Years of consistent records
You hold several thousand or more similar documents or cases with known outcomes, such as reviewed contracts or coded claims.
Solutions
Train a Model on Your Own Data, When It's Worth It
A model trained on your own records makes sense for some firms. We check first whether connecting your documents to an agent would do the job, as it does for most.
Trusted by businesses like yours
Most Firms Should Start Somewhere Simpler
Training a model takes a large set of clean, consistent examples, and the model needs retraining as your work changes. For most firms the better first step is an AI agent that searches your own documents and cites its sources, which we can build in weeks. A trained model earns its cost when you have years of consistent records and one narrow task repeated at volume, and we'll tell you which case you're in before you spend anything on training.
Where a General Model Struggles
- 01
It doesn't know your field
A general model writes well but misses the terms and formats your specialists use every day, so its drafts need heavy editing.
- 02
The data can't leave
Client agreements or regulators require some records to stay in systems you control, so staff can't use outside AI tools on that work at all.
- 03
The same task, thousands of times
Classifying or pulling details from the same kind of document all day on a large model costs a lot per item, and the bill grows with the volume.
From Assessment to a Model You Own
Every step runs in a cloud account you own, and your data stays there.
- 1
Check whether it's worth it
We compare what a trained model would gain you with what an agent connected to your documents could do now. If the agent is enough, that's what we'll recommend.
- 2
Prepare the data
We gather the records and remove anything that shouldn't be there, then label good examples and hold some back for testing. All of it stays in your cloud.
- 3
Train or fine-tune
In most cases we fine-tune an existing model on your examples. Training a smaller model from scratch only suits very narrow tasks.
- 4
Test against simpler options
We score the model on the held-back examples and compare it with a general model and a document-connected agent on the same tasks. We only recommend it if it does better.
- 5
Run it and keep it current
We set up monitoring for quality and cost, and agree with you when to retrain the model as new records come in.
What You Keep
- A written assessment of whether a trained model is worth it
- A cleaned, labelled training set that stays in your cloud
- A held-back test set and the scores the model has to beat
- The trained or fine-tuned model, owned by you
- A side-by-side test against a general model and a document agent
- Monitoring for quality and cost once the model is running
- Docs on how it was trained and when to retrain it
Where a Custom Model Fits
One narrow task at volume
The same classification or extraction runs many times a day, so a smaller trained model could do it faster and for less.
Data that has to stay put
You need the data and the model to stay in a cloud account you own, where you decide who can see them.
Answers from your own files
If your team mainly needs answers from its documents, an agent that searches and cites them is the quicker start, and the assessment will say so.
Common Questions
Still deciding? On a free 30-minute call we'll tell you what fits.
Do we need our own model?
Probably not at first. An agent that searches your own documents and cites its sources covers what most firms need, and a trained model helps when you have many consistent examples of one task.
Where does our data go?
It stays in a cloud account you own. We prepare the data and train the model there, and you decide who on your team can see it.
How much data do we need?
It depends on the task. As a rough guide we look for several thousand consistent examples with known good answers, and the assessment will tell you if you're short.
How long does it take, and what does it cost?
The assessment takes about 2 weeks. The training work takes about 6 to 10 weeks and is quoted once we've seen the data. You pay your own cloud and computing costs.
Who owns the model?
You do, along with the training data and the test results. It runs in your account, and we document how it was built so your team or another firm could retrain it.
Related Research
How it works · 7 min read
How AI knowledge assistants work, and where they break
A company AI brain searches your files, hands the closest passages to a language model and has it write an answer with citations. Your sharing settings decide what it can see, and your documents decide how often it's right.
Case studies · 4 min read
Company AI brain case studies: what the firms with results did first
Morgan Stanley, Steno, BCI and others have published results from internal AI assistants. Most of the numbers come from vendors, and the firms with the clearest results did similar groundwork before launch.
Free 30-minute call
Find Out Whether a Custom Model Pays Off
On a free 30-minute call we'll ask what data you hold and what you want a model to do, then tell you plainly whether to train one or start with your documents.