Build or buy: AI for a law firm
By the ChamberOne team 5 minute read
A partner returns from a conference excited. A developer friend says a working prototype could be built over a weekend with a public model and a few lines of code. Why pay a subscription for something the firm could own? It is a fair question, and the weekend prototype will probably look impressive. The trouble starts in month three.
Key points
- The model is the small part. Security, access control, records and support are the real work.
- Most firms should buy the platform and build only a thin layer for what is truly distinctive.
- Count staff time and turnover risk in the cost of building.
- Run a short, scored comparison on three real tasks before deciding.
What building actually means
People use the word build for very different things. Writing a prompt template is not the same as running a system. A real internal AI tool needs a way to upload and store documents, access controls that follow your matters, a record of use, a way to switch models, error handling when the provider is down and someone to answer the phone when it breaks.
The model itself is the smallest part. The work is everything around it. Most of that work is invisible in a demo and very visible on a bad day.
The case for buying
For most firms, buying wins for dull reasons. A vendor spreads the cost of security reviews, upgrades, support and testing across many customers. You borrow that effort for a subscription. You also get improvements you did not have to plan, and a party you can hold to account.
Buying also fits the skills of a law firm. A ten-lawyer firm has no spare engineer. A forty-lawyer firm may have an IT manager who is already busy with devices and email. Asking either to become the maintainer of a document-processing system is asking for a single point of failure.
When custom work does make sense
There are honest cases for building, and they are narrower than enthusiasts suggest.
- You have a distinctive process that no product handles, and that process is a source of competitive advantage rather than a convenience.
- You hold a body of know-how, such as a large precedent bank, that you want to make searchable in a specific way.
- You have, or are prepared to hire, people who can maintain the system for years, not months.
- A client or regulator requires a particular arrangement that no vendor offers.
Even then, a hybrid is often wiser. Buy the platform for matters, documents, billing and access control. Build a thin layer for the one thing that is truly yours, using the platform's connections where they exist.
- A vendor spreads security, upgrades, support and testing across many customers
- You get improvements you did not have to plan
- You have a party you can hold to account
- A distinctive process that is a source of advantage
- A body of know-how, such as a large precedent bank
- People who can maintain it for years
- A client or regulator requirement no vendor meets
The hidden costs of owning it
Think of a prototype as a puppy. The purchase price is the smallest cost.
- Model changes: providers retire and update models. Your prompts may behave differently after an update and someone must test them.
- Security: you own the review, the patching, the keys and the incident response.
- Access control: client confidentiality and ethical walls must be built into search and retrieval, or the tool will happily show a person documents they should never see.
- Record keeping: you will want a log of who used what. That is its own build.
- Staff turnover: the person who wrote it leaves, and the firm discovers that nobody else understands it.
- Support: lawyers will report problems at the worst moment, such as the night before a hearing.
A short evaluation you can run in a month
Instead of debating in the abstract, run a small, honest comparison.
- Pick three real tasks, such as summarising a contract, extracting dates and drafting a first-pass letter.
- Run them through a bought product and, if you must, through a quick prototype.
- Have a lawyer who was not involved in building either score the output against the source documents.
- Count the hours spent on setup, checking and fixing. Include the developer's time at a real rate.
- Ask what would happen if the person who built the prototype left tomorrow.
Most firms find the exercise answers the question quickly. The bought product usually wins on setup time and the prototype on flexibility, and the real debate is whether you need that flexibility enough to pay for it every year. Write down the answer you reach and the assumptions behind it, so that the decision can be revisited when the assumptions change.
Questions to ask when buying
- Where is our data stored, and which AI providers receive it?
- Can we choose the model, or switch AI off entirely?
- Is there an audit record of AI use that we can export?
- How is access restricted for sensitive matters?
- What can we take with us if we leave, and in what format?
- What does the product say it cannot do?
Be wary of any vendor who cannot answer the last question. Every product has limits, and honest ones are easier to plan around.
A scenario: the weekend prototype in month six
Suppose a twelve-lawyer firm builds a contract summariser. It works well on the first fifty contracts. Then the provider updates its model and the summaries change style. A lawyer notices that dates are now occasionally formatted differently, and a diary entry is wrong. The developer is on holiday. Nobody knows which version of the instructions was used on which contract, because the prototype kept no record. The firm now has to re-check months of output.
Nothing in this story is exotic. It is what ordinary maintenance looks like when no one has been given the job. A bought product does not remove such problems, but the vendor carries part of the burden, and a decent one will tell you when the model underneath changes.
Decide as a partnership, not as an enthusiast
Build-or-buy decisions are often made by whoever feels most strongly. Write the choice down instead. State what problem you are solving, who will own it, what it will cost over three years including people, and the date on which you will review whether it is working. A decision with a review date is far less likely to turn into an orphan system that nobody dares to switch off.
- State the problem you are solving and who will own it
- Count the cost over three years, including people
- Set a date to review whether it is working
- Let whoever feels most strongly decide
- Leave the choice unwritten and without a review date
This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.
Read next
How a small law firm can start with AI without a big project
Pick two low-risk uses, set review rules and measure before expanding, with a list of what to postpone.
5 minute readA 90-day AI pilot plan for a law firm
A week-by-week plan to choose a use case, set measures, run a small trial, review errors and decide.
5 minute readChoosing AI models and firm-level controls: what a practice should decide
Model choice, an off switch and a usage record are three decisions a firm should make deliberately.
4 minute read