Every company has contracts. Almost no company has a CLM.
Some businesses are contract-heavy, some are contract-light, but almost none have no contracts. At a minimum, they need to handle purchase orders, general terms, and a handful of employment or service agreements.
For a while, a shared drive is fine. Then somebody asks which agreements are governed by the laws of England and Wales, which expire this year, or which supply contracts carry an uncapped indemnity, and a folder structure stops being an answer. That is the problem contract lifecycle management software was built to solve, and it solved it: categorise contracts, track obligations, run deadlines, keep a hundred vendors on one template rather than a hundred. Underneath sits the point most people skip. A governed contract portfolio is worth more than one that isn't. Rights you cannot find are rights you do not exercise.
So the case for CLM is sound. It just never really won.
Roughly one company in ten runs a CLM company-wide. Most of them have a love-hate relationship with it. The rest are on SharePoint, Google Drive and Dropbox. Shared drives are far cheaper, and they work about as badly as a poorly maintained, complex and overpriced CLM.
The true cost of CLM is data entry, not licenses.
Every incoming contract has to be uploaded, tagged, categorised, linked to its parent and updated when it is amended. Do that diligently and you have a genuine single source of truth. Miss it once and something worse than an error happens: the user learns to check. The first time somebody opens the underlying PDF because they did not trust the field, the field is finished. Trust in a system of record is binary, and it does not come back.
Metadata fails in four predictable ways, and each one is human. Somebody keys it in wrong. Two teams tag the same thing differently. The record is captured once at signature and never touched again. And nobody ever goes back to check whether a field is still true. The consequences are the ones CLM users complain about constantly: obligations that are hard to track, related documents that are hard to link, contracts that are hard to find inside the CLM itself, and people quietly working around the system.
Then the deployment. Migration, template preparation, integrations, workflow redesign, training, and deciding who enters what and who chases whom. Every time an in-house colleague has told me about a CLM implementation, it has sounded like a horror story. Timelines that slipped and slipped. Budgets that went well past the number anyone signed off on. Whole activities pushed back outside the system because the system would not let them do it the way the business actually worked. And the less technical members of the team quietly wondering whether they should leave, or change profession altogether.
So CLM became large company software. It works if you can afford the discipline. The bitter joke is that the discipline required to make a CLM worth having is the same discipline whose absence made you want one.
A CLM with an AI add-on is still the same CLM, with the same problems
Through 2025 and 2026, every legacy leader in the category shipped their AI upgrade: conversational front ends, clause extraction, automated metadata capture on upload. It helped. It extended the architecture's life expectancy rather than changing it.
These assistants live inside the system, and they are reactive: you go to them, you ask, they answer. CLMs do have workflow features, but those workflows are built once and rigid, so the process bends to the software rather than the other way round. Operators are compelled to follow the configured path even when they can see a better one. And the real process runs past the boundary anyway, into other tools, other people, and now other AI assistants and agents. The board approving the deal does not log in. Neither does the procurement lead doing the negotiating. An assistant inside the CLM cannot run a process that mostly happens outside it. That takes an agent sitting outside the CLM. Meanwhile, at a moment when everyone already has three chat windows they did not ask for, it was one more application and one more login, oblivious to everything the person had been doing everywhere else.
Systems of record exist to translate documents into structured, human-readable data, because no one can read four thousand contracts to answer one question. Adding AI made that translation faster. It did not ask whether the translation was still needed.
The contract is the record, not the database
A human needs the table. A machine does not. An agent will happily open the relevant fifty agreements and read the indemnity clause in each, because it takes seconds rather than a fortnight.
Which exposes the category error. The CLM treated the derived table as the system of record. It never was. The contract was always the record. The table was a workaround for human reading speed, and we mistook it for the truth.
Two things follow.
Data entry dies. Nobody should be entering a contract into the CLM, any more than anybody should be entering a prospect into the CRM. The agent does it, or does not need to. It may keep an index, and at real volume it should, because sweeping an entire corpus on every question is neither instant nor free. But it maintains that index itself, and it can always go back to the document to check. That is the difference between a database that is authoritative because somebody typed carefully and an index that is disposable because the source is one step away. It also answers the trust problem that killed the CLM. A field can drift from the contract because it is a copy. An answer read out of the contract cannot.
And the interface built on top of that database dies with it. The CLM’s screens exist because a table underneath needs viewing, filtering, and manipulating. Dashboards, saved views, report builders: all of it is downstream of the assumption that a human has to go somewhere and look. Once you can simply ask, none of it has a reason to exist. A dashboard becomes something you request, in the shape you want, on the day you want it, not a legacy screen that constrains how you are allowed to ask.
The record survives all of this, and gets better. What dies is the obligation to feed it by hand and the screen you had to visit to use it.
Your agent might read a contract, or query a CLM. That is her problem, not yours
We get asked where an agent like this sits: a new category, a CLM substitute, or somewhere on a quadrant with other legal AI tools. It is the wrong question.
Before agentic AI, software was something you gave people so they could work faster and so you could see what was happening. Then assistive AI made individuals dramatically more productive, which is why so many legal ops leaders quietly stopped delegating downwards. Both are questions about equipping people. Agentic AI is not. We assume enterprises will move beyond assistive intelligence and implement solutions that commit to workflow results instead. People will steer the system and supervise execution, rather than doing the work themselves.
So the buying question changes. It used to be: which tools do I give my team, so that when I ask which contracts expire this year, somebody can answer. Now it is: which work would I be happy never to do again, and does my agent have the access it needs to do it? Notice that the second question does not mention a CLM. Your agent might read the contracts. She might query a CLM. She might go and ask the person in procurement who already knows. That is its problem, not yours, in exactly the same way it is a junior associate’s problem and not yours. You never asked the associate which system they used. You asked for the answer.
An assistant answers the question. A coworker does the work
Do not evaluate an agent by how well it answers in a chat window. Evaluate it by whether you can hand it a piece of work and get the finished thing back.
Assign it a task in the tool you already use. Forward it the email. Then see whether the renewal notice actually goes out, whether the vendor questionnaire comes back completed, whether procurement’s request is resolved without anybody in legal touching it, and whether every step is on the record afterwards.
That is what a coworker does. You don't ask a colleague which internal systems they used, and you don't buy them a licence per screen. You give them work.
Arna is your AI legal coworker. Your lawyers instruct her once: which template, who to loop in, who signs off. Then anyone in the company can assign her a task, and she runs the process start to finish inside the tools you already use, with every step on the record. If you are weighing a CLM implementation, or wondering why the one you have isn't being used, that is the conversation we would like to have.