We build AI inside youroperation, then run it.
You'd be buying one system, built into your operation and run for a monthly fee. Before we build, whoever owns the budget signs what the work costs today. Most jobs we look at aren't worth automating.
Before
Plain emails, scanned PDFs, the odd Word attachment, worked into the case system by the carrier's most experienced adjusters.
After
A claim arrives. Twelve fields come out, each with a confidence score.
The loss amount is unreadable, so it waits for a person instead of getting a guess.
Approved with one fix. Written back, read back, logged.
of adjuster time per file, and manual data entry across the desk fell by half.
The case platform's vendor rated the bug that blocked us a P3, after agreeing we were right. We shipped around the vendor, with retries timed around its nightly downtime.
The receipt.
This job saved 22 hours this month. Here is the log.
The baseline was taken before anything changed, while the old way was still running, and signed by whoever owns the budget. The number is allowed to go down, and the reason prints beside it.
Illustrative run — not a live customer record
- 01
What was the baseline?
- 02
Who measured it, and when?
- 03
Does the running system report its own numbers?
- 04
Is the number able to go down?
We answer all four in writing. Ask them of whoever else is in the running, and the answers will tell you which numbers were measured and which were quoted.
Client email was eating billable hours. An agent drafts the replies, and an attorney approves every send.
The right crews in the wrong place every morning. Matching runs on what each crew is certified and inducted for, and the ops manager confirms every dispatch.
Ziani.ai watches UK committee hearings live, on the backend we built. Conventional monitoring lands the next day.
Ziani.ai is the one client that agreed to be named. The rest stay anonymous, which is the answer you'd get about your own deployment. So what we publish is the method instead of the names: how the baseline gets taken, who signs it, and the four questions above.
The pattern.
Every deployment on this page has the same four things inside it: an agent doing the work, a named person approving what leaves, a baseline taken before anything changed, and a receipt out of the run log. Only the work changes.
If your second job matches a pattern we've built before, it's configuration rather than a build, and priced that way. The rules and the memory stay yours, exportable the day you leave.
Your options.
Hiring is usually the right answer. A person handles what nobody anticipated, and you can tell them to stop. It stops being the right answer when the process is the problem, because a new hire runs that same process.
It holds until the person who knows how it all works hands in their notice.
Fully loaded, a year before they're good, and it all leaves with them.
Most teams have them. A seat with no skills and no memory is the subscription that changed nothing.
None of the Big Six publish one. Ask whoever quotes you who measured the before, and when their number last went down.
An agent inside your systems, your person over it, and a diagnostic allowed to say don't build.
The number arrives whether or not it flatters us.
Start with the working sessionThe people.
Cho Yin Yong
Teaches Engineering Large Software Systems at the University of Toronto, the youngest lecturer his department has promoted. Two AI patents. Joined a regulated engineering company as its third employee and left it at eighty.
Mahmoud Halat
Employee two at Verto Health, on the platform behind a quarter of Canada's daily COVID vaccinations. A patent in data harmonization. $100k Lovable Shipped grand prize, solo, against 5,800 builders.
Ten years of building automations in health, where a wrong answer is a compliance event. That work predates XY Space. In the UK we work with North Stack on FCA-regulated insurance, and neither of these two hands your build to a junior.
What would you handto one more person?
Start with a working session: ninety minutes, free, no deck. Then a £9,000 diagnostic measures what that work costs you today and scopes the build, with half the fee credited into it.
It comes back with every candidate job priced in hours and money, and most of them marked not worth automating. Finding the few that are is most of the value.
A month after the job goes live you get the first receipt: hours back, at what accuracy, worth how much.
Less than a year of that one more hire, and what it learns stays. It pays where the work repeats. If the bottleneck is a one-off, or nobody will put their name to the output, we're the wrong firm and will say so on the call.
Common questions.
Our people already use ChatGPT or Copilot. Isn't that this?
That's the raw model, the one part you already had. A bare seat has no skills, no memory, and nothing wired into the systems your work lives in, which is why it changed nothing. 78% of people using AI at work brought the tool themselves (Microsoft and LinkedIn, 2024). We install the rest, inside something your compliance team has approved.
Are we buying software, or a service?
A service that includes the software, rather than software with services attached. We scope it, build it in your own cloud account, run it, and report what it returned. The code, the rules and the data are yours. The underlying pattern stays ours, and nothing of yours ships to another client.
Will it work with our legacy systems?
That's most of the job. The incumbent stays authoritative, write-back goes through its own API, and every write is read back and logged. Where a system has no API, an agent operates it the way your staff do, under approval.
Who is accountable when an agent gets it wrong?
You are. You can't sue a model, but you can hold a named person and a named process to account. So the limits are agreed in scoping and then sit in the system connections and the approval steps. Anything written into a prompt can be talked around.
Does this affect our professional indemnity cover?
We are not insurance advisers, and how your policy responds is a question for your broker rather than for us. What we can tell you is that through 2026 professional-liability insurers started writing AI exclusions into E&O and D&O wordings, and that what these policies turn on is who reviewed the work. So we build the review lane first and the automation behind it: a named person approves anything that leaves, on a hash-chained ledger you can export and put in front of your broker. Last checked August 2026.
Where does our data run?
Enterprise endpoints where your obligations allow, open-weight models served in the UK or Canada, or fully inside your own cloud account, which is also the Canadian residency answer. Changing the model later means re-testing it on your samples, so we schedule that as work rather than flipping a setting.
Another vendor is quoting us a much bigger saving.
Put the four questions above to them. The tell is the second one: if the before-number came from the vendor's own model rather than from someone measuring the work while the old way was still running, the saving is a quote and not a measurement. It might still be right. Neither of you can check it.
Why now?
Harper raised $46.8m in February 2026 to run AI-native commercial insurance broking for mid-sized businesses. Lawhive raised a €50m Series B as an AI-native law firm. Both are funded to serve your clients directly, and both are UK-facing. The part of the work that stays yours is the judgement, and your people only get to spend their time there if the document-heavy end stops taking their week.
Book a call.We'll come back with specifics.
Start with your people or with the work. One job at a time, on probation, measured in hours and money. Everything we build stays yours.