Legal AI workbenches pay off when the work is redesigned first

Our view is that the return from legal AI comes from redesigning the work, and that a workbench is often the best place to run the redesigned process. We tested that view against a framework we largely agree with, against the research on where AI returns come from, and against an invented example of one legal process.

The framework comes from a LinkedIn post on 21 October 2026, in which Uwais Iqbal, founder of the legal AI consultancy simplexico, sets out three levels of AI maturity in legal work [1]. He plots legal AI tools on a two-by-two chart: general to specific on one axis, with the three levels running along it, and vendor-owned to sovereign on the other. The post ends: "ROI from legal AI comes from Level 3." His framework is a useful way to sort tools, and we agree that a workbench is often where good legal AI ends up. We differ on where the return comes from: the workbench, or the redesign it carries out. Below are his three levels, our own framework, and an example of the difference.

Exhibit 1

The three levels in the post

On his second axis, general chat assistants and legal platforms sit on the vendor-owned row, and a workbench deployed in the firm's own cloud environment sits top right. Source: [1].

Our framework: redesign first, then two questions

We start one step earlier and split his first axis in two, because whether a tool was built for legal work does not decide how much of the work it holds, or where the data goes.

1. Redesign the work firstChoosing a tool before this step automates the old process.
2. How much of the work does the tool need to hold?One answer, a few steps, or one whole piece of work with its documents, people and steps in one place.
3. Where does the data go at each step?From the provider's own service to a model running inside your organisation, with two common positions in between.

Exhibit 2

Most AI tools hold one answer on someone else's service; redesigned work usually needs more

Where common types of AI tool sit on our two questions. No position is the best one: a team picks the cheapest that fits its redesigned work and its confidentiality rules.

1. 1General chat assistant
2. 2General workflow tools
3. 3Legal chat assistant
4. 4Legal workflow features
5. 5Legal platform's workbench
6. 6Saved prompt in your own project
7. 7A skill in your own project
8. 8What our course participants build
9. 9Firm-built workbench

A tool can sit in a different place depending on its settings and connections. The dashed box marks where tools built around a redesigned piece of work usually sit.

What his framework gets right

• A chat window is a poor fit for most legal work. As the post puts it, a matter has several documents, several people and several steps.
• A workbench can hold all of that in one place. It is what participants on our course build.
• Where a tool runs affects confidentiality. Lawyers should ask it of every tool: does the data stay with a provider, pass through a legal platform that calls another company's model, or stay inside the organisation?

Does the level of the tool decide the return?

The return depends on the work as a whole, and the levels describe only how much of the work a tool holds. The work as a whole means how long a request takes from start to finish, how much lawyer time it uses, and how often it goes wrong. A tool can hold a whole workflow and leave all three unchanged. What changes them is a different process, and a workbench is often the best place to run one. The research on AI value supports this order.

None of these studies compares chat assistants, workflow tools and workbenches directly. They show that the change in the work goes with the return. The tool, workbench included, is how the change is carried out. More detail is in Why AI pays off only when the work is redesigned.

What happens when a workbench is built around the old process?

The review gets faster and the process stays slow. Take an invented in-house team that reviews about 40 non-disclosure agreements (NDAs) a month. Requests arrive by email, often without the counterparty or the purpose. They wait about two days for a lawyer, who reads the NDA, emails the business for the missing details, reads it again and replies.

Exhibit 3

The same workbench pays off only on the redesigned process

An invented NDA process, three ways. Figures are illustrative.

Working days from request to reply

• Old processabout 4
• Workbench on the old processabout 4
• Redesigned, then run in a workbenchabout 1

Lawyer minutes per NDA

• Old process90
• Workbench on the old process30
• Redesigned, then run in a workbench20

The workbench cuts the review from 90 minutes to 30. The business still waits two days in the queue, still gets an email asking for the counterparty's name, and still gets its reply on about the fourth day. The review is faster; the steps that should not exist are still there. Lawyers who supervise juniors know the pattern: a quicker first draft does little when the instructions arrived without the facts.

The workbench also adds costs a chat assistant does not have:

• someone has to build it, test it after every change and pay for it;
• someone has to keep its access and security up to date, and the closer it runs to the organisation's own environment, the more of that falls on the organisation;
• a lawyer still has to check its output. Being built for legal work does not make a tool accurate: leading AI legal research tools gave incorrect answers 17% to 33% of the time in a Stanford-led study [5].

A lawyer checks a workbench's output the way she checks a junior's markup: line by line, against the team's standard positions.

Redesign first and the result is different. A short request form asks for the counterparty, the purpose and the term, so the email exchange goes. The tool lists the clauses that depart from the team's standard positions, and the lawyer reviews only those. Requests arrive complete and take less lawyer time, so the queue shortens. This is the job the workbench does: the new process has a request form, a comparison against standard positions, a list of departures and a rule for when to escalate, and a chat window cannot hold those together. The redesign sets what the process should be, and the workbench runs it.

When is a workbench the right answer?

Often. Redesigned legal work usually has several documents, people and steps, and that is what a workbench is for. Still, ask how much the tool needs to hold after the redesign, because some redesigned work is simpler than it looked.

The redesigned work involves	Often enough
One document and one person, done the same way each time	A saved prompt or template in a chat assistant
A repeated procedure with a checklist	A saved skill in a workflow tool
Several documents, people and steps, with a record of who decided what	A workbench

A lower level is cheaper to build, quicker to change and easier to test. A workbench is worth its cost when the redesigned work needs one, and on our course most groups find that theirs does.

The order we teach

1. Map the workAs it runs today, with times.
2. Redesign itRemove the steps that add nothing; decide which steps a tool takes and where a lawyer checks.
3. Decide what the tool holdsA saved prompt, a skill or a workbench.
4. Decide where it runsAnd where the data goes at each step.
5. Test, change, test againThen try it on real work alongside the current process, with a lawyer checking every output.

Steps 3 and 4 are the two questions in our framework, and his three levels fit inside step 3. A workbench built after steps 1 and 2 is the combination we teach. Starting with the workbench puts the tool before the work.

Where we could be wrong

The case for starting with the tool has a fair point. In our opinion, a team whose process is already sound may get most of its return from a good workbench alone, and for that team the post's order works. A team that waits for a perfect redesign before building anything also learns nothing from real use.

The evidence has limits we want to state plainly. No study we found compares the three levels on the same legal work. Our view is an inference from research on redesign, mostly surveys, much of it from firms that sell process redesign. We have a stake too: our course teaches redesign. The post's claim is also an inference, and its author's consultancy designs and builds AI systems for law firms [1]. A legal team can test it for itself: measure one process before and after, and see what changed.

What we'd ask

If you lead a legal team or a firm:

• Redesign the work before choosing the tool: map it as it runs today, with times, and remove the steps that add nothing.
• Pick the cheapest tool that fits the redesigned work and your confidentiality rules. A saved prompt or a skill is sometimes enough.
• Budget for what a workbench costs after it is built: testing after every change, and keeping its access and security up to date.
• Measure one process before and after, and see what changed.

If you work in a legal team:

• Ask of every tool where the data goes at each step.
• Check every output. Being built for legal work does not make a tool accurate.
• Try a new tool on real work alongside the current process before relying on it.

Why we built the course

The return from legal AI goes to teams that redesign their work and then build the tool to run it, and someone in each team has to do both. We built AI Women in Law so that more of the people doing it are women lawyers. Over five weeks, participants redesign one piece of their team's work and build a first version of a tool for it. See the course.

References

1. Uwais Iqbal, There are three levels of AI maturity in legal, LinkedIn post, 21 October 2026. Background: British Legal IT Forum speaker profile.
2. McKinsey & Company, The state of AI: How organizations are rewiring to capture value, 12 March 2025. 1,491 respondents, surveyed July 2024.
3. BCG, AI at Work: Why strategy matters more than tools, 2026.
4. Bain & Company, From Pilots to Payoff: Generative AI in Software Development (Technology Report 2025), 23 September 2025.
5. Magesh, Surani, Dahl, Suzgun, Manning and Ho, Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, Journal of Empirical Legal Studies 22(2), 2025.