TL;DR
Four things share the name: assistants, copilots, agentic workflows and decision intelligence. Only two change how work gets done.
Four different things share the name. Assistants, copilots, agentic workflows and decision intelligence. Only two of them change how work gets done — and the difference is whether the AI sits beside the work or inside it.
Get that distinction wrong and you buy a very capable thing that leaves your process exactly as it was.
So what are the four kinds?
| Kind | Where it sits | What it changes |
|---|---|---|
| Assistant | A separate chat window | Your speed. You ask, it answers, you act |
| Copilot | Inside the application you already use | Your speed, with less switching. Suggests in place |
| Agentic workflow | Inside the process itself | The process. It executes steps and escalates exceptions |
| Decision intelligence | On the data behind the process | What gets decided. It surfaces the choice, with reasoning |
The first two make a person faster. The second two change what the organisation does.
That is not a quality ranking — a good assistant is genuinely useful and cheap to adopt. It is a ranking of consequence. If your problem is that a process takes eleven days, a faster person inside that process changes the days very little.
What do assistants and copilots actually do well?
Compress the individual task. Draft a clause, summarise a contract, answer a research question, explain an unfamiliar provision.
They are the right purchase when the constraint is an individual's throughput — and they are the easiest thing in this market to adopt, because nothing about your process has to change.
Their ceiling is that nothing about your process changes. Three consequences follow:
- The output still needs somewhere to go. A summarised contract is a better artefact in the same queue.
- Consistency does not improve. Two people using the same assistant still reach different conclusions, because the assistant has no view on your positions.
- Nothing accumulates. Each session starts cold. There is no record that makes the next one shorter.
A copilot fixes the switching cost of an assistant and inherits everything else — which is still worth having, and is not a process change.
What makes a workflow "agentic", and why does it matter?
An agentic workflow does not answer questions. It performs steps and stops when it should.
The distinction that matters operationally is not autonomy — it is whether the system holds the process. An agentic workflow has four properties an assistant does not:
- It is triggered by an event, not by a person opening a window. A contract arrives; the work starts.
- It executes a sequence — read, match against your rules, score, propose, route — rather than answering one prompt.
- It knows what it cannot decide, and escalates on a defined trigger rather than on a confidence number.
- It leaves a record that is queryable afterwards — what it saw, which rule applied, what it proposed.
That fourth property is what turns work into data, and it is why agentic workflows compound while assistants do not.
And the honest constraint: an agentic workflow requires that your rules exist. It runs against a playbook of positions, not against general legal knowledge — so the prerequisite is a decision exercise, not a software purchase.
What is decision intelligence, in legal terms?
Surfacing the decision, with the reasoning attached, from data the organisation already has.
Not "here is a summary" and not "here is the answer". *It is "this is the choice in front of you, here is what it depends on, and here is what we did last time."***
In legal operations that looks concrete rather than abstract:
- Which of these renewals should we let lapse, given usage and spend?
- Where are we accepting terms we normally refuse, and with which counterparties?
- What is our aggregate exposure on a clause type across the portfolio?
None of those is answerable from a document. They need contracts turned into obligations and fields — which is why decision intelligence is downstream of the agentic workflow rather than an alternative to it. The workflow produces the data the decision needs.
Which kind do you actually need?
Ask what your constraint is. The answer follows directly:
| Your constraint | What to buy |
|---|---|
| One person is a bottleneck | Assistant or copilot. Cheapest, fastest to adopt |
| The same work is done inconsistently | Agentic workflow — it applies one set of rules |
| Work waits in queues between people | Agentic workflow — the process is the problem |
| You cannot answer questions about your own portfolio | Decision intelligence, and the workflow that feeds it |
| You do not know which of these you have | Measure first. Time the stages |
Most teams buy an assistant because it demonstrates well and adopt it happily — and their actual constraint is the third row. A faster reviewer and an unchanged queue produce an unchanged cycle time.
What does legal AI not do?
It does not decide, and it does not remove the reviewer. It moves them.
Three boundaries worth stating plainly, because the market states them rarely:
- Nothing here gives legal advice. A system applies rules a legal team wrote; it does not substitute for the judgement that wrote them.
- A workflow that produces work nobody reviews is worse than no workflow. Capacity to review has to exist before drafts arrive faster.
- None of it works on rules that were never written down. This is the honest gate on the whole category, and it is why our most common recommendation is to write the playbook first.
That is the shape of the market. Assistants and copilots make people faster. Agentic workflows and decision intelligence change what the organisation does — and both of the second pair depend on rules you have already decided.
About Cognilium Cognilium builds AI optimization apps for Microsoft Dynamics 365 — companion apps that optimize the pricing, inventory, warehouse and planning decisions your ERP manages but can't optimize. Dynamics is your system of record. Cognilium is your system of intelligence. https://cognilium.ai · https://www.linkedin.com/company/37180269/
Legal AI Ops. We transform legal workflows with agentic AI, copilots, agentic workflows and decision intelligence — built into core workflows rather than beside them, to raise productivity and cut operational overhead. Contract Review Copilot is the contract-review app in that family. It ships as Paralegent AI, in production today. How we build Legal AI Ops — custom AI capabilities on top of legal work, against your playbook and your Dynamics 365.
Not sure which of the four you need? Bring your constraint to a 15-minute call — the answer is often the cheapest of the four, and sometimes it is none of them yet.
Sources
No external source is cited. Every claim is our own definition or design position, labelled as such.
Sources and fact-check
| # | § | Claim | Tier | Source | Verdict |
|---|---|---|---|---|---|
| 1 | 1 | Four kinds share the name; two change how work gets done | T2 — ours, a market taxonomy stated as ours | Internal definition | PASS |
| 2 | 2 | Assistants and copilots compress the task and leave the process unchanged | T2 — ours. No vendor named | Internal definition | PASS |
| 3 | 2 | Three consequences — output still queues, consistency unchanged, nothing accumulates | T2 — ours | Internal definition | PASS |
| 4 | 3 | The four properties of an agentic workflow | T2 — ours, and it describes our own architecture | Profile §1, §2 | PASS |
| 5 | 3 | An agentic workflow requires rules that exist | T2 — ours, and it is the article's honest gate | Profile §4.1 | PASS |
| 6 | 4 | Decision intelligence surfaces the choice with reasoning, and is downstream of the workflow | T2 — ours | Internal definition | PASS |
| 7 | 5 | The constraint-to-purchase mapping | T2 — ours, judgement | Internal definition | PASS |
| 8 | 6 | Three boundaries — no legal advice, review capacity required, no unwritten rules | T2 — ours, and it is the disclosure boundary | 03-proof/DISCLOSURE-RULES.md | PASS |
Tier summary: 0 × T1, 8 × T2 — 0 × T4.
Entirely T2 and correctly so — this is a definitional article about a market category. It names no competitor, makes no comparative claim about a named product, and carries no figure of any kind: no throughput, no accuracy, no time saved, no customer count. "Eleven days" in §1 is illustrative and attached to nobody.
Disclosure: §6 states plainly that nothing here gives legal advice, that review capacity must exist first, and that none of it works without written rules. The CTA says the answer is "often the cheapest of the four, and sometimes none of them yet."
