Put Practical AI to Work on the Questions, Documents and Enquiries Your Team Handles by Hand

Not always, and we will tell you when it is not. AI earns its place where the work involves language, judgement or messy documents — answering varied questions, reading bills that arrive in twenty formats, sorting complaints by urgency. If your process is a fixed rule that always produces the same outcome, a rules-based workflow or a straightforward integration will be cheaper to build, easier to audit and more reliable. The assessment stage exists to make that call honestly before you spend on a model.

It uses exactly the data you approve, and nothing else. Your documents and records go into a retrieval store that sits in your own cloud account or in infrastructure we run for you, under access controls you define. We sign an NDA before discovery begins, redact personal data where it is not needed for the task, and tell you in writing which provider processes what and under which settings. Where your data must not leave your own infrastructure at all, we can build on self-hosted open models instead — that changes the cost and the capability, and we will walk you through the trade-off before you decide.

Yes, and this is how we build by default. The assistant is restricted to a defined set of your documents and records, each with an agreed current version and a named owner. When a question falls outside that set, it says it does not have the answer and offers to connect the person to your team, rather than filling the gap with something plausible. Topics you want it to refuse outright — clinical advice, legal opinions, pricing exceptions — are blocked explicitly.

We build an evaluation set from your real questions and real documents, with the correct answers agreed by your own team, and score the assistant against it before go-live and again after every significant change. You see the failures, not just the successes. No assistant is right every time, so the more important design decision is what happens when it is wrong: low-confidence responses hand over to a person, risky categories never answer alone, every reply carries its source so a mistake is visible rather than hidden, and the conversation log lets us find and fix the weak area.

Yes, and for anything that commits your business we insist on it. You decide the line: some businesses let the assistant answer routine questions directly and send only pricing and refunds for approval, others start with every outgoing message reviewed by an agent and relax it once they have seen the quality. Draft-and-review is a normal mode — the assistant prepares the reply with the relevant record attached, and your agent sends it after a glance.

Yes, wherever the system offers an API, database access or a scheduled export. We connect to Tally, Zoho, SAP Business One, custom ERPs, SQL databases and helpdesk tools so the assistant can look up a live order, consignment or invoice status and write captured enquiries back as proper records. Where a system is closed or very old, we work through file exchange or a database bridge instead, and we confirm what is possible during assessment rather than assuming it.

It is normal, and it is work we plan for rather than a reason to stop. Most businesses come to us with the current rate list in one person’s mail, three versions of the same SOP in a shared drive, and policies that were never written down. During knowledge preparation we collect what exists, identify duplicates and superseded versions, and get one owner to confirm what is current. The only genuine blocker is when the information does not exist anywhere in writing — then that has to be created first, and we will scope it with you honestly instead of building on top of nothing.

The main drivers are the number and complexity of tasks, the condition of your documents and data, how many channels the assistant must serve, how many systems it must read from and write to, how much human-review workflow is required, and whether you need hosted models or self-hosted ones. Timeline moves with the same factors, plus two that sit on your side: how quickly document owners confirm approved versions, and how promptly your reviewers work through evaluation results. A single-task assistant on one clean knowledge base is a much shorter piece of work than a multi-channel assistant wired into three systems. After the assessment we give you an itemised proposal and a phased schedule, with no hidden charges.

There are three parts, and we size all of them before you commit. Model usage is billed by volume — messages answered, pages read — so it rises as adoption rises; hosting and storage are a steadier monthly figure; and support covers monitoring, content refreshes and accuracy re-testing. During assessment we estimate your expected monthly volume from your actual enquiry and document counts, and we build in usage caps and alerts so a busy season cannot produce a surprise bill. Where volume is high or data must stay in-house, self-hosted models shift the cost from per-use to fixed infrastructure, and we will show you where that crossover sits for your numbers.

Yes, all of it. You receive the full source code repository, the prompts and configuration, the retrieval setup and the evaluation sets, with complete intellectual property ownership. Cloud and model provider accounts are created in your name and billed to you, so you can change vendor, change model or bring another team in without asking us for permission or paying an exit fee.

Four things. A single decision-maker who can approve scope and settle disagreements. Access to the people who actually do the work, for a few sessions each. The real material — documents, past conversations, sample bills, including the messy examples rather than only the clean ones. And two or three reviewers who will work through the evaluation results and tell us which answers are genuinely correct. Projects that run late almost always do so because document owners and reviewers were not available, not because of the technology.

Yes. Training covers what the assistant is reliable at, what it must never be trusted with, how to read a cited source, and how to take over a conversation — delivered as short sessions with plain-language guides your team can keep. After launch there is a warranty period, followed by optional support covering conversation monitoring, correction of weak areas, content updates when your documents change, re-testing when a model version changes, and cost and usage reporting.