Comparisons
AI Agents vs CLM Software: Replace, Layer or Ignore?
AI agents vs CLM software for in-house teams: what Ironclad-style CLMs do that agents cannot, where agents win, how MCP connects them, and three architectures.
On this page
Keep the CLM as the system of record and put agents in front of it. That is the answer for most in-house teams in 2026. A contract lifecycle management system holds the repository, the approval routing, the signatures and the audit trail; an agent reads unstructured contracts, redlines against a playbook and drafts, and it does those things better than any CLM's native AI. Connect the two through a vendor connector rather than choosing between them.
The question, reframed
Jason Boehmig co-founded Ironclad, the CLM that secondary sources put at about $200m in annual recurring revenue. In June 2026 he left to lead legal at OpenAI. By then Ironclad was already a connector inside Anthropic's Claude for Legal, and Ironclad itself had shipped Intake and Redlining agents (13 November 2025) followed by "Jurist", a set of Manager, Drafting, Editing, Review and Research agents that moved inside Word in June 2026. The CLM vendor was building agents; the agent vendors were connecting to CLMs; and the CLM's co-founder went to work for the model company.
So "agents versus CLM" is the wrong frame. The categories are converging from both ends, and the question an in-house team actually faces is which system is the record, which system does the reading, and how the two talk. Gartner's warning about "agent washing", that only around 130 vendors of the thousands claiming agents actually ship them, applies to CLM marketing as much as to anyone's; the same Gartner note predicts more than 40% of agentic projects cancelled by the end of 2027. The definition of an AI agent on this site is the test to apply before believing either category's brochure.
What a CLM does that an agent cannot
| Function | What the CLM does | Why an agent does not replace it |
|---|---|---|
| Repository and metadata | One store of executed contracts with parties, dates, values and clauses indexed | An agent has no persistent state by default; GC AI's review of the Claude plugin listed "context resets per session" as a gap |
| Templates and clause library | Approved starting points, version-controlled | An agent can draft from a template but does not govern which template is current |
| Approval routing | Delegation-of-authority matrix, escalation, sign-off records | A probabilistic system should not decide who may approve; Anthropic's own plugins gate board consents behind a human review step for this reason |
| E-signature | Integrated execution, sealed copies | Agents send drafts; signing belongs in a system built to evidence it |
| Obligations and renewals | Calendar, alerts, reporting | Scheduled agents can watch renewals, and Claude for Legal ships one, but the source of truth for dates still has to be a record |
| Audit trail | Who changed what, when, in which version | Anthropic's Compliance API keeps six years of session transcripts but omits thinking blocks and MCP configurations; a CLM's version history is designed for the auditor |
| Blast radius | Bounded by role permissions | In February 2026 a venture capitalist let Claude Cowork "organise" a desktop and an rm -rf removed 15 years of photos; Super.com's general counsel put the in-house version plainly: "If you give it access to start deleting files, that's concerning" |
None of this is an argument against agents. It is an argument for keeping a deterministic system of record underneath them, whether that is Ironclad, Agiloft, DocuSign CLM, or, for a small team, a permissioned document store and a spreadsheet you actually maintain.
What agents do that a CLM never did well
CLMs historically templated, routed and stored. They did not read. That is the gap agents fill, and the 2026 product releases are all variations on it.
Anthropic's commercial-legal plugin reviews MSAs, NDAs and SaaS agreements against a playbook and writes the result to Word tracked changes, keeps amendment history, runs a 90-day renewal tracker, and ships renewal-watcher, deal-debrief and playbook-monitor agents; the renewal-watcher is one of five that can run headless as a Managed Agent. Legora's aOS, launched on 7 May 2026, promises overnight counterparty redlines "reviewed, flagged and responded to by the time lawyers arrive". Spellbook's Autonomous Contract Management, in early access since 30 June 2026, has agents review and redline documents that arrive through DMS and email integrations. LegalOn's five agents (February 2026; 9,000-plus customers) include a Triage agent that auto-approves low-risk contracts and escalates the rest. Harvey's Contract Intelligence has extracted 20 million-plus terms and Thomson Reuters' Tabular Analysis runs 10,000 documents against 100 questions.
The quality is real and it is not finished. Anthropic reports Opus 5's first-turn NDA redlines at nearly double Opus 4.8; GC AI's RedlineBench rose to 57.0 with Fable 5.1. But in Vals' February 2025 study lawyers still beat Harvey on redlining, 79.7% to 65.0%, and the contract review page explains why a playbook matters more than the model. Two Reddit data points from 2026 bracket the range of experience: one lawyer reported that "our firm was able to replace Harvey and Ironclad with Claude code", and the reply called that "shadow AI chaos with no governance, no consistency, and no audit trail"; an in-house lawyer without budget running Copilot and Power Automate predicted that "CLMs will have a hard time renewing" and named token usage as the main problem. Both are right about something.
How they connect: MCP, connectors and the confirmation step
The join between an agent and a CLM is a connector, and in 2026 that increasingly means the Model Context Protocol. Anthropic donated MCP to the Linux Foundation's Agentic AI Foundation on 9 December 2025; the 2026-07-28 specification made it stateless, hardened its OAuth and added "multi round-trip requests" so that a server can pause mid-call and ask the user to confirm before it acts, which is exactly the primitive a contract workflow needs before an agent writes to the record. Claude, ChatGPT, Gemini and Copilot Studio are all first-party clients.
For contracts specifically, Ironclad and DocuSign are among the 20-plus connectors in Claude for Legal, Zapier's MCP and SDK expose 9,000-plus apps and Zapier's own legal team runs intake, triage ticketing and AI-assisted contract analysis on it, and iManage's and NetDocuments' MCP servers run under the user's own credentials so ethical walls and matter restrictions apply automatically.
The caveat is security, and it is not theoretical. A June 2026 audit of 2,614 public MCP servers found 82% vulnerable to path traversal, 43% to command injection, 53% using static API keys against 8.5% using OAuth, and 492 exposed with no authentication at all; one compromised server in a five-server setup gave a 78.3% attack success rate. The rule for anything touching signed contracts: use the CLM vendor's own maintained connector with OAuth, never a community-built server, and read the agent security guide before you switch it on.
Three architectures for an in-house team
| Architecture | What it looks like | Fits | Main risk | Cost shape |
|---|---|---|---|---|
| A. CLM-centred | Ironclad, Agiloft or DocuSign CLM as the workspace, using the vendor's native agents (for example Ironclad's Intake, Redlining and Jurist agents) | Teams with a mature CLM, high volumes of standard paper and a delegation-of-authority policy to enforce | Locked to the CLM vendor's model choices and release cadence; native AI may lag the frontier by a generation | CLM licence plus AI add-on |
| B. Agent-first, CLM as record | Claude for Legal's commercial plugin, Harvey, Legora or Spellbook ACM does the reading and redlining; the CLM stores executed contracts, approvals and obligations, connected by a vendor MCP connector with confirmation steps | Teams that want the best available model on review and drafting while keeping an existing CLM's audit trail | Two vendors and a connector to govern; who approved what must be visible in the record, not only in a chat transcript | CLM licence plus agent seats or tokens |
| C. No CLM | A permissioned document store (SharePoint, iManage, NetDocuments), an e-signature tool, a playbook-driven review agent and a scheduled renewal-watcher with a human gate; automation glue in n8n, Zapier or Make | Small teams under a few hundred contracts a year, or a legal function of one to three people | No system of record unless you build the discipline yourself; audit gap; "shadow AI" if lawyers improvise | General-tier seats plus automation tooling; a three-lawyer German firm runs its intake and deadline capture on self-hosted n8n for €20–30 a month |
Architecture B is the default recommendation for a team that already has a CLM, because it keeps the record deterministic and lets the reading layer change as the models change. Architecture C is legitimate for small teams and is where the Zapier and Make guide and the in-house guide start; Deloitte's June 2026 survey found 61% of legal departments piloting agentic AI and predicts agents could handle 30% of in-house work within three to five years, which is the volume that eventually forces a team from C to B.
Governance you must keep whichever you choose
Four controls survive every architecture.
- An approval gate before anything leaves. Anthropic's plugins refuse to send a demand letter until a human confirms; Harvey's agents "flag moments requiring user confirmation"; Lexis' Legal Intelligence Engine presents an executable plan for approval before it runs; ChatGPT Work has Plan mode and approval gates. The human-in-the-loop entry explains where the gate goes.
- An audit trail in the record, not only in the chat. Anthropic's Compliance API retains transcripts six years but not MCP configurations; a CLM's version history is the artefact an auditor or a regulator will ask for.
- Hard limits the agent cannot override. Zenity's 2026 research on zero-click hijacking of browser agents concluded that "the fix isn't to keep asking the AI to behave. It's to put hard limits around the agent that the agent cannot override": no send permission, no delete permission, no signature authority. The OWASP Top 10 for Agentic Applications lists goal hijack, tool misuse and memory poisoning as the first, second and sixth risks, and a contract inbox is a delivery channel for all three; the prompt injection entry covers the mechanism.
- A liability review. Clifford Chance's briefing "Agentic AI: the liability gap your contracts may not cover" (9 February 2026) flags agent authority, hallucinated commitments, indemnities and insurance as gaps in most vendor and customer paper; if an agent negotiates, someone has to decide in advance what it is authorised to concede.
A 90-day plan
- Days 1–30. Inventory the contract types by volume and risk, and write a one-page playbook for the top three. Pick the architecture from the table above and name an owner for each agent you intend to run.
- Days 31–60. Connect one CLM or DMS object through the vendor's OAuth connector, with write access off. Run the review agent on fifty closed contracts against the playbook and score it on positions applied, positions invented and whether it stopped when unsure. Module 4 of the AI Lab for Lawyers course builds a playbook-review workflow with a confirmation step live, in n8n and in Claude, which is the fastest way to see what a gate looks like in practice.
- Days 61–90. Turn on one scheduled agent, usually a renewal-watcher, that reads the record and emails a human 90 days before each renewal; give it no permission to change anything. Measure hours saved and errors caught, and decide whether to extend write access to a second, narrowly scoped task.
Where this leaves you
The CLM is not obsolete and the agent is not optional. A CLM without agents leaves the reading to lawyers, which is where the hours go; agents without a CLM leave the record to chance, which is where the audit findings come from. Layer them, connect them through a maintained connector with a confirmation step, keep a human on every send and every signature, and revisit the architecture in a year, when the CLM vendors' own agents and the labs' plugins will have converged further still. The NDA review at scale workflow is a worked example of Architecture B end to end.
Frequently asked questions
Can AI agents replace contract lifecycle management software?
Not for the functions a CLM exists for. Agents do not reliably keep a single source of truth, route approvals under a delegation-of-authority matrix, integrate e-signature, maintain an obligations calendar or produce a version-controlled audit trail. They do read, review, redline and draft unstructured contracts better than a CLM's native tools. For most in-house teams the answer is to layer agents on the CLM through connectors, and to drop the CLM only if the team is small.
What do AI agents do better than a CLM?
Unstructured work at scale. As of September 2026, Anthropic's commercial-legal plugin reviews MSAs, NDAs and SaaS terms to Word tracked changes against a playbook; Legora's aOS answers overnight counterparty redlines before lawyers arrive; Spellbook's Autonomous Contract Management reviews documents that arrive by email or DMS; LegalOn's Triage agent auto-approves low-risk contracts and escalates the rest; Harvey has extracted more than 20 million contract terms. CLMs historically templated and routed; they did not read.
How do AI agents connect to Ironclad, Agiloft or DocuSign CLM?
Through vendor-maintained connectors, increasingly built on the Model Context Protocol. Ironclad and DocuSign are among the 20-plus connectors in Claude for Legal, Zapier exposes 9,000-plus apps through its MCP and SDK, and Ironclad's own Jurist agents run inside its platform and Word. The July 2026 MCP specification adds mid-call confirmations, so a server can pause and ask a human before it writes to the record. Avoid community-built servers for anything touching signed contracts.
Should a small in-house team buy a CLM or just use AI agents?
Below a few hundred contracts a year, a team can run without a CLM: a document store with permissions, an e-signature tool, a playbook-driven review agent and a scheduled renewal-watcher that emails a human 90 days before each renewal. The risks are the absence of a system of record and audit trail, so write down where executed contracts live and who approves what. Above that volume, or with a delegation-of-authority policy to enforce, buy the CLM.
What governance do AI agents need in contract work?
An approval gate before anything is sent, filed or signed; an audit trail that records what the agent read and changed; hard limits the agent cannot override, such as no send permission and no delete permission; and a named owner for each agent. Clifford Chance's February 2026 briefing warns that agent authority, hallucinated commitments, indemnities and insurance are gaps most contracts do not yet cover. Treat every agent output as a draft for a lawyer.
Put this into practice
AI Lab for Lawyers: a live, hands-on course for practising lawyers
4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.
Related reading
Use Cases
AI Contract Review: How Agents Redline Against a Playbook
AI contract review works when an agent applies your playbook of positions, fallbacks and escalations. Tools, accuracy numbers, setup and failure modes for 2026.
8 min read
Guides
AI Agents for In-House Legal Teams: A Practical Guide
AI for in-house counsel in 2026: the agents legal departments run, how to govern a Copilot stack, how to win AI savings from law firms, and a 12-month roadmap.
9 min read
Glossary
Model Context Protocol (MCP) — Definition for Lawyers
Model Context Protocol (MCP) is the open standard that lets Claude, ChatGPT and Copilot read your DMS or docket feed through one connector. What a server exposes and what to check.
4 min read
Guides
Agent Security for Lawyers: Prompt Injection, Data Leaks and Guardrails
AI agent security for law firms: prompt injection via hostile documents, the Grok Build and Cowork incidents, and the allowlists and gates that contain them.
11 min read
Playbooks
Automate Contract Review with a Playbook: Positions, Fallbacks and Escalations
A contract review playbook an AI agent can apply: ten sample rows, encoding for Claude, Spellbook and Word's Legal Agent, plus a 20-contract calibration method.
12 min read
Tools
Claude for Legal Plugins Explained: The 12 Practice-Area Plugins and 20+ Connectors
Claude for Legal gives paid Claude users 12 open-source practice-area plugins, 150+ skills and 20+ connectors to iManage, NetDocuments and CoCounsel.
8 min read