FAQ
A healthy doubt about AI deserves a straight answer.
We've answered the questions a careful owner should ask before putting AI into their business.
Most AI jargon describes simple ideas.
Models, agents, harnesses and workflows can sound like gobbledygook. Each one is a means to the same two ends: jobs that get done faster, and work that costs less to do.
Model
The engine that reads and writes text, such as OpenAI's GPT or Anthropic's Claude. It knows a lot in general and nothing about your firm. You pay for it as you use it, so its cost rises and falls with the work.
Prompt
The instructions a model gets before it starts. We write and test them once, so your team never has to work out what to type.
Assistant or agent
A model given one job, your documents and a few tools, such as reading a mailbox or filling a template. An agent can take steps on its own, like filing an attachment, before a person checks the result. This is the part that takes hours off a job.
Workflow
A fixed set of steps that runs the same way every time, such as filing each document a client emails and ticking it off the checklist. Many steps need no AI at all, and an assistant handles the ones that need reading or writing. Workflows remove the copying and retyping that cause mistakes.
Harness
Everything around the model that keeps it on task: the instructions, the fixed rules in code, the documents it may read and the approval step. The harness makes an assistant reliable enough for real work, and it's most of what we build.
Connector
The link between an assistant and a system you use, such as your email, your files or your accounting software. Connectors let figures move once, automatically, so nobody types them in twice.
Hallucination
When a model states something false with confidence. We guard against it by making the assistant work from your documents, keeping rules in code and sending every output to a person.
Human in the loop
A person who approves the assistant's work before it goes anywhere. Your people keep the decisions and hand over the typing.
Running cost
What Microsoft, Google or the model provider charges each month for the assistant to run. We name it in the scope before you sign, so you can set it against the hours the job takes today.
Every part we build has to make a job faster or cheaper to run.
Trust
AI makes things up. Why would we trust what it writes?
It does get things wrong, and an assistant that's left to decide will eventually decide badly. We give it the parts of a job it's good at: reading your documents, gathering facts and drafting from your templates. Your fixed rules run in code it can't change, and a named person approves every output. Checking a good draft takes far less time than writing one, and that saving is where the opportunity lies.
We tried ChatGPT and it wasn't much use.
A general chatbot knows nothing about your firm, so it gives general answers. We build each assistant around one job, with your templates, your past work and your rules. That's the difference between a clever stranger and a trained junior who already knows how you work.
Our work needs judgment. AI can't do it.
Most jobs that need judgment also involve hours of work that needs none. We split the job so code applies the rules, the assistant gathers and drafts, and your people make the calls. Your most expensive people then spend their time on the part only they can do.
Who's responsible when it gets something wrong?
The same person who's responsible today. Every output goes to a named approver on your team, and nothing leaves the building without their sign-off. The difference is that they start from a draft instead of a blank page.
Data and control
Our client data is confidential. We can't hand it to an AI company.
That's the right instinct. What we build runs in your own accounts, under your admin. Your admin decides who can use each assistant and which AI models it may call, and you can read each provider's data terms before you sign. Your data stays where your team already keeps it.
We're regulated. Can we use AI at all?
Regulators expect you to stay in control, and our builds keep you there. Fixed rules sit in code you can inspect, a named person approves every output, and written notes explain how each part works, so your compliance team can review it. We run assistants in a regulated investment business ourselves, Cur8 Capital.
What if you disappear, or it breaks after you've gone?
That's a fair question for any supplier. Everything we build sits in your accounts, and you hold the logins and written notes, so any competent developer can pick it up. You're never locked in to us.
Cost and timing
It's all hype. We'll wait until it settles down.
A lot of it is hype. The jobs we automate are ordinary ones that AI already handles well: drafting from templates, filing documents and moving figures between systems. We build so you can switch AI models without rebuilding, so waiting for better models gains you nothing. The firms that start now get the hours back first.
This will cost more than we think, and the costs will creep.
That happens when the scope is vague. You get one fixed fee in writing before we start, and the monthly running costs are named in the scope before you sign. If the work takes longer than we planned, that's our cost. You can weigh the price against the hours the job takes today.
We're too small for an AI project.
Small teams feel lost hours the most, because there's no one to pass the work to. We start with one job, and you have a written scope with one price within a week of the first call. Once the first assistant is running, adding a second is usually much smaller work.
People
Will this replace our staff?
We aim it at the jobs your people like least: chasing documents, retyping figures and redrafting last year's report. Your team keeps the client work and the decisions. The usual result is that the same team takes on more work.
My team won't use it.
People ignore tools that add steps. We build into the email, documents and approval flow your team already uses, so the assistant shows up where the work already happens. A draft that's waiting for them is easier to adopt than a new system to learn.
Bring us your hardest objection.
Book a 20-minute call and tell us the job you'd most like to stop doing by hand.