Scope & boundaries
This guide covers the AI that answers and resolves employee requests — and the consolidation that removed the independent specialists, across four pricing units that cannot be compared to one another.
It does not cover the system of record underneath — tickets, catalog, CMDB and the workflows themselves (Enterprise Service Management (ESM) Tools), the assistant that helps an employee do their own job faster (Role-Specific AI Copilots for the Enterprise), finding the answer across every system, with no service desk underneath (Enterprise Search & Agentic RAG Platforms), or customer-facing calls replacing an IVR (AI Voice Agents & IVR Replacement).
Executive Summary
The independent AI service desk is close to extinct as a category. Three of its founders now answer to somebody else's domain, and that changes what a three-year commitment is actually a commitment to.
AI service desk agents answer employee questions and resolve IT, HR and finance requests in Slack, Teams, email or a portal, so fewer requests reach a human. The market divides into the AI built into the ITSM you already run, AI-native platforms that replace the ITSM, knowledge and search layers that answer questions without a service desk, back-office assistants that also serve HR and finance, and assistants that write into whichever ITSM you run. The specialists that defined the category now sit inside larger platforms. AI is charged for in three ways: a monthly allowance of AI conversations included in the plan and metered above it, an AI add-on priced per human agent, or a price per AI worker or per resolved outcome. In production, products differ on what the assistant can reach, what it may do without a human, and what happens when it is wrong.
Three specialists that defined this category now sit inside larger platforms. Moveworks states that ServiceNow has acquired it. A request for aisera.com is answered by automationanywhere.com. A request for espressive.com is answered by resolve.io.
So the buying decision has changed shape. It is no longer specialist versus incumbent, because most of the specialists are now inside an incumbent. It is a choice between the AI already bundled into the ITSM you run, an AI-native platform that intends to replace that ITSM, and a knowledge layer that does a surprising amount of the deflection without being a service desk at all. Those three are priced on units that cannot be compared, which is the second thing this guide is for.
Why Is the AI Service Desk Shortlist Out of Date?
The employee service desk was the first place generative AI produced a number a CFO recognized. Ticket volume is measured, deflection is measured, and the cost per ticket is a line somebody already owns. That made it the easiest AI purchase to justify in 2024 and 2025, and the category filled with well-funded specialists selling one thing: an assistant in Slack or Teams that answers employee questions and closes tickets without a human.
That phase ended by acquisition rather than by anyone winning. Each specialist went to a company that already owned the workflow underneath it: the ITSM records, the automation runtime or the orchestration layer. The assistant was the part they were missing, and it turned out to be the cheaper half to buy.
For a buyer, the practical consequence is about term rather than quality. Nothing that was good about these products stopped being good. But a three-year commitment to a product inside a larger platform is a commitment to that platform's roadmap and its commercial model, and both of those change after an acquisition in ways that are not visible at signing. If you are shortlisting a company that was independent when your requirements document was written, the first thing to establish is whether it still is.
Which type of AI Service Desk & Employee Support Agents fits your organization?
Nobody builds this any more, because the hard part was never the conversation. It was the integrations — the identity provider, the MDM, the HRIS, the ticketing system — and the permission model that stops the assistant from telling one employee about another's salary review. Leena AI states that its AI Colleagues arrive pre-built, pre-trained and pre-integrated to the back-office systems already in use, and that sentence is the entire value proposition of the category compressed into one line.
The real sourcing question is which layer you are buying at, because the three answers are mutually exclusive in practice. Buying the AI inside your ITSM means the records, workflows and approvals already exist and the assistant reads them. Buying an AI-native platform means replacing the ITSM, and vendors in this camp say so plainly rather than hiding it. Buying a knowledge layer means accepting that it will answer questions well and not do anything — which is fine if your ticket mix is mostly questions, and useless if it is mostly access requests.
| Layer | What you are buying | What it costs you |
|---|---|---|
| AI inside your ITSM | An assistant that reads the records, catalog and workflows you already maintain | Ceiling. It is as good as the platform's AI roadmap, and you do not get to shop for a better one separately. |
| AI-native service platform | A replacement ITSM built around agents rather than around forms and queues | A migration. Atomicwork describes itself as an ITSM and ESM platform run by AI Coworkers and governed by the customer’s IT team. Buying it is a system-of-record change. |
| Knowledge and search layer | Answers drawn from your documents, wherever they live, without a service desk underneath | Actions. It resolves the question and cannot provision the access. |
| Assistant over your existing ITSM | A conversational front door that writes into the ITSM you already run | Two vendors and one integration surface between you and every incident. |
| Build it yourself | A bot over your own retrieval and your own connectors | The connectors and the permission model, forever, and neither is where your team's time creates value. |
How Do You Evaluate Agentic AI for the Service Desk?
Deflection rate is the metric everyone quotes and the least useful one to compare, because no two vendors compute it the same way and none of them will tell you the denominator. Score these products on the things that determine whether deflection happens at all: what the assistant can reach, what it is allowed to do without a human, and what happens when it is wrong.
Four vectors separate these products once the demos end. The first is action versus answer — whether the assistant can actually provision access and reset a credential, or only explain how. The second is the permission model, which decides whether the assistant can be trusted with HR content at all. The third is where the assistant lives: an assistant employees have to visit is an assistant employees forget, which is why every serious product in this category ships into Slack and Teams first. The fourth is what happens on a miss, because an assistant that hands off badly costs more than no assistant — the employee has now waited twice.
| Capability | What it does | Buyer translation |
|---|---|---|
| End-to-end resolution | Completes the request rather than routing it | The difference between deflection and delay. Siit describes agentic AI that connects systems, routes requests and automates end-to-end workflows. |
| Permission inheritance | Answers only from content the asking employee may see | Non-negotiable once HR content is in scope. Ask whether permissions are inherited at query time or synced on a schedule. |
| Channel presence | Lives in Slack, Teams, email and the portal | Adoption is a function of where the assistant is. A portal-only assistant competes with the habit of emailing a person. |
| ITSM interoperability | Reads and writes the system of record you keep | Decides whether this is an addition or a migration. Rezolve.ai works with ServiceNow, Freshservice and Jira. |
| Model choice | Which model runs, and whether you pick | Atomicwork lets customers bring their own agent harness and models from OpenAI, Anthropic and Gemini — which matters for data residency more than for quality. |
| Multi-department reach | Serves HR and finance from the same assistant | Where the economics improve. Leena AI describes an agentic AI platform with pre-built AI Colleagues for the back office. |
| Governance surface | What each agent may do, and the record of what it did | Ask this before the pilot, not after. Kore.ai states that every agent is defined, tested and validated before deployment. |
Vendors in AI Service Desk & Employee Support Agents
Representative vendors include Atlassian, Atomicwork, Glean, Leena AI, Moveworks and Rezolve.ai. The list is alphabetical, is not a ranking, and is not for sale. How we choose vendors
The camps below describe where a product came from, which in this market still predicts what it does well. Note that the first camp is not a place to buy from so much as a place to check your shortlist against — if a name on your list is in it, the diligence question changes from product to platform.
One caution about reading these pages. Every vendor here now describes end-to-end autonomous resolution, and the copy has converged to the point where the homepages are nearly interchangeable. The distinctions that survive contact with production are not in the copy at all: they are the integration list, the permission model, and whether the pricing unit moves with your ticket volume or your headcount. Ask each vendor which ticket types they would not attempt to resolve autonomously. A vendor with a real answer is describing a product; a vendor with no answer is describing a demo.
| Vendor | Approach | Where it fits |
|---|---|---|
| Moveworks | Acquired specialists | ServiceNow customers, since that is now the relationship being bought |
| Atomicwork | AI-native platforms | Organizations willing to replace the service desk rather than bolt onto it |
| Leena AI | Back-office specialists | Shared services covering HR, IT and finance from one assistant |
| Atlassian | ITSM incumbents | Teams already standardized on Jira Service Management |
| Rezolve.ai | ITSM-agnostic assistants | Buyers keeping their ITSM and adding a conversational front door |
| Glean | Knowledge-led | Estates where the ticket mix is dominated by questions with documented answers |
One representative of each approach is named here; the category runs to roughly two dozen vendors, and the acquisitions of the last eighteen months have moved several of them between camps. 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.
Is AI Included in the Base Price or Sold as an Add-On?
Several vendors here publish prices, each on a different unit, and they show the three ways AI is charged for. Freshworks publishes a price of $19 per agent per month, billed annually, described as solid foundations for rapid growth. Freshworks publishes a price of $99 per agent per month, billed annually, for unified IT service management with AI. Freshworks prices Freddy AI Copilot at $29 per agent per month. Atlassian includes 1,000 Virtual Agent assisted conversations per month at no additional billing cost. Atlassian charges from $0.30 per assisted conversation per month for Virtual Agent conversations at 1,001 and above. Atomicwork publishes a starting price of $25,000 per year, billed annually. Atomicwork charges $499 per worker per month for additional AI Coworkers. Atomicwork publishes a price starting from $3 per outcome.
Those are four different units — the human agent, the conversation, the AI worker, the resolved outcome — and the choice between them is a bet on which number in your organization grows fastest. Per-agent pricing is stable and rewards you for deflecting, because deflection lets you hold headcount flat while volume grows. Per-conversation pricing does the opposite: every deflected ticket is a conversation you pay for, so the bill rises exactly as the product succeeds. Per-outcome pricing sounds like the fairest of the three and is the hardest to forecast, because you are agreeing to a unit price before anyone knows the unit count.
Two costs sit outside all of these. The first is knowledge remediation, which is the real project: the assistant is only as good as the documents behind it, and somebody has to find, merge and retire the four conflicting answers before deflection reaches the number in the business case. Budget it as a content project rather than a deployment task, because that is what it is. The second is the integration surface. Every system the assistant must read or write is an authentication, a permission model and a failure mode, and the vendors that quote fastest are usually quoting against a shorter list than yours.
| Basis | You are charged for | Grows with | Where it goes wrong |
|---|---|---|---|
| Per human agent | Each service desk staff member with a license | Support headcount | Nothing much — and it rewards deflection, because the bill does not follow volume. |
| AI add-on per agent | AI features, priced per human agent again | Support headcount | Large desks, where you pay the AI premium on staff who never touch it. |
| Per assisted conversation | Each conversation the virtual agent handles | Employee adoption | Success. The better the assistant, the more conversations you are billed for. |
| Per AI worker | Each AI agent deployed, like a headcount line | How many roles you automate | Estates that want many narrow agents rather than a few broad ones. |
| Per outcome | Each resolved request | Ticket volume | Forecasting. The unit is clean and the count is the thing you were trying to change. |
| Platform floor plus usage | An annual minimum, then metered consumption | Whatever the meter counts | Small pilots, where the floor dominates and the unit economics never get tested. |
| Bundled in the ITSM | Nothing visible, until the tier upgrade | The platform's pricing, not yours | Renewal, when the AI you adopted turns out to sit one tier up. |
Where a vendor publishes a rate, this guide quotes the published figure and says who published it. Several vendors in this category publish nothing at all — Moveworks asks buyers to connect with a Moveworks expert to request a custom quote — and that is recorded rather than estimated.
How long does implementation take for AI Service Desk & Employee Support Agents?
The sequencing failure in this category is universal and avoidable: teams deploy the assistant, watch deflection come in below the business case, and conclude the product is weak. Almost always the product is fine and the knowledge behind it is not. Front-loading the content work is what separates a deployment that reaches its number from one that gets quietly de-scoped in month nine.
Pull the top fifty ticket drivers and find the documented answer for each. The count that cannot be answered from current documentation is your deflection ceiling, and it is knowable before you sign anything. This is also the number that should shape the business case, rather than the vendor's benchmark.
One department, unedited ticket text, in the channel employees already use. Measure containment and escalation quality separately — a high containment rate hiding bad answers is worse than a low one, and only the escalation transcripts show which you have.
Move from answering to doing, one workflow at a time, starting with password reset and access provisioning because they are high volume and low ambiguity. Each integration is its own permission review; treat them as sequential rather than parallel.
HR and finance are where the economics of this purchase actually work, and they are also where the permission model gets its real test. Re-audit knowledge quarterly — documentation decays continuously and deflection decays with it.
An assistant answering employee questions is not making consequential decisions about people, so the heavier obligations generally attach elsewhere. Two things do attach and are routinely missed. The assistant reads HR content, which makes permission inheritance a data protection control rather than a feature. And where the assistant touches hiring, performance or termination workflows — which back-office deployments reach faster than expected — the classification changes, because the surrounding process is the regulated thing rather than the chat interface.
Classified under the EU AI Act's risk tiers, as they apply to an employee-facing assistant and to the workflows it is later extended into
What should you ask vendors about AI Service Desk & Employee Support Agents?
The first question below is the one that has changed most in the last two years, and it is the one most requirements documents still do not ask.
-
Is every vendor on your shortlist still independently owned?Yes Proceed on product merits. Confirm it again before signing — this market moved three times in eighteen months.No Re-run diligence on the parent. You are buying their roadmap and their commercial model, not the one you evaluated.
-
Are you willing to replace your ITSM?Yes The AI-native platforms are in scope, and they are the only camp built around agents rather than around forms.No Choose between your ITSM's own AI and an ITSM-agnostic assistant. The AI-native camp is a migration wearing a chatbot.
-
Is your ticket mix mostly questions or mostly requests?Yes Questions — a knowledge and search layer will deflect more per dollar than a service desk product will.No Requests — you need the integrations and the action surface, and knowledge tools will stall at explaining.