A good number of US law firms still run their talent data on systems installed years ago. Usually an HR module sitting alongside practice management, configured during an implementation project that nobody involved remembers fondly, and upgraded rarely because an upgrade is another project.

The data is almost always there. Review scores, engagement responses, tenure, practice group, hours. The question is not whether the firm holds it. It is what the firm can ask of it, and how long an answer takes.

What on-prem actually constrains

The constraint people expect is speed, and that is the least of it. The real ones are structural.

The instrument is frozen at the point it was configured. A system upgraded every three to five years holds the questions that mattered when somebody specified it. Adding a question about supervision quality, or splitting a cohort differently, becomes a change request with a cost and a queue position. So firms stop asking new questions, and the survey slowly becomes a measure of last decade's concerns.

History gets lost at the seams. Migrations and archive policies are where five years of comparable data goes. Year-on-year movement is the most useful thing a talent dataset holds, and it is also the first thing to break when a system is replaced or a retention schedule fires.

Reporting is somebody's side project. When the route to an answer is an export into a spreadsheet, analysis depends on one person having the time and the skill. That person leaves eventually, and the analysis stops being produced without anyone deciding to stop producing it.

Integration is custom work, so it does not happen. Connecting review data to matter and billing data is a piece of work on every system pair, which means it gets scoped, deferred, and left out of the business case.

On-prem has real arguments in its favor and they should be said plainly. The security boundary is known and the firm controls it. There is no vendor dependency and no renewal risk. For a firm with strong internal IT and a stable set of questions, that can be the right trade. The cost is not inefficiency. It is that the system answers the questions it was built for and resists the rest.

What changes with cloud, and it is not AI first

The first gain is cadence. An instrument that can be reconfigured between cycles lets a firm ask this year's question this year, which matters because the thing a firm needs to measure does change. A firm that lost three fourth-years to the same practice group has a specific question, and a system where that question costs a change request will not get it asked.

The second is that history stays in one place. Comparing this cycle to the last four becomes a query rather than a reconstruction from archived exports. A firm with the history in one place can see that engagement fell in one cohort for two cycles running, rather than only that it fell.

The third is that integration becomes routine. Review data read against matter and billing data tells you who actually supervised whom, rather than who remembered working with whom when asked. SRA integrates with billing systems either directly or through middleware, and the practical effect is that rater pools come off the record rather than off a partner's recollection.

Where AI earns its place

Specifics matter more than enthusiasm here, because the phrase covers both a model that tidies a sentence and a model that decides who gets promoted.

The work that pays off first is written feedback. Vague and inconsistent evaluation language is the most common defect in a review cycle, and catching it while the reviewer is drafting is worth more than catching it in a calibration meeting six weeks later. Artemis does AI-assisted drafting, consistency checks on written feedback, and AI-assisted review of reports and feedback before they reach anyone.

The second is reading at volume. Matter debriefs and account health across a whole client relationship, free-text responses across a firm, and the patterns in them that nobody reading sequentially will hold in their head. Artemis derives insights from that material and carries them into goals and development plans, which is the part that most often falls over when it is left to a document and a calendar reminder.

What still needs a person is the decision a career turns on. A model can put it in front of leadership that three cohorts raised the same supervision problem, and that is useful. Whether it becomes a development conversation, a staffing change or a compensation decision is a human call with an author who can be asked about it. The distinction worth holding onto is between AI that assembles the evidence and AI that makes the decision, because only one of those can be appealed.

One question is worth putting to any cloud vendor before the security review reaches encryption: does our data train your models? Artemis’s answer is never. A firm that cannot get a clear answer to that question has learned something about the vendor.

How this reaches retention

The chain from system to retention is specific, and it runs through three things rather than through better dashboards.

Earlier detection. Attrition shows up in engagement and supervision data before it shows up in resignations. SRA's benchmark puts engagement at 82 in year one, 68 in year two, 47 across years three and four, and 55 across years five and six. The trough is not a surprise to anyone who has worked in a firm, but knowing which of your cohorts is in it this year, rather than which was in it last year, is the whole of the opportunity to act.

Attribution precise enough to act on. A firm-level engagement score tells leadership there is a problem. A cut by practice group and supervising partner tells them where, which turns a management committee discussion into a conversation with one partner about how they run a team. Reporting that cut safely needs threshold discipline, which is a design question rather than a software feature, and it is covered in the methodology.

Follow-up that survives the cycle. Development actions agreed in a readout and recorded in a document get revisited when someone remembers. The same actions held in the system, with an owner and a revisit date, get revisited because the system asks. Most of the value in a review program is in the part after the report, and that part is where an annual spreadsheet process tends to fail.

The growth half

Retention is the half firms ask about first. The same architecture carries two things that bear on growth.

Client feedback read across a relationship rather than gathered at the closing conversation with the partner who ran the matter. An account where responsiveness has been slipping for three matters is a renewal risk that is visible in a running record and invisible in a year-end survey.

And lateral integration measured while it can still be changed. The signals that predict whether a lateral partner hire works are behavioral and early: introductions promised and not made, work not shared, the sponsor who has stopped having the meetings. A system holding 30, 90 and 180 day check-ins surfaces those in the first year. A firm judging integration on whether the book followed finds out in year three, when the answer is already settled.

Questions worth asking before replacing anything

Five, and they are more useful than a feature comparison:

  • Can we change the instrument between cycles without raising a project?
  • Where does the raw response data sit, and who inside the firm can reach it?
  • Does our data train the vendor's models?
  • Can we get a cut by practice group and supervising partner without exporting to a spreadsheet, and what happens when a group is too small to report?
  • What happens to five years of comparable history when we migrate?

A vendor who answers the fourth one by explaining what they will withhold, and why, is describing a program worth buying. One who answers it by showing you a dashboard has not understood the question.