AI & LLM Systems
Document search, extraction and drafting with human review.
AI tools can take the first pass at a lot of office work: reading documents, pulling out details, sorting requests and drafting replies. We build those features into the way you work, with the controls around them: who can use them, what they can see, a log of what they did, and a person who approves the result where the stakes are high.
Send an enquiryIllustrative example
Extract the details. Show what needs checking.
A useful first pass keeps the source alongside its draft and makes missing information explicit.
“We need the inspection report for the north workshop. The last one was done in March. Can you send it to the new site manager?”
Keep the original wording available to the reviewer.
- Document
- Inspection report
- Location
- North workshop
- Report date
- March — year not given
- Recipient
- Unconfirmed
Ask before sending. Confirm the report and recipient's access. The model's draft does not authorise disclosure.
When this helps
Your team reads the same kinds of documents all day.
Invoices, applications, referrals or contracts are opened, read and summarised by hand.
Enquiries pile up.
Routine questions wait behind complex ones because everything lands in one inbox.
Staff are already using AI tools on their own.
Nobody knows what's been pasted into them, or whether anyone checked what came back.
What we can build
Document reading and extraction
Details pulled from forms, emails and PDFs into your systems, with anything flagged as uncertain sent to a person to check.
Triage and routing
Incoming requests sorted by type and urgency, and sent to the right person or queue.
Drafting assistants
First drafts of replies, summaries and reports for a person to review, edit and send.
Search over your own documents
Answers drawn from your policies, procedures or records, each with a link to its source so people can check it.
Agents for well-defined tasks
Software that carries out a repetitive, rules-based task inside limits you set, and hands anything outside them to a person.
How we approach it
- 1.
Look at the task first
What goes in, what comes out, how often it goes wrong today, and what an error would cost.
- 2.
Decide where the person sits
Each step is sorted: suits an agent; AI-assisted, person approves; or stays with a person. Money, safety, complaints, people's jobs and regulated decisions stay with a person.
- 3.
Try it on your own examples
We run it on examples of your work and show you where it gets things wrong, so you can judge whether it's good enough to use.
- 4.
Log, limit and review
Features are designed to log what they did, work inside limits you set, and pass anything uncertain to a person. Under a support agreement, we review the exceptions and adjust the limits as your work changes.
What you receive
- Working AI features
- Built into the systems your team already uses, with the limits and approval steps written down.
- A record of the trial
- The examples it was tried on, where it got things wrong, and what changed as a result.
- Logs you can read
- What each feature did, when, and who signed off the result where sign-off was required.
- A runbook
- How to pause each feature, change its limits or switch it off.
- A data register
- Which AI services your information passes through, and the country each one is in.
What we need from you
- Examples of the work, with names and identifying details removed wherever we can.
- A person who will own each feature's results and agree its limits.
- Your decision on which AI services your information may pass through.
- Access to the systems involved, through their own invite or permission settings.
Before committing
Scope depends on the systems, data and access available. These are the constraints we discuss with you.
Risks we check
- Personal or sensitive information going into an AI service your agreement doesn't name.
- Output that sounds confident but is wrong, reaching a customer before a person has seen it.
- AI making or shaping decisions about people's jobs, eligibility, money or safety.
- Business information pasted into personal AI accounts.
What we won't recommend
- Letting AI send, pay, sign off or lodge anything without a person where the stakes are high.
- AI where a simple rule would do the job.
Where we're not the right fit
- If a rule, a template or a better form would solve the problem, we'll say so before anyone mentions AI.
- We don't train AI models from scratch. We build with established AI services and your own documents.
- We don't give legal or employment advice about AI at work. Where roles could change, get your own HR or legal advice first.
Questions about ai & llm systems
What if the AI gets something wrong?
It will, sometimes. The approach above sets out the logs, limits and hand-offs we design for that. Where a mistake would matter, a person signs off the result before it goes anywhere.
Where does our information go?
Your agreement names every service that stores or processes your information, AI services included, and the country each one is in, including any overseas. You decide what's allowed before anything is built.
Which AI service do you use?
We choose per project, based on the task and on where your information may be processed. Your agreement names the service and the country it's in.
Can AI make decisions on its own?
For repetitive, low-risk, rules-based tasks, an agent can work inside limits you set. Decisions about money, safety, complaints, people's jobs or regulated matters stay with a person.
Will this replace my staff?
We report on tasks, never on named individuals. What you do with the time is up to you, and any change to someone's role should follow your own HR or legal advice.
Do we need a workflow review first?
Not always. If you already know the task, we can start there. If you'd like to find out which tasks suit AI and which don't, a workflow review sorts them before anything is built.
Related capabilities
Have a project in mind?
Send a short description of the problem and the systems involved. We will review fit and availability.