AI integration for
the software you already run.

We add AI search, drafting, replies and agents to your own products and workflows: connected to your data, behind your permissions, with a person approving what matters.

Answers from your own documents.

Search that answers with its sources, drawing only on documents the person asking may see.
Based on the legal research platform we built.

Has the appeal court accepted late notice as a defence in supply contract disputes since 2019?

Yes, but narrowly: only where the contract makes timely notice a clear condition of the right to claim.1

Court of Appeal · 14 March 2021

Halden Foods v. Mercer Packaging

¶ 42Late notice will be a defence only where the contract makes timely notice a clear condition of the right to claim …

Has late notice been a defence in marine cargo claims since 2019?

No supporting passage was found in your matters, so no answer is given.

Instead of a guess

  • Has late notice been a defence in claims against a road haulier?
  • Search again across all years

One restricted matter was not searched. Its documents never reach the model.

How much should the AI do on its own?

Drafts for your team and replies to your customers, at the level you set for each task.
On the AI sales agent we built, drafts are the default.

The AI suggests.
A person writes.

It reads the conversation and points to the next step of the sales script.

Is the 15% code still on? The Lune trousers are a bit over my budget this month.

Next step in the scriptHandle the price objection

A person writes the reply

The AI drafts.
A person reviews.

It writes the whole reply in the brand's voice. A person sends it, edits it or discards it.

Is the 15% code still on? The Lune trousers are a bit over my budget this month.

That code has expired, but I can offer you free delivery on the Lune trousers. Shall I send you the link?

A draft, waiting for a person

The AI sends,
within your rules.

It replies on its own only where your rules allow. Everything else waits as a draft.

When does the Mira dress come out?

It's in the shop on Monday at 10:00! Want the link when it's live?

Sent by the AI, within your rules

to find and confirm a source, down from hoursLegal research platform
~5 sec
to reply to a customer, down from about 20 minutesAI sales agent for social media
3–4 sec
the share of conversations that end in an orderAI sales agent for social media
2×

Where does the model get what it needs?

Instructions, retrieval, fine-tuning or an agent: four ways to give a model what it needs for your work. Most systems combine them.

Watch for

Answers drawn from general knowledge instead of your own records.

Watch for

Document quality, permissions that must stay in sync with the source, and scans that need OCR first.

Watch for

Enough good examples, a model that can be fine-tuned, and new training when the style or the base model changes.

Watch for

Instructions hidden in what it reads, tools with too much access, running costs and errors that build up over many steps.

Agents that act in your systems.

An agent can read email and web pages, call your tools and hand work to another agent.
Based on the AI agent platform MVP we built.

A research agent reads a web page

To a person it reads as news. The hidden part speaks to the model.

Hidden text,
written as a command.

  • Pages, emails and files an agent reads can hide instructions aimed at the model.
  • With wide tool access, an agent could act on them: export records, send email.
  • No filter is known to catch every prompt injection.

One task first. Then the build.

An AI feature can only be judged against a clear definition of a good result. We agree it first, with examples from your own work, and measure against it before and after launch.

  1. 1One task

    A task, a measure, a price.

    We agree one task, how success is measured and the examples to test it on. You get a scope with a fixed price, and decide whether to build.

  2. 2Build and release

    Built in, tested, released.

    We build the feature into your product, with the permissions and approval steps it needs, test it on your examples, then release it and hand it over.

Where should AI
help your team?

Tell us the task: what people search, write or answer today, and which systems hold the data. We will come back with questions and a first view of what to build and how to measure it.

Discuss your project

Frequently Asked Questions

  • It depends on the task and how strictly its answers must be checked; your data (volume, formats, scans that need OCR and how permissions are set up); whether the AI only suggests or also acts in other systems; the models and hosting, which also set the running cost of each request after launch; and how much evaluation the use case needs. After a first conversation we outline the work, timing and budget. We agree a fixed price for a clear scope, and you approve the cost of any additional work before it starts.

  • We agree the timeline with the scope. For reference, not as a standard length: the legal research platform and the AI sales agent each went live after two months of development and then had a month of support; the AI agent platform took two months from brief to a working MVP.

  • Yes. The first phase covers one task: we agree how success is measured, collect examples from your work to test against, and map the data, permissions and models it needs. You receive the success measures, the test examples, a data and access map and a scope with a fixed price for the build, then decide whether to go ahead. You can also order this phase on its own, as AI integration consulting.

  • Yes. Most AI features fit into what you already run: a search panel, a draft next to a message, an agent with access to selected tools. We review the product, the data and the permissions first, then agree a focused scope that fits your existing code and release cycle.

  • It can. Language models sometimes give wrong answers with confidence, so we design and test for it. Where answers come from your documents, the search finds the relevant passages first and the model writes from them with citations. Citations can still be wrong, so they open the source for a person to check. When nothing relevant is found, the system can say so instead of guessing. Questions it should decline are part of the evaluation before launch, and after release we watch how it answers in real use.

  • We agree external services and data access upfront, exclude your data from AI provider training, and support hosting within your infrastructure.

    This page covers AI inside your product or workflow. For how our own team uses AI to build software, see How we use AI.

  • We choose the model for each task. The AI agent platform we built lets each agent run on its own model: one from OpenAI or Anthropic, or an open-source model such as Llama, Qwen, Mistral or DeepSeek on cloud GPUs. The legal research platform uses a managed cloud model, and the AI sales agent a model fine-tuned on the agency's conversations.

  • Yes, when a conversation is the right interface. For the AI sales agent we connected the agency's social and messaging channels to one queue. The agent drafts replies in each brand's voice and follows the agency's sales scripts. It sends on its own only where the agency's rules allow; everything else waits for a person. For rules-based steps that need no AI, see workflow automation.

Let’s build your next advantage

Tell us what you want to launch, improve or automate. Let’s discuss the scope and the next step.

Get in touch

Prefer email?

sales@cursopic.com

Cursopic

Tell us about your project

What would you like to launch, improve or automate? A few lines are enough to start.

We use what you send only to reply to you about your request. Privacy policy (opens in a new tab)

Prefer email?

sales@cursopic.com