AI Agent vs Chatbot: What Changes for Contract Review

An AI agent vs chatbot comparison usually gets reduced to a marketing slogan, but for a legal team the distinction is practical, not semantic. A chatbot answers the question you just asked and stops. An AI agent takes a goal such as "flag the deviations in this NDA against our playbook," breaks it into steps, pulls the clause library, compares terms, and comes back with a redlined draft, without you prompting it again at every step. That difference decides whether a tool saves you ten minutes or two hours on a Tuesday afternoon contract queue.

What actually separates an AI agent from a chatbot when you're reviewing a contract

Three things separate an agent from a chatbot, and none of them are about how "smart" the underlying model is.

Autonomy. A chatbot waits for your next message after every answer. An agent keeps working across several steps toward a goal you set once, adjusting its plan as it goes.

Tool use. A chatbot answers from what it was trained on, or from a document you paste into the window. An agent connects to your clause library, your contract repository, sometimes your CRM, and pulls what it actually needs instead of relying on your copy-paste.

Memory. A chatbot session resets, more or less, with every new context window. An agent carries what it found in step one into the decision it makes in step four, so a flagged limitation-of-liability clause actually informs the fallback language it drafts.

A short definition, for the record Chatbot: answers one question per prompt, no persistent task, no external tool access. AI agent: pursues a multi-step goal autonomously, uses external tools (clause library, repository, CRM), retains context across steps.

Voilà ce que cette distinction change concrètement: a chatbot is a very good research assistant for a single question. An agent is closer to a junior associate who has already read the whole file.

Where a chatbot still does the job better

Skip the agent when the task is genuinely a single question. "What does this indemnification clause mean in plain language?" "Summarize this three-page SLA." "Is this jurisdiction clause standard for a SaaS vendor contract in Switzerland?" A chatbot answers these faster than an agent would plan for them, because there is no multi-step task to plan. Before you sign, it vaut mieux savoir que routing a one-off question through an "agentic" tool usually adds latency and cost for zero extra accuracy.

A chatbot is also the safer default when you don't want a tool touching your live repository. If the task doesn't require reading five documents and cross-referencing a playbook, don't hand it the keys.

There's also a cost argument that rarely makes it into the sales pitch. Agent runs are typically credit-metered because they burn far more tokens per task than a single chatbot exchange: reading source documents, planning steps, calling tools, checking its own output. Routing a one-line question through an agent framework doesn't just add latency, it adds a bill you didn't need to pay.

What an agent adds once it can act on your contract, not just talk about it

En pratique, cela signifie que the agent stops being a chat window and starts being a workflow. Contract review is the case that shows the difference most clearly: an agent can take a counterparty's paper, map every incoming clause against your approved clause library, tag the deviations by risk level, and produce a redlined draft with suggested fallback language, all before you open the file yourself.

That's not a hypothetical. Legal operations teams already run this pattern for routine agreements: standard NDAs, order forms, renewal riders, anything with a predictable structure and a clear playbook. The agent does the first pass; a human reviews what gets flagged. Ironclad's breakdown of AI agents in legal operations puts it plainly: a chatbot answers one question per prompt, while an agent takes a goal, breaks it into steps, and adjusts its approach based on what it finds along the way.

The same logic applies past the review stage. An agent can watch a portfolio of executed contracts for renewal windows and opt-out deadlines and trigger an alert before one slips, something a chatbot simply cannot do because it has no persistent task to run in the background.

Agentwashing: why most tools sold to legal teams as "agents" are still chatbots

Ce n'est pas un risque théorique: c'est ce que les juristes voient régulièrement quand a vendor renames a chat feature "AI Agent" without adding autonomy, tool use, or memory to it. Industry analysts have started calling this pattern "agentwashing": relabeling an assistant as an agent when it still waits for a new prompt at every step and never operates on its own.

Forrester's 2026 research on agentic AI found that three-quarters of enterprise leaders say they're adopting agentic AI, but only a small minority have it running in meaningful production beyond what the firm bluntly calls "agentish chatbots": tools carrying agent branding with chatbot-level autonomy underneath. That gap between the label on the pricing page and the actual capability is exactly what a legal buyer needs to test before signing a contract for the tool itself.

For a legal buyer, the practical test is simple: ask the vendor what happens after step one without your input. If the answer is "you type the next prompt," you're looking at a chatbot with a new name. If the answer involves the tool reading your repository, comparing clauses, and producing a deliverable across multiple steps, that's an agent.

Where Harvey, Spellbook, Kira Systems and Legalysis sit on the chatbot-to-agent spectrum

None of these tools sit at a single fixed point; most combine both modes depending on the task.

Harvey leans toward agent behavior for research-heavy workflows: multi-document review, memo drafting that pulls from case law and firm precedent across several steps. Spellbook works largely inside Word as a drafting copilot, closer to the chatbot end for most everyday tasks, though its clause suggestions increasingly pull from a firm's own playbook automatically. Kira Systems built its reputation on extraction at scale across large document sets, which is agent-shaped work (multi-document, tool-connected, no per-clause prompting) even though the interface reads more like a dashboard than a chat window.

Legalysis sits closer to the agent end for its core task: clause-by-clause contract analysis that reads the whole document, flags risk by clause type, and explains what each flagged term means in practice, without the analyst re-prompting for every clause. Where it doesn't try to compete is generative drafting from scratch or case-law research; that's still better served by tools built specifically for it.

None of these four tools are interchangeable, and a comparison table that scores them all on the same axis (speed, accuracy, price) misses the more useful question: which stage of the contract lifecycle is each one actually built for. Kira and Legalysis both concentrate on the review side, before signature, where the cost of missing a clause is highest. Harvey spreads across research and drafting. Spellbook stays close to the drafting desk itself. Fréquemment dans les SLA tech, teams end up running two of these tools in parallel rather than picking a single winner, because review and drafting are genuinely different jobs with different risk profiles.

What to check before you let an agent touch a live contract

Trois types de vérifications méritent une relecture attentive avant d'accorder un accès à un agent.

Data handling. Confirm the agent's underlying model provider runs on a zero-retention policy for your inputs, and that your contract data isn't used to train someone else's model. Ask about sub-processors explicitly; the answer is rarely volunteered.

Human checkpoints. A well-built agent workflow drafts or flags, a human approves or overrides, and the system logs both actions. If a vendor pitches full autonomy on anything above a standard NDA, that's a red flag, not a feature.

Audit trail. Every clause the agent suggests and every routing decision it makes should be logged in a record you can pull during a dispute or an internal review. Without that, you're trusting a black box with documents that carry real liability.

None of this replaces your own legal judgment on anything material. À vérifier avec votre conseil quand l'enjeu dépasse une clause standard: an agent narrows what needs your attention, it doesn't decide what's acceptable risk for your business.

Should your team use a chatbot, an agent, or both right now

Most legal teams don't need to pick one and abandon the other. A chatbot stays useful for one-off questions, quick summaries, and plain-language explanations of a clause you're staring at right now. An agent earns its place on anything repetitive: standard NDA triage, vendor SLA comparison against your playbook, renewal-date monitoring across a large contract portfolio.

Three concrete moves worth making this quarter:

Start with your highest-volume, lowest-risk contract type, standard NDAs are the usual choice, and measure whether an agent workflow actually cuts turnaround time before expanding it to anything more complex.

Ask every "AI agent" vendor on your shortlist to demonstrate a full multi-step run on one of your own documents, not a demo script, before you sign anything.

Keep a chatbot in the stack for the questions an agent shouldn't be trusted with alone: nuanced jurisdiction calls, anything novel enough that there's no playbook entry to compare against yet.