Automation
AI & automation
Work your team repeats by hand costs hours every week and lets mistakes through until a client spots them. We automate it: assembling documents, routing requests, sending instant alerts. If it won't save hours or money, we'll say so.
What we do
AI & automation, at Kazap, means taking a task your team repeats by hand and having software do it. Most of those tasks follow rules: a request about property goes to the property team; a form with a company number fetches that company's details. Rules are cheap, predictable and easy to check, so we use them first.
AI integration comes in where a rule cannot do the job: reading a free-text email to decide where it should go, pulling fields out of documents that arrive in different layouts, or drafting a first reply for a person to check. In each case a large language model (LLM), the technology behind assistants such as Anthropic's Claude, does one named task inside a process you already run, and a person stays responsible for the result. We build with Claude every day.
The automation we have delivered for clients is rules-based, and it carries real work. A legal-document platform assembles each document from saved data and produces a signed PDF. A legal-advice portal for a London law firm routes requests to lawyers in 45+ countries and sends instant alerts, in place of email threads with 50+ recipients; turnaround became two to three times faster. A lease and licence platform cut a sign-off between five and seven parties from up to two weeks to two hours, once everyone agrees a time.
Our own product works the same way. YLB Score runs automated, passive checks on a local business's website and returns a 0–100 score with a plain-English report. It is free, and has been live since September 2026.
What you get
- A list of the repetitive tasks, with the time each one takes now
- A recommendation per task: a rule, AI, or leave it alone
- A prototype run on a sample of your own data
- A pilot with go/no-go criteria agreed before it starts
- Version one in daily use, hosted and supported
Prototype → pilot → version one
AI projects stall when a demonstration is mistaken for a product. We use three stages, each ending in a decision.
- Prototype (proof of concept, or PoC). The task run on a sample of your own data, to measure how often it gets the answer right.
- Pilot. The automation working inside the real process for a fixed period, with a person checking its output, against criteria written down before it starts: accuracy, time saved and cost per task. At the end, go or no-go. A no-go is a useful answer; it costs less than a version one nobody uses.
- Version one (minimum viable product, or MVP). The automation in daily use, with the checks that proved it in the pilot kept in place, hosted and supported by us.
How it runs
- List the tasks. We sit with your team and time the repetitive work. Gate: you choose one task worth automating.
- Prototype. Gate: is it accurate enough on your data to be worth a pilot?
- Pilot. Gate: go or no-go against the criteria agreed in advance.
- Version one and support. Gate: we review it with you after launch and adjust or switch it off.
Who it is not for
- Anyone after an "AI strategy" with no task attached. We start from a named task and the hours it costs.
- Decisions nobody will check. If a wrong answer could reach a client or a regulator unseen, that step stays with a person or with rules.
- A chatbot added because a competitor has one.
Question
Do we need AI, or would plain automation do?
Usually plain automation. Kazap Technologies uses rules wherever a task follows them, because rules are cheaper, more predictable and easier to check, and adds AI only where a rule cannot do the job, such as reading free-text requests or documents in mixed layouts. Every AI task goes through a pilot with go/no-go criteria agreed in advance. If it won't save hours or money, we say so.