AI Agents & Automation — AI agents and automation with approvals for ERP
AI Agents & Automation lets a company build, run and govern AI agents that do real work in its ERP: reading and acting on people, pay, finance and sales records through the same tool catalogue as Copilot. Every agent acts with the permissions of whoever runs it, and every change it proposes waits in Needs you for someone allowed to approve it, then runs as the person who started the run. Every run is traced and audited. Administrators connect outside tools, chain agents into flows, test agents with evaluation suites and cap daily usage. Limits today: model routing, rollouts, memory, skills and guardrail switches are previews that agents do not follow yet, the cost and return-on-investment figures are estimates, and flows start only when someone presses Run.
Last updated
No credit card · Ready in about a minute · Your data stays private
66 features in 9 areas · works with 11 other modules · 3 team workflows
Full feature list
Agent estate
See every AI agent your company runs, how it is doing and how the agents relate.
- Agents overview — One page shows how many agents you have, the runs in the last 24 hours and the success rate, what waits for your approval in Needs you, your squads, and a table of every agent you can pause or resume.For: Tenant admin, Manager
- Every agent has its own page — Open any agent from the overview, the map or the deploy banner and see its status, squad, owner, autonomy and what it may do without asking, with tabs for its instructions and tools, its runs (each opens its trace) and its versions. Arrows walk to the next agent.For: Tenant admin, Manager
- Search and sort agents — Search the agent table and sort by runs, success or name.For: Tenant admin, Manager
- Agents grouped into squads — Agents belong to squads, so a large estate stays understandable. Every workspace starts with five squads, and you can add one from the overview or from the squad picker while you build.For: Tenant admin
- Synapse map — A live diagram shows agents as connected nodes. Drag nodes to arrange them; positions are saved for everyone, and clicking a node shows its model, connected tools, and shortcuts to its trace, its settings and Pause.For: Tenant admin, Manager
- Live signal feed — Watch run started, step and finished events as they happen on the map.For: Tenant admin, Manager
- External agent registry — Track agents that run on other frameworks, such as LangGraph or CrewAI, so they are listed beside your own. They are tracked for visibility only and cannot be run here.For: Tenant admin
- Demo estate included — New workspaces start with a sample estate of five squads and fifteen agents to explore.For: Tenant admin
Building agents
Create an agent in minutes by describing its job.
- Describe the job and build — Type what the agent should do, or start from a template tile, and the builder drafts its name, squad and instructions, which you can then adjust.For: Tenant admin
- Write instructions, name and squad — Set the agent's identity and the instructions it follows.For: Tenant admin
- Choose the AI provider and model per agent — Pin an agent to a provider and model, or let it use the company's default AI provider.For: Tenant admin
- Three levels of autonomy — Choose Autonomous, Supervised or Gated. A supervised agent's changes wait for approval. A gated agent asks before it starts and again before each change. An autonomous agent's changes also wait, unless its owner marked that change and started the run.For: Tenant admin
- Test run before deploying — Run the agent on a real task and read the answer before others use it. Testing an agent that is already live runs its saved version and changes nothing about it.For: Tenant admin
- Test a new agent as a private draft — Test run saves a new agent once as a private draft that only you can see and run. Anything a test proposes to change waits for your approval. Deploy then makes that same agent active and opens it in the agents table.For: Tenant admin
- Edit and delete agents — Change an agent later, or delete it; its run history is kept. Every change is recorded.For: Tenant admin
- Pause and resume an agent — Pause stops an agent at once and cancels the changes it has waiting for approval; the screen shows Paused only after the change is saved. Resume brings it back.For: Tenant admin
- Create a squad while you build — The squad picker in the builder has New squad, so a new workspace can create its first agent without leaving the page.For: Tenant admin
- Let chosen changes run without asking — On an autonomous agent, its owner ticks the write tools that may run without asking on the agent's Autonomy tab; one confirmation names who stops reviewing the change, and unticking is immediate with Undo. It applies only when the owner starts the run. Tools that delete never qualify, high-risk ones say so, and outside tools need approval rights.For: Tenant admin
- Template marketplace — Browse starter templates by category, open one in the builder with Use and adapt it.For: Tenant admin
- Sensible sizes for names, instructions and tasks — An agent's name, role, instructions and tools, and the task you hand it, have clear limits (for example a task of up to 20,000 characters), the boxes stop at the limit, and a name can never be saved blank.For: Tenant admin
Tools and connections
Give agents safe access to your business data and outside tools.
- Use your ERP modules as tools — Agents can read and act in HR, payroll, attendance, accounting, CRM, inventory, sales, documents, projects, recruitment, social, gamification, learning, activities and notifications through the same tool catalogue Copilot uses.For: Tenant admin
- Agents act with the caller's own permissions — A tool call runs as the person who started the run, so an agent can never do more than that person could, and a refused call is explained rather than crashing the run.For: Manager, Tenant admin
- Connect outside tool servers — Add any server that speaks the Model Context Protocol (MCP), preview the tools it offers and attach them to agents. Access keys are stored encrypted.For: Tenant admin
- Probe and re-check a connection — Test a connection again at any time to refresh its list of tools.For: Tenant admin
- Connector catalogue — A catalogue of common business tools can be connected through its own MCP address and shows how many tools it brought.For: Tenant admin
- Safeguards on outside addresses — Connections to internal or unsafe network addresses are refused.For: Tenant admin
- Let outside AI tools use your ERP — Your company's ERP tools are also published to outside MCP-capable assistants through an access token you control.For: Tenant admin
- Outside tool servers on your own network, when your administrator allows it — A tool server on your company's private network can be connected once the platform administrator allows private addresses for your company (on-premises installs have it on). Otherwise only servers on the internet over https are accepted, and the refusal says who can change that. The platform itself, the cloud metadata address and our internal services are never reachable, and the address is checked again on every call.For: Tenant admin
Running agents and flows
Hand an agent a task and see what it did.
- Run console — Pick an agent, type a task and see the answer, with recent runs listed beside it.For: Manager, Tenant admin
- Agents that use tools — With tool use on, an agent can look things up and act in your modules for up to four reasoning turns and eight tool calls per run.For: Manager, Tenant admin
- Multi-step flows — Chain agents and tools into a flow you run on demand; each step's result is passed to the next. Checkpoints drawn in a flow do not stop a run yet, and the counters come from real runs.For: Tenant admin
- Flow editor — Build flows with an ordered step list, choose each step's type and agent, set a trigger and view the pipeline as a diagram. A flow starts when someone presses Run, or by itself from a schedule or an event in another app (Agents → Schedules).For: Tenant admin
- Run trace — Open a run to see every step, the tools called and their results, and whether each change was approved, declined or ran without asking, with filtering and export. You see the runs you started; seeing other people's runs is its own permission, and even then a step that read data you cannot reach yourself is hidden.For: Manager, Tenant admin
- Run history — List your runs (or everybody's, with the permission to see all runs) with agent, who started it, status, time and size.For: Manager, Tenant admin
- Runs always finish — A run never stays running or waiting for ever: an interrupted run ends, and a failure is a plain sentence in your language saying what happened and what to do next.For: Tenant admin
- Your own AI provider — Runs use the AI provider and key your company set up, in the language you configured.For: Tenant admin
- Paused agents refuse to run — A paused agent returns a clear message instead of running.For: Tenant admin
- Agents that start by themselves, on a schedule or when something happens — Choose weekdays, a time and a time zone, or pick an event such as a new help-desk ticket, a customer's reply on one, a message in the Social inbox, a new hire, an invoice becoming overdue or a payment reminder falling due, and an agent starts on its own. It runs as a named person with that person's permissions, any change it proposes waits in that person's Needs you list, each schedule time or event starts exactly one run, and pausing the agent stops its triggers.For: Tenant admin
- Publish an agent to your website chat, WhatsApp or the Help desk — Paste one line of code on your website, or pick the chat in Website builder settings, and visitors can chat with an agent you choose. With Social or Help desk the agent can also answer a WhatsApp number or a desk's new tickets — as an internal note a colleague sends on, or straight to the requester. The agent answers as a colleague you name and can only use the few tools made for visitors. If you allow it to make a change (a call-back lead, a help-desk ticket, a follow-up activity), every change waits for that colleague's approval and the visitor gets the answer once it is decided. A customer asking where their order is must give the order number and the e-mail or phone on it, or the status link you sent them; they see where the order is and when it is promised, never prices, addresses or anyone else's order. Each visitor has an hourly limit, the website must be on your list, every message is a recorded run that the Run trace labels with the channel's name, and what visitors typed is deleted after 90 days.For: Tenant admin
- Runs in the background — A run or flow starts at once and keeps going if you close the page; the run console shows each step as it happens, and the result stays in the run trace. A run whose server stops is marked failed after 15 minutes.For: Tenant admin, Manager
- Connect a model on your own network (on-premise) — On an on-premise installation, a company with no AI model yet can connect one that runs on its own network from the Agents page: choose Ollama or LM Studio, type the machine's address, see what it holds, test it and save it. Only a private network address is accepted, never the server itself, and the hosted service never offers it.For: Tenant admin
Governance and safety
Keep humans in control of what AI does.
- Needs you: changes wait for approval — Every change an agent proposes waits in Needs you, on the overview and on Governance. Its card shows the agent, the change in words, every value that will run and the person it will run as; approving runs it as the person who started the run, with their permissions. A gated agent asks once to start and again for each change. Nobody approves their own gated start (they may decline it to withdraw it) unless the company has switched that rule off, and what a run found reaches the approver only as far as they could open it themselves.For: Approver
- Approve, decline or edit what an agent wants to change — Approve, decline with a reason, or, if you started the run, change a value first. High-risk changes and changes sent to an outside server ask for a second confirmation, and a change that was updated meanwhile cannot be approved by mistake.For: Tenant admin
- Approvers are told — The people who may approve get one in-app notification each time a run waits, and the owner hears once when a run made changes without asking.For: Tenant admin
- Guardrail switches — Create a guardrail that keeps agents out of chosen modules (for example Payroll): while it is on, a run that tries one of their tools stops with a warning step and the tool does not run, and only someone allowed to approve agent actions can turn it off. The other listed switches (spend ceiling and so on) are recorded but not followed yet.For: Tenant admin
- Daily run limit — A company-wide cap on runs per day (500 by default, 0 for no limit) stops runaway spend.For: Tenant admin
- Cost and monthly budgets — Set a price per million tokens for each model and every run that ends is priced from its tokens; Cost & ROI shows spend by agent and month in your currency. Local models cost nothing, and a model with no price is counted, never guessed. Give an agent a monthly budget: it is flagged at 80 % and paused at 100 % with the reason stated.For: Tenant admin, Manager
- Staged rollouts — Record a new agent version's stage, from shadow to canary to full traffic, or roll it back one step. A preview: the stages are not enforced yet.For: Tenant admin
- Versions and restore — Every deploy of an agent is a version. Rollouts shows each version with who made it and what changed (instructions, tools added and removed, provider, model), and Restore puts an earlier version back as a new one, re-checking which tools may run without asking. Each run records the version it ran.For: Tenant admin
- Model routing rules — Routing rules runs follow: pick a model for some agents (or all of them) and their runs use it, with the run saying which rule chose it; a model your AI provider cannot serve is not forced. The prose rules shown below them are still a preview.For: Tenant admin
- Kill switches — The platform can switch agent running, and separately tool use, off for a company or for everyone.For: Tenant admin
- Separate permissions — Viewing, editing, running, editing flows and approving are five separate permissions, so a manager can run agents without being able to change them.For: Tenant admin
- Audit trail — Every agent change and run is written to the audit log, including when someone removes an agent's memory. Admins read it under Governance: who did what, to which agent, with an approval's tool and outcome, filtered by action, person, agent and date and exported to CSV.For: Tenant admin
- Personal data hidden from the AI provider — When an agent reads a record and your AI provider is not on your own machines, Emirates ID, passport, IBAN, bank account and salary values reach the provider as hidden, and IBANs and Emirates IDs in any text are hidden too. People still see the real values. You choose the field names; only a person can change the list.For: Tenant admin
- Pilots: who may run an agent — Put an agent in a pilot and name the people who may run it: until you open it to everyone, only its owner and those people can start a run — by hand, on a schedule or by approving a gated start. Only the owner or someone allowed to approve agent actions can change it.For: Tenant admin, Manager
- Hand an agent to another owner — Pick a new owner by name in the builder: they approve the agent's changes and decide what it may do without asking, the old owner no longer may, and the old 'may run without asking' marks are cleared so every change asks until the new owner marks it. Audited.For: Tenant admin
Memory and skills
Make agents remember and follow your procedures.
- Agent memory — Store facts, preferences and past episodes, add or edit entries and export them. A preview: agents do not read these memories yet.For: Tenant admin
- Forget on request — Delete a memory entry for good, as a right-to-erasure step that is audited.For: Tenant admin
- Skill library — Write a skill as a numbered procedure, publish it, and import or export skills as files. A preview: agents do not use skills yet; write the steps into an agent's instructions instead.For: Tenant admin
Quality and improvement
Test agents and improve them from real runs.
- Evaluation suites — Write a suite for an agent — golden cases (a scenario and the correct answer) and a pass threshold — and run it: the agent answers each case, a second AI step judges pass or fail, and the score is saved with three real metrics: quality, safety (protection hits over 30 days) and the monthly budget left.For: Tenant admin
- Deploy gate — A draft agent deploys only after each of its evaluation suites passes its threshold against the agent's current settings. The Builder says which suite holds it and why; someone allowed to approve agent actions can deploy anyway with a reason, which is audited.For: Tenant admin
- Run every suite at once — Run all suites in order with progress shown, and stop when you wish.For: Tenant admin
- Self-improving suggestions — The optimiser reads recent runs, groups failures and proposes up to five fixes. Apply a fix to add it to the agent's instructions with a date. It reads only the runs you could open in Run trace, and leaves out the request and answer of someone else's run that read data you cannot see.For: Tenant admin
Assistant and platform
How agents fit with Copilot.
- Ask Copilot about agents — Copilot can list and describe your agents and their runs, explain how to build one, and run an agent for you. It cannot approve an agent's change: a person does that in Needs you.For: Tenant admin
- Arabic and right-to-left screens — All agent screens are translated into Arabic with right-to-left layout.For: Every user
Learn it in the app
A guided tour inside Agents teaches the work by having you do it.
- Hands-on guided tour — 4 short chapters (Know your agents; Build an agent; Give an agent a task; Stay in control) that start by themselves the first time you open Agents. Each step waits for you to really do it, such as opening a list or saving a first record, and any chapter can be paused, resumed or replayed from the last row of the Agents menu or from Tours in the account menu.For: Tenant admin, Manager, Approver, Every user
Works with other modules
Agents and Copilot share one tool catalogueAgents call the same ERP tools Copilot does, using your company's AI provider.Agents read and act on HR recordsAn agent searches employees, reads requests and, within the runner's permissions, approves or updates them.Agents read accounting figuresAn agent fetches ledger and invoice figures the caller is permitted to see.Agents work with leads and dealsAn agent lists leads and opportunities and drafts follow-ups using the caller's CRM permissions.Outside AI tools use access tokensAn access token created by an administrator lets an outside MCP assistant call the ERP tools it is allowed.AI agents can read and act on projectsAgents you build in AI Agents can use project actions such as sprint burndown and team utilisation, always under the permissions of the person or schedule that runs them.The Copilot and your AI agents use one catalogue and one company briefAI agents call the same actions the Copilot does, read the same company knowledge, and use the strong model when you have set one, while the Copilot can list, run and review agents.
Done by more than one person
An agent's change waits for an approverA gated agent asks to start, then again for each changeAn admin builds an agent and a manager uses it
What each role gets
- Manager — Runs agents, and saved flows, from the run console by typing a task, and reads the answer with recent runs listed beside it. Opens the trace of any run they started to see every step, the tools it called and what came back. Every run acts with the manager's own permissions, so an agent can touch only what they could touch themselves, and every change an agent proposes during the manager's run waits in Needs you for someone allowed to approve it, then runs as the manager. The manager can decline their own waiting requests. Building, changing and pausing agents stays with the tenant admin.
- Tenant admin — Owns the agent estate: builds an agent by describing its job or from a template, picks its squad (or creates one), provider, model and level of autonomy (Autonomous, Supervised or Gated), tests it as a private draft and deploys it. Gives agents the same tool catalogue the Copilot uses, across HR, payroll, accounting, CRM and other modules, plus outside tool servers that speak the Model Context Protocol (MCP), and chains agents and tools into flows. Approves or declines the changes agents propose in Needs you (never their own gated start while the company keeps that rule on), pauses an agent at once when needed, may let an autonomous agent's chosen changes run without asking when they start the run themselves, sets the daily run limit, runs the supplied evaluation suites and applies the fixes the optimiser proposes. Reads every run in the company, with any step that read data they could not reach themselves hidden. Model routing, rollouts, memory, skills and guardrail switches are previews under Advanced. Every change and every run is written to the audit trail.
Frequently asked questions
Do AI agents act without human approval?
Not unless you let them. Every change an agent proposes waits for someone allowed to approve it, unless the agent is autonomous and its owner marked that change and started the run. Tools that delete always ask, and every run is audited.
Can an agent see data a user cannot?
No. Tool calls run with the permissions of the person who started the run.
Which AI providers can agents use?
The provider and key your company sets up, and an agent can be pinned to another provider.
Can agents connect to outside systems?
Yes, through MCP servers you add. Their access keys are stored encrypted.
Do agents show real cost?
Yes. You set a price per million tokens for each model; every run that ends is priced from its tokens, and Cost & ROI shows spend by agent and month in your currency. Local models cost nothing; a model with no price is counted as a run without a price, never guessed. It is your price list, not your provider's invoice.
Can I stop runaway usage?
Yes: a monthly budget per agent (flagged at 80 %, paused at 100 % with the reason stated), a daily run limit, a pause per agent, and platform kill switches.
Can I create my own evaluation suites?
Yes. Create a suite for any agent, add up to 12 golden cases, set the pass threshold and run it. While a draft agent's suites have not passed, Deploy waits.
Are the messages an agent run shows in my language?
Yes. When a run cannot start — the task is empty, the agent is paused, the automation has no steps, or a request was already answered — the message says what to do next, in English or Arabic to match your workspace.
Will I know if an agent fails at night or the agents are about to stop?
Yes. When a scheduled, event or website-chat run fails, the agent's owner and the person it ran as get a notification that says why in plain words and opens the run's trace; by default only the first failure of each agent each day, with the rest counted on Home. Admins are told when today's runs reach 80% of the daily run limit and again when the limit stops the agents, an owner is told when an agent has spent 80% of its monthly budget or is paused at 100%, and the owners of agents that use an outside tool server are told when it stops answering. The same items wait on Home under Waiting for me.
Part of one platform
AI Agents & Automation is one of 59 apps (36 business apps and 23 ISO standards) that share the same sign-in, the same people, the same customers and the same AI copilot. See all apps or how Tarleaks ERP compares with Odoo, ERPNext and Zoho.