AI integrations · vectorel

What it looks like

An answer you can check
before anyone acts on it.

1
Answers from your own files
Retrieval finds the relevant clauses in your documents, not the public internet.
2
Every claim has a source
Each statement links to the page it came from, so checking takes seconds.
3
A person approves
Nothing reaches a customer or changes a record until someone on your team signs off.

Contracts assistantLegal · 1,240 documents

What's the notice period to terminate the Northwind MSA, and has it changed?

1

90 days' written notice before the renewal date 1. It was reduced from 120 days in the 2024 amendment 2. The next renewal is 1 March 2027, so notice is due by 1 December 2026.

2

1Northwind MSA.pdf§14.2 · p. 11

2Amendment 2 (2024).pdf§3 · p. 2

3 Draft · needs approval Reject Edit Approve

Illustrative example

Why pilots stall

Four reasons AI pilots never ship,
and what we do about each.

Problem
It answers from the internet, not from you.
A general model doesn't know your contracts, your pricing rules, or your last three years of tickets. The answers are plausible and wrong.

What we do
Ground it in your documents.
Before any model work, we make your files and records reliably searchable. Most of the quality difference between AI systems is decided here.
Problem
It has no hands.
A model that can't read your CRM or write to your ticketing system isn't a system. Someone still copies the output across by hand.

What we do
Connect it to where work happens.
It reads from and writes to your CRM, inbox, ticketing system or internal tool. No new tab to remember.
Problem
Nobody trusts it.
Without sources and a review step, the team quietly stops using it, usually after the first confident mistake.

What we do
Show sources, keep a person in the loop.
Every answer links to where it came from, and a person confirms anything that reaches a customer or changes a record that matters.
Problem
It was built as a demo.
Pilots are optimised to impress in a meeting. Production needs error handling, monitoring, cost control and someone accountable when it breaks.

What we do
Measure it and cap the cost.
Accuracy is tested against your real cases before and after launch. Every request is logged, and per-request cost is monitored and capped.

What we build

Five kinds of AI system
we put into production.

Retrieval over your own documents

Ask a question in plain language, get an answer with the source attached. Built on your contracts, manuals, policies and past work, so the answer can be checked rather than trusted blindly.

Document processing at volume

Extract fields, classify, summarise and route across every document, not a sample: invoices, contracts, applications, tickets, forms. What took an afternoon per batch runs continuously.

Drafting and triage inside your tools

Replies drafted in your inbox, requests sorted into the right queue, records filled from unstructured text. Your team reviews and approves rather than starting from a blank page.

AI features in your own product

Search, recommendations and in-product assistance built on your data, designed to hold up in front of customers, with the latency and cost profile that requires.

Agents with a defined scope

Multi-step work: look something up, check it against a rule, update a record, notify a person. It runs within limits you set, with every action logged.

Use cases

Where teams use it first

Customer support
Draft replies grounded in your help centre and past tickets. Route by intent and urgency. Surface the three most similar resolved cases to whoever picks it up.
Sales and revenue
Summarise calls into CRM fields, flag risk in open deals, and prepare account briefs before a meeting from everything already recorded about that customer.
Operations and back office
Read incoming documents, extract what matters, validate it against your rules and post it into the system of record. Exceptions go to a person; the rest goes through.
Legal, compliance and procurement
Search across contracts for clauses, dates and obligations. Compare an incoming document against your standard terms and get a list of what differs.
Internal knowledge
One place to ask how something works, answered from your own documentation, with links to the source. New hires stop interrupting senior people to find out.

How AI projects run

The same three phases as every engagement,
with an evaluation set at the centre.

01 / Assessment

Is AI the right tool?

We review the process, the data behind it, and what an acceptable answer looks like. You get a written scope, a fixed price, a running-cost estimate, and an honest read on whether AI fits. Sometimes it doesn't.

About one week

02 / Delivery

Built against known answers

We collect real examples with known correct answers from your team and test every version against them. You see the results each week, including the misses.

Usually 4–10 weeks

03 / Handover

Logged, monitored, yours

Logging, cost limits and alerts are in place from launch. The code, the evaluation set and the infrastructure access are yours to keep.

Support optional

Built on

OpenAI
Anthropic
Hugging Face
LangChain
Python
FastAPI
PostgreSQL

How every engagement works Full stack

FAQ

Before you trust it
with real work.

Will it make things up?

Do you use our data to train models?

What does it cost to run?

What happens if a model or provider changes?

AI that works on your data,inside your existing tools.

An answer you can check
before anyone acts on it.

Answers from your own files

Every claim has a source

A person approves

Four reasons AI pilots never ship,
and what we do about each.

It answers from the internet, not from you.

Ground it in your documents.

It has no hands.

Connect it to where work happens.

Nobody trusts it.

Show sources, keep a person in the loop.

It was built as a demo.

Measure it and cap the cost.