AI employees · checked October 29, 2026
AI employees that work inside your projects
An AI employee is an AI agent with a job: a name, a role, a set of tools it may use, the people it answers to, and work that is assigned to it the way work is assigned to a person. Gartner expects 40% of enterprise applications to include task-specific AI agents by the end of 2026, up from less than 5% in 2025 [2], and also expects over 40% of agentic AI projects to be canceled by the end of 2027 over cost, unclear value or inadequate risk controls [1]. The difference is usually not the model. It is whether the agent has a clear role, narrow permissions and a person who approves what matters. Kavanah lets you hire AI employees, such as an AI CFO, an AI COO or an AI chief of staff, scope each one, give it scheduled work and assigned tasks, seat several in a daily executive meeting, and hold every action to a policy and a ledger. This page explains what that looks like, what to require from any vendor, and how five products compare.
#Why it is different
Why most "AI employees" are a chat window with a job title
Every work tool now ships agents. Asana offers more than 30 prebuilt AI Teammates [7], ClickUp sells Super Agents you can @mention and assign tasks to [9], monday.com shows ready-made agents such as a contract reviewer and a deal scorer [11], and Lindy calls itself the AI employee that plugs into your tools [13]. Gartner warns that much of the market is relabelling: it estimates only about 130 of the thousands of agentic AI vendors are real [1]. A job title on a chatbot does not make it an employee. An employee shows up without being asked, owns work, reports to someone, stays inside what it is allowed to touch, and leaves a record.
The risk is the same one security people already name. OWASP lists Excessive Agency, an LLM application performing damaging actions in response to unexpected or manipulated output, among the top risks for LLM applications, and traces it to three causes: excessive functionality, excessive permissions and excessive autonomy [4]. Its mitigations are the ones you would apply to a new hire: limit permissions to the minimum necessary and have a human approve high-impact actions before they are taken [4]. NIST's principle of least privilege says the same about any process acting on behalf of users [5], and Anthropic's own guidance on agents recommends the simplest design that works, guardrails, and checkpoints where the agent pauses for human feedback [3].
So the useful question is not whether a product has agents, but whether its agents can be given a narrow job and held to it. A small finance desk that runs a C-suite of AI officers on Kavanah put the result this way: "Kavanah gave us a C-suite of AI officers that actually show up every morning. By the time I've had my coffee, the pipeline, the stale deals and the three decisions only I can make are already waiting for me." (Henry Augustine, Managing Partner, Baird Augustine). The three decisions are the point: the AI officers prepare them, and a person makes them.
#What to require
What to require before you hire an AI employee
Use this list on every product you shortlist, including ours. Each item is something a person in the same seat would be expected to have, and each maps to a risk the frameworks above name.
01
A role that changes what it can do, not only how it talks
A role should narrow the tools an agent can reach, not just its tone. OWASP's first cause of excessive agency is functionality the task does not need [4]. Ask whether an AI CFO can be kept away from email sends, and an AI COO away from the books.
02
Permissions no wider than the person behind it
Least privilege applies to processes acting for users [5]. Asana says its AI runs on a user's existing permissions [6], ClickUp that its agents only access data they are explicitly allowed to see [9], and monday.com that you define which data an agent can read, edit or create [11]. Ask whose permissions an unattended run uses when nobody is signed in.
03
Approval for the actions that matter
OWASP's mitigation for excessive autonomy is a human approving high-impact actions before they happen [4]. Lindy says nothing irreversible happens without your approval [13]; monday.com offers a simulation mode before an agent is switched on [11]. Ask which actions wait, who approves, and whether the agent can approve its own request.
04
A record of every action, and a way back
ClickUp describes full audit trails for every action [9] and monday.com a real-time log of what each agent did and when [11]; Lindy lists audit logs on its Enterprise plan [14]. Ask for the log on the plan you will buy, and whether a mistaken action can be undone rather than only read about.
05
Work it does without being asked
An employee that only answers when prompted is an assistant. monday.com agents can produce status updates and reports on a set schedule [11], and Lindy takes repeat work on a schedule [13]. Ask what can run on a schedule, on an assigned task, or on a trigger, and where the output lands.
06
Someone accountable for it
Decide which person each AI employee answers to and which decisions stay with people. An agent that recommends a decision is useful; one that records a decision on a human's behalf is a governance problem. Ask whether the product can tell those apart.
07
A price you can forecast
Agent pricing comes in requests, credits and seats. Asana includes 5 AI requests per user a month, up to 50 or 250 per account by plan [8]; ClickUp pools AI credits per user by plan [10]; monday.com includes 1,000 to 3,000 AI credits a month by plan [12]; Lindy's Team plan is $49.99 a month for 5,000 credits [14]. Gartner names escalating cost as a leading reason agent projects are canceled [1]. Ask what one scheduled run costs and whether you can cap unattended spend.
08
A choice of model
Different jobs suit different models. Asana selects the model for you [6], ClickUp routes to models by intent [9], monday.com connects to Claude, ChatGPT, Copilot and Gemini [11], and Lindy lets you choose a model per task on its Team plan [14]. Ask whether a finance role can run on a stronger model than a triage role.
#How Kavanah does it
What an AI employee is in Kavanah
Kavanah is an agentic project management platform: projects, tasks, a Deal Room and planning, with AI employees that do coordination work inside limits a person sets. Each item below describes the product as it ships today. Anything that runs on its own or spends AI credits is off until an admin turns it on.
A named employee with a job description
An admin creates each AI employee with a name, a role title such as Chief Financial Officer, instructions in its own voice, an avatar, and its own model from the model picker (Claude, GPT, Gemini and others). Each carries its own know-how, vision and hard nos on top of the workspace charter; it can add a hard no but never remove one the workspace set.
Learn moreTools scoped to the role
Each employee can be limited to the capability groups its job needs, from tasks and projects to email and calendar, CRM, finance, reports or knowledge. A restricted employee never sees a tool outside its scope, in chat, on a schedule or on an assigned task.
Only the people you choose can use it
An employee is open to the whole workspace by default, or restricted to named members plus admins. A member off its list cannot chat with it, @mention it, bind it to a schedule or learn from the org chart that it exists. Only admins can create, edit or re-scope one.
Assign it a task, and it works the task
AI employees appear in every assignee picker. Once an admin turns on "AI employees work their assigned tasks", a task assigned to one starts a run. The run borrows the role and project visibility of the person who assigned it, never more, narrowed further by the employee's own scopes and model, and is capped at 20 runs an hour per workspace. With the setting off, the assignment is a label.
@mention it in a comment, a discussion or chat
Type @ and the employee's name in a task comment or discussion post and it answers there, signed as the employee with "(AI)". A chat message that opens with the name starts a conversation with it. A person's name always wins: an employee that shares a member's name cannot be summoned.
Scheduled work, delivered by email
Bind a recurring schedule to an employee: a morning operating check, an end-of-day wrap, a weekly forecast. Each run uses that employee's scopes and model and can email its result to the people who need it. A schedule bound to an employee that has been switched off does not quietly run as someone else.
An org chart and a daily executive meeting
Employees report to one another in an org chart the database enforces: same workspace, no cycles. Seat two to six of them in a boardroom and, once an admin enables exec meetings, they meet on a schedule: each gives an update from read-only tools, they discuss for up to three rounds without tools, and the chair closes with action items that become tasks for people or employees. A financial-services org template sets up a C-suite, its reporting lines and its meeting series in one step.
They prepare decisions; people make them
An AI employee recommends but never makes, approves or records a decision reserved for a person. With the daily decision window turned on, an employee asks a named person to decide on the task's Approvals section, with at most three open asks per employee, one digest a day per reviewer and one escalation after 48 hours. No agent tool can approve or reject; only a person can.
Policies, safe mode, a ledger and undo
Action policies, set for the workspace or for one employee, decide per category (sending messages, moving money, deleting, membership, deploying, writing to connected tools, spending AI) whether an action runs, runs up to a daily limit, waits for approval or is blocked. Safe mode removes risky tools entirely. Every action lands in an action ledger, and many can be undone.
Learn moreColleagues, a Water Cooler and Slack
An employee can ask another for its view, the CFO on the numbers or the CRO on the pipeline, and relay the answer by name. Two to four can debate a topic on the Water Cooler, without tools, once an admin enables it. An employee can also answer people in your connected Slack who are not Kavanah members, with no tools and no workspace data, from a playbook you write; it is created switched off.
Learn moreSee its week
The roster's This week view shows one employee's last seven days on one screen: scheduled runs, skips and failures, assigned-task runs, outcomes, escalations and open nudges.
Spend you can see and cap
AI is metered in credits, about a cent each, at the price of the model that ran. Unattended runs are billed in credits like any other AI use, never run free, and an admin can set one daily token ceiling across every unattended run: schedules, assigned tasks, meetings and the Water Cooler. Pay as you go has no seat fee; seat plans start at $3 a seat a month with 100 credits per seat pooled.
Learn more#In practice
A day on an AI C-suite: how a small finance desk runs it
AI COO and the managing partner
1. Before the partner's first coffee: the morning operating check
- A weekday-morning schedule bound to the AI COO reads open tasks, overdue work and what the team is waiting on, using only the tools in the COO's scope.
- It writes one brief: what moved yesterday, what is stuck and who owns it, and the short list of decisions only a person can make.
- The brief is emailed to the partner. The COO names the decisions; it does not take them.
- Anything the COO wants decided goes on the task's Approvals section as one ask, not a new task, when the daily decision window is on.
AI CRO and AI CFO
2. Pipeline and forecast
- A schedule bound to the AI CRO checks the Deal Room for deals that have sat in a stage too long and sends a pipeline-velocity alert naming them.
- The AI CFO's schedule sends a revenue forecast email, with charts, built from the deals' stages and amounts.
- Each employee sees only what its scope allows: the CRO the CRM, the CFO the finance tools.
- Outbound mail to people outside the firm waits for approval wherever the send policy says so.
AI CEO, CFO, COO, CRO and CMO, with the partner reading the minutes
3. The daily executive meeting
- An admin enables exec meetings and arms the boardroom; creating a boardroom never arms it.
- Each officer gives an update from read-only tools, so a number in the meeting comes from the system of record.
- They discuss without tools, then the chair closes with action items.
- Action items become tasks for named people or employees. A commitment that is ambiguous about which person owns it stays unassigned rather than being handed to an AI.
AI COO, AI chief of staff and the team
4. Routing the work and ending the day
- A routing schedule assigns new tasks to owners against the team's limits, so nobody's queue silently overflows.
- Tasks assigned to an AI employee are worked by it, if an admin has turned that on, as the person who assigned them.
- An end-of-day wrap reports what closed, what slipped and how the day moved the workspace's North Star metric.
- The partner reviews the open approval asks in the next morning's digest. The ledger shows every action each officer took.
#Compared
AI employees in Kavanah, Asana, ClickUp, monday.com and Lindy
Each competitor cell describes only what that vendor's own public page said when we checked on October 29, 2026: Asana's AI, AI Teammates and pricing pages [6][7][8], ClickUp's Super Agents and pricing pages [9][10], monday.com's AI and pricing pages [11][12], and Lindy's home and pricing pages [13][14]. "Not described" means the page did not say, not that the product lacks it. Kavanah cells describe the product as it ships.
| Criterion | Kavanah | Asana AI Teammates | ClickUp Super Agents | monday.com AI agents | Lindy |
|---|---|---|---|---|---|
| What they are called | AI employees | AI Teammates, prebuilt agents that work alongside your team [7] | Super Agents [9] | AI agents; an "AI agent workforce" on paid plans [11][12] | "The AI employee" [13] |
| Roles | You define each one; a financial-services template sets up a C-suite | 30+ prebuilt Teammates [7] | An agent catalog anyone can extend: PM, sales, coding, design [9] | Ready-made agents, e.g. contract reviewer, deal scorer, risk analyzer [11] | Marketing, sales, support, operations, finance and more; 40+ skills [13] |
| Assign tasks / @mention | Both; assigned tasks are worked once an admin turns it on | Assign tasks [6]; @mention not described | @mention, assign tasks, message directly [9] | Not described | @mention in a channel it has been added to [13] |
| Runs on a schedule | Schedules bound to an employee, results by email | Not described | "Work around the clock"; schedules not described [9] | Status updates and reports on a set schedule [11] | Repeat work on a schedule [13] |
| What it can reach | Tool scopes per employee; an unattended run borrows the assigner's access, never more | Runs on the user's existing permissions [6] | Permissions-aware; inherited, explicit or manual access [9] | Data access and read, edit or create, set per agent [11] | Not described; 1,500+ integrations [13] |
| Who may use it | Whole workspace, or named members plus admins | Admins choose who can create, modify and collaborate with Teammates [7] | Not described | Not described | Not described |
| Human approval before acting | Per-category policies per employee: allow, daily limit, approve or block; agents cannot approve | Not described | Not described | Simulation mode before an agent is activated [11] | Nothing irreversible without approval; writes and sends can wait [13] |
| Record of actions | Action ledger on every plan; many actions can be undone | Usage dashboard by user and Teammate [8] | Full audit trails for every action [9] | Real-time log of every agent action [11] | Audit logs on Enterprise [14] |
| Agents working together | Org chart, daily exec meetings, colleague questions, Water Cooler | Not described | "Agents have managers" (a heading; not explained) [9] | Work moves between people and agents [11] | Not described |
| Model choice | Per employee, from Claude, GPT, Gemini and other models | Asana selects the model [6] | Routes to models by intent [9] | Connects to Claude, ChatGPT, Copilot, Gemini [11] | Choose a model per task on Team [14] |
| How AI is priced | Credits, about a cent each; 100 or 250 per seat pooled, or packs from $10 per 1,000 | 5 requests per user a month, up to 50 or 250 per account; $0.50 per prepaid request [8] | Pooled credits: 1K per user on Core ($9), 1.5K on Business ($17); $10 per 10K extra [10] | 1,000 to 3,000 AI credits a month, Basic to Pro; no AI on Free [12] | Team $49.99 a month for 5,000 credits; $10 per 1,000 extra [14] |
#Questions
Questions buyers ask
- What are AI employees?
- AI employees are AI agents set up like members of staff: each has a name, a role, a defined set of tools and data it may use, people it answers to, and work it does without being asked, such as scheduled reports and assigned tasks. The difference from a chat assistant is ownership. An AI employee owns recurring work and leaves a record of what it did; a person still owns the decisions.
- Can an AI CFO replace a human CFO?
- No, and it should not try. An AI CFO in Kavanah prepares the work a finance lead reviews: a forecast email from the pipeline, variance and runway questions answered from connected finance tools, flags on stale deals. It can be limited to finance tools only, and it recommends rather than decides. Approving spend, signing off numbers and anything reserved for a person stays with a person.
- What does an AI COO or AI chief of staff actually do?
- The operating rhythm. In Kavanah that usually means a morning check of overdue and stuck work, routing new tasks to owners, chasing what the team is waiting on, an end-of-day wrap, and a daily executive meeting with the other AI officers that ends in action items. The chief-of-staff job of connecting work streams and keeping a leader briefed is the part agents do well; the judgement calls stay with the leader.
- Do AI employees run on their own?
- Only where you switch it on. Scheduled runs happen on the schedule you set. Working assigned tasks, daily executive meetings, the daily decision window and the Water Cooler are each a workspace setting that is off until an admin turns it on. Each unattended run is billed in credits, and an admin can set a daily ceiling across all of them.
- How do you stop an AI employee doing something it should not?
- Four layers. Scope its tools to its role. Restrict who can use it. Set action policies, for the workspace or for that employee, so sending messages, moving money, deleting or writing to connected tools runs, runs up to a daily limit, waits for a person's approval, or is blocked. And use safe mode to remove risky tools entirely. Every action is recorded in a ledger, and many can be undone.
- Whose permissions does an AI employee use?
- When it works a task assigned to it, it borrows the role and project visibility of the person who assigned the task, never more, and its own tool scopes narrow that further. An AI employee has no login of its own, so it cannot reach a project the person behind the run cannot see.
- Can AI employees talk to each other?
- Yes, in three ways. One employee can ask another for its view and relay the answer by name. Two to six can meet as a boardroom on a schedule, give updates, discuss and close with action items. And two to four can debate a topic on the Water Cooler without tools. Their discussion never executes anything on its own; action items become tasks with owners.
- Does anyone actually run a business on AI employees?
- Yes. A small finance desk runs an AI CEO, CFO, COO, CRO and CMO on Kavanah every day: a morning operating check, pipeline-velocity alerts, a revenue forecast email, a daily brief against its North Star metric, task routing and an end-of-day wrap. Its managing partner, Henry Augustine of Baird Augustine, said: "Kavanah gave us a C-suite of AI officers that actually show up every morning. By the time I've had my coffee, the pipeline, the stale deals and the three decisions only I can make are already waiting for me."
- How much do AI employees cost in Kavanah?
- There is no per-employee fee. AI employees use credits, about a cent each, at the price of the model that ran. Pay as you go has no subscription and no seat fee, with credits in packs from $10 per 1,000. Seat plans start at $3 a seat a month, $2 billed yearly, with 100 credits per seat pooled across the workspace. A daily executive meeting or a frequent schedule is the biggest driver, so set a daily ceiling on unattended spend.
#Go deeper
One question at a time
- What Are AI Employees? A Plain Definition, and What They Are NotEvery work tool now sells AI agents, and several call them teammates or employees. Here is a plain definition of an AI employee, the four things that separate one from a chatbot with a job title, and the questions that expose the difference.
- AI CFO: What an AI Finance Officer Can and Can't DoMost finance functions already use AI, and many CFOs say they have deployed AI agents. Here is what an AI CFO is good for on a small team, where it has to stop, and how to scope one so it prepares decisions instead of making them.
- How a Small Finance Desk Runs on an AI C-SuiteOne small finance desk runs an AI CEO, CFO, COO, CRO and CMO every day. Here is the shape of that day, from the morning operating check to the end-of-day wrap, and the rules that keep five AI officers from making a human's decisions.
- AI COO: Running the Daily Operating Rhythm With AgentsThe COO and chief-of-staff job is mostly rhythm: the morning check, routing work, chasing what stalled, the meeting and the wrap. Here is how to hand that rhythm to an AI COO without handing it your judgement.
- How to Give an AI Employee Guardrails: Scopes, Approvals, Audit and a BudgetOWASP traces the main agent risk to excessive functionality, permissions and autonomy. Here is how to answer each one for an AI employee: a scope, an access boundary, an approval policy, a ledger and a daily spend ceiling.
Related reading
- What are AI employees, and what they are notA plain definition, the four things that make an agent an employee, and the label traps.
- AI CFO: what an AI finance officer can and cannot doWhere AI helps a finance lead, where it must stop, and how to scope one.
- How a small finance desk runs on an AI C-suiteThe daily rhythm of five AI officers, hour by hour.
- AI COO: running the daily operating rhythm with agentsMorning check, routing, the exec meeting and the end-of-day wrap.
- How to give an AI employee guardrailsScopes, approvals, audit and a spend ceiling, mapped to OWASP and NIST.
- Agentic project managementWhat agentic PM is, the autonomy ladder, and the controls that make it safe.
- Project management software for financial servicesArchiving, supervision and deal workflow for regulated firms.
Sources
- [1] Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 — Gartner, 2025
- [2] Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, Up from Less Than 5% in 2025 — Gartner, 2025
- [3] Building effective agents — Anthropic, 2024
- [4] LLM06:2025 Excessive Agency — OWASP Gen AI Security Project, 2025
- [5] least privilege (glossary) — NIST Computer Security Resource Center, 2026, updated
- [6] Asana AI & Agentic Work Management — Asana, 2026, updated
- [7] AI Agents for Work: Asana AI Teammates — Asana, 2026, updated
- [8] Asana pricing: Personal, Starter, Advanced, Enterprise — Asana, 2026, updated
- [9] ClickUp Super Agents: AI agents that work like human teammates — ClickUp, 2026, updated
- [10] ClickUp Pricing and Plans — ClickUp, 2026, updated
- [11] The AI Workspace for People & Agents — monday.com, 2026, updated
- [12] monday.com pricing and plans — monday.com, 2026, updated
- [13] Lindy: Hire an AI Employee That Gets Work Done — Lindy, 2026, updated
- [14] Pricing — Lindy, 2026, updated
Sources and the comparison were checked on October 29, 2026. Product details change; check each vendor's own pages before you decide.
Run it on your own work.
Free to start, priced per seat from $3 a month with AI usage metered in credits.