Scope & boundaries
This guide covers the role-specific assistants an employee invokes to do their own work — suite copilots against domain specialists, and the seat-versus-credit decision that determines whether next year's bill is forecastable.
It does not cover internal support: deflecting and resolving what employees ask IT and HR (AI Service Desk & Employee Support Agents), the retrieval and permission layer that decides what any assistant can see (Enterprise Search & RAG Platforms), the runtime that executes multi-step agent work without a person in the loop (AI Agent & Agentic AI Platforms), which foundation model to license, and on whose terms (Generative AI & LLM Platforms), or brand governance across everyone who writes, and the content workflow around it (Enterprise AI Writing & Content Generation).
Executive Summary
The question is not whether the suite copilot is as good as the specialist. It is whether the difference is worth a second seat on the same employee — and for most roles in most organizations, it is not.
Every major platform you already license now includes an assistant. Microsoft includes usage of agents published to Microsoft 365 Copilot in the Microsoft 365 Copilot license. SAP describes Joule as bringing assistants and agents into a unified workspace that runs end-to-end workflows. Salesforce and Google make the equivalent case for their own estates. Against that, a set of role specialists argue that general-purpose assistance is not good enough for legal research, revenue forecasting or regulated financial analysis, and for a narrow set of roles they are demonstrably right.
What makes this hard to buy is that the two camps are priced on incompatible units and both bills land on the same employee. Suite copilots are per seat, which is predictable and charges you for everyone whether or not they use it. Specialists and agent platforms increasingly meter consumption — Microsoft sells Copilot Studio as tenant-wide packs of 25,000 Copilot Credits at $200.00 per pack per month, Notion prices its AI agents at $10 per 1,000 credits after a free trial, Sierra prices on outcomes — which is fair and impossible to forecast in year one. Choosing well means deciding which of those two failure modes you would rather explain.
Why This Purchase Keeps Being Made Twice
Copilots arrived through two doors at once and organizations rarely noticed. The suite copilot came in through an existing enterprise agreement, often as an upgrade discussed with procurement rather than with the department that would use it. The role specialist came in through the department — legal, revenue, finance — bought on a departmental budget against a specific frustration. Both are now live, both are billed, and in most organizations nobody has compared them because they were never evaluated in the same room.
The overlap is real and rarely measured. An employee with a suite copilot, a horizontal work assistant and a role-specific tool has three subscriptions pointed at overlapping work, and the marginal one is often the one bought most recently and defended most loudly. Before adding a fourth, the useful exercise is unglamorous: list the assistants each department already pays for and ask which tasks each is genuinely better at. In most organizations that list has not been written down once.
There is also a governance dimension that gets discovered late. These tools read broadly by design, and the reach is the product — Glean describes specialized AI coworkers that use company context to take action across a customer's systems. That reach is exactly what makes them useful and exactly what makes an over-permissioned document store a live problem. The copilot does not create the permission error; it finds it, at scale, and surfaces it to whoever asked. Organizations that have never audited file permissions discover this in week two.
Which type of Role-Specific AI Copilots for the Enterprise fits your organization?
Nobody should build a general copilot, and the reason is not model access — that is a commodity. It is the connector and permission surface, which is the same reason the service desk category consolidated. What some organizations should build is a narrow agent for a workflow no vendor covers, and the platforms for that are now sold by the same companies selling the copilots. Microsoft states that the Copilot Credit pack and the pay-as-you-go meter differ only in how the customer pays, not in features or capability, which is a fair description of how these builder platforms are now packaged: the capability is settled, and what you choose is a billing posture.
The genuine sourcing decision is between four things, and the deciding factor is reach rather than quality. A suite copilot sees the suite. A horizontal assistant sees whatever you connect. A role specialist sees a curated corpus plus your documents, and is usually the only camp that treats citation as a first-class requirement. A builder platform sees whatever you wire it to, at the cost of somebody owning it.
| Camp | What you are buying | What it will not do |
|---|---|---|
| Suite copilot | An assistant inside the productivity or business suite you already license | Reach outside its own estate. The work that spans four vendors is the work it cannot see. |
| Horizontal work assistant | One assistant connected across every system, independent of any suite | Match a role specialist's depth in a regulated domain, or its citation discipline. |
| Role specialist | Domain-trained assistance for legal, revenue or financial work | Justify itself outside that role. It is a departmental purchase and should be priced as one. |
| Builder platform | The tooling to make agents for workflows nobody sells | Run itself. Every agent is software somebody maintains, and the meter runs whether or not it is good. |
| Nothing new | The copilot bundled into a license you already hold | Impress anyone. It is also the only option with no incremental cost, which is a real argument. |
How do you evaluate Role-Specific AI Copilots for the Enterprise?
Model quality is the wrong axis. The frontier models are close enough that in blind tests on ordinary knowledge work most users cannot rank them, and every vendor here is using one of the same handful. What differs is what the assistant can see, what it is allowed to do with what it sees, and whether it can show its work.
Four vectors matter once the demos end. Context reach is first and decides most outcomes: an assistant that cannot read the system where the answer lives will lose to one that can, regardless of model. Permission fidelity is second, and the question is whether access is evaluated at query time against live entitlements or against a synced index that lags. Provenance is third and is where the role specialists earn their premium — in legal and financial work an answer without a traceable citation is not usable output, it is a draft nobody can sign. Action capability is fourth: whether the assistant drafts the email or sends it, and what governs the difference.
| Capability | What it does | Buyer translation |
|---|---|---|
| Context reach | Which systems the assistant can actually read | The single largest determinant of usefulness. Notion describes completing complex, multi-step tasks using context from Notion, connected apps and the web. |
| Permission fidelity | Whether answers respect live entitlements | Ask whether permissions are evaluated per query or synced. A synced index is a permission model with a lag. |
| Provenance and citation | Whether an answer can be traced to a source | Decisive in regulated work. Harvey describes purpose-built agents that execute complex legal work end to end, in a domain where the citation is the deliverable. |
| Action capability | Whether it drafts or executes | Writer describes executing a described task from start to finish — which is a different governance question from drafting. |
| Cross-suite operation | Whether it works outside its parent's estate | SAP positions Joule as spanning SAP and non-SAP systems, which is the claim every suite vendor now makes and the one to test hardest. |
| Metering transparency | Whether you can predict next quarter's bill | Consumption pricing without a usage dashboard is an unbounded commitment. Ask what the reporting looks like before the contract, not after. |
| Data handling | What is retained and what trains a model | Google states that your data is not used to train Gemini models or for ads targeting; get the equivalent in writing from every vendor on the list. |
Which vendors lead in Role-Specific AI Copilots for the Enterprise?
The camps below are organized by what the assistant can see, because that predicts where each one is genuinely strong. Vendors move between camps — the suite copilots are all extending reach outward, and the horizontal assistants are all adding depth — but the origin still shows in the product.
A caution about how this market presents itself. Every vendor now describes autonomous end-to-end work across all your systems, and taken literally the category contains a dozen products that all do everything. They do not. The claims are usually true in the sense that a capability exists under each heading, and misleading in the sense that depth is concentrated where the company started — the suite copilot is deepest in its suite, the legal specialist in legal work, the revenue platform in the pipeline. Ask each vendor which category of work they would send you to a competitor for. The ones that answer are describing a product; the ones that do not are describing a category.
| Vendor | Approach | Where it fits |
|---|---|---|
| Writer | Builder platforms | Teams building governed agents around their own content and standards |
| Salesforce | Business-application copilots | Estates where the customer record and the work both live in the CRM |
| Glean | Horizontal work assistants | Organizations whose knowledge is scattered across many unrelated systems |
| Microsoft | Productivity-suite copilots | Organizations already committed to Microsoft 365 across the workforce |
| Harvey | Regulated-domain specialists | Legal and professional services work where the citation is the output |
| Gong | Revenue-role platforms | Revenue teams whose most valuable corpus is their own customer conversations |
One representative of each approach is named here; the category runs to several dozen vendors and most large organizations already hold two or three without having compared them. The camps were written before the vendors were chosen, and no placement here is for sale. Any vendor in this category can speak for themselves in the Spotlight below.
How much should you budget for Role-Specific AI Copilots for the Enterprise?
Published pricing is thin here and the pattern is itself informative. Microsoft publishes its builder-platform rates in detail — tenant-wide packs of 25,000 Copilot Credits at $200.00 per pack per month, with a pay-as-you-go alternative that differs only in how you pay. Notion publishes $10 per 1,000 credits. Beyond that the numbers thin out fast: the Glean pricing page redirects to the Glean home page, Salesforce states that subscriptions other than Starter are generally paid annually in advance without printing the rate, and the enterprise copilot pages of the largest vendors are quote-driven. Where a rate is not published, this guide says so rather than estimating one.
The unit matters more than the rate, and there are only two. A seat is predictable, budgets cleanly and charges for a population rather than a behavior — which means the finance risk is paying for people who never log in. A credit or outcome is fair, tracks actual value and is close to unforecastable in year one, because the count depends on adoption you have not yet seen. Neither is wrong. What is wrong is signing one and budgeting as though it were the other, which is the most common commercial mistake in this category.
Three costs sit outside every table. Permission remediation comes first and is frequently the largest: the assistant reads what the employee may read, so any over-permissioned repository becomes visible immediately, and cleaning that up is a project rather than a configuration step. Change management is second, and it is not optional — the difference between a copilot deployment at fifteen percent weekly use and one at sixty is almost entirely training and worked examples, not product. Third is the overlap you already own: before adding a license, check whether the same employee is already covered by a suite copilot bundled into an agreement signed by someone else.
| Basis | You are charged for | Grows with | Where it goes wrong |
|---|---|---|---|
| Per seat | Every licensed employee, used or not | Headcount | Bimodal adoption. You pay the same for the daily user and the one who tried it in March. |
| Per seat, tiered | Different seat classes with different capability | How many people need the full tier | Tier creep, as the useful features migrate upward at renewal. |
| Consumption credits | Actions and responses, metered | Automation, not headcount | Forecasting. The first year has no baseline and the second is priced against it. |
| Prepaid credit packs | Capacity bought up front | Whatever the meter counts | Unused capacity at period end, and overage rates nobody read. |
| Per outcome | Completed work rather than access | Volume of the work itself | Defining the outcome. The definition is the negotiation, and it happens once. |
| Bundled in a suite license | Nothing incremental, until the tier moves | The suite's pricing | Renewal, when the bundled assistant turns out to sit one tier up. |
| Departmental specialist | A role-specific tool on a departmental budget | That department's headcount | Invisibility. Nobody compares it against the enterprise copilot because they report to different budgets. |
Pricing in this category is predominantly quoted rather than published, and several vendors publish nothing at all — Glean's pricing URL resolves to its home page. This section describes the metering bases rather than rates, and quotes a published figure only where the vendor prints one.
How long does implementation take for Role-Specific AI Copilots for the Enterprise?
The failure pattern in this category is not technical. Deployment is straightforward, the assistant works on day one, and usage collapses in month three because nobody was shown what to use it for. Treat this as an enablement program with a software component rather than the reverse.
Before any assistant reads your content, find out what is over-shared. The copilot will find it immediately and surface it to whoever asks, and discovering this after rollout turns a productivity launch into a security incident. This phase is unglamorous and it is the one that most often gets skipped.
Pick two or three roles with well-understood daily work and measure task completion against a baseline. Volunteer cohorts self-select for enthusiasm and produce adoption numbers that do not survive a general rollout. The point of the pilot is a defensible per-role case, not a positive one.
For each piloted role, compare the bundled suite copilot against the specialist on the same tasks with the same graders. This is the comparison almost nobody runs and the one that determines whether you are paying twice. Expect the answer to differ by role.
Publish worked examples from your own organization rather than the vendor's, refresh them as capability changes, and review consumption monthly if you are on a credit meter. Adoption is a maintained number, not an achieved one.
A general assistant drafting an email carries little regulatory weight. The same assistant inside a hiring workflow, a credit decision or a claims process sits within a regulated process, and the obligations attach to that process rather than to the copilot. Because these tools spread by role rather than by project, the classification can change without anyone re-opening the assessment — the copilot approved for the marketing team is the same license the recruiting team starts using in the following quarter. Tie the assessment to workflows rather than to the product.
Classified under the EU AI Act's risk tiers, as they apply to the workflow a general-purpose assistant is used inside rather than to the assistant itself
What should you ask vendors about Role-Specific AI Copilots for the Enterprise?
The first question is the one most organizations cannot answer today, and answering it usually changes the shortlist.
-
Do you already pay for a copilot covering this employee?Yes Compare against it on the same tasks before buying. The bundled option costs nothing incremental and often loses on merit rather than on price.No Establish where the work lives. Reach decides more than model quality, and reach is a function of connectors.
-
Is the role's output externally reviewed or regulated?Yes A specialist with real provenance and citation is worth the premium. A general assistant produces drafts nobody can sign.No The suite copilot or a horizontal assistant will very likely be sufficient, and one of them is already paid for.
-
Are you buying assistance or automation?Yes Automation — expect a consumption meter, and negotiate the reporting before the rate.No Assistance — per-seat is right, and the number that matters is weekly active use rather than licenses issued.