honest recommendation if the process does not warrant AI at all
- Agreed source code, data and infrastructure handed over according to the project scope and terms
- Handover documentation and any agreed support arrangements
Technical depth
Retrieval
Your content chunked, embedded and indexed, with retrieval measured against a question set drawn from real use rather than assumed.
Grounding
Where grounded answers are required, responses can be constrained to retrieved context and sources carried through to the interface.
Evaluation
A held-out set with expected answers, scored on accuracy, refusal behaviour and citation correctness, re-run when a prompt or model changes.
Boundaries
Tool access can be allow-listed and consequential actions routed to a human approval gate, scoped to what each process warrants.
Privacy
Provider and deployment choices follow the agreed data and privacy requirements, including private or local inference where the content calls for it.
Operations
Prompt and response logging, cost and latency monitoring and versioned prompts, at the depth the agreed scope calls for.
Often combined with
What usually sits next to this.
- Data & analytics →Assisted answers are only as good as the records underneath them.
- Software & systems →The assistant usually needs a real workflow to sit inside.
- Analysis & research →When the decision needs evidence weighed rather than generated.
Tell us the task, not the technology.
Describe the work you think AI might help with. If it would not help, we will say so.