AI agents that do the HR work, inside limits you set
Ascend's AI agents draft onboarding plans, training and reviews from your own documents and records, check what people understood, and take real actions when asked, always with the asking person's own permissions and never further than your autonomy settings allow.
Several agents, each with one job.
Ascend is not one chatbot with a long prompt. Each agent owns a narrow piece of work, and the ones that write something about a person hand it to a person before it counts.
- The AI assistant is the one most people meet: it takes requests in plain language, answers from the company brain with sources, and carries out actions through a fixed list of 42 tools, 18 of which change something.
- The context agent builds the company brain: it reads the documents you upload, splits and indexes them, and keeps each passage visible only to the people you shared the document with.
- The drafting agents write first versions: training programs with a critique pass on every block, onboarding templates per job role with a reason for every gate, and manager reviews drawn from a person's verified skills and training record. A person approves each before it publishes.
- The insights agent checks your data on a schedule and raises findings nobody asked for: people below the skill level their role requires, overdue training that has been escalated, and repeated failed certification attempts.
- The examiners grade rather than write. One marks Understanding Checks criterion by criterion and re-explains what a failed answer missed. In certification conversations, one agent plays the customer, a separate monitor checks the simulation stayed in character, and a scorer that never spoke in the conversation grades it against the rubric.
Ask for the work, not for instructions on doing it.
The assistant only offers the tools the asking person's role allows, so an employee, a manager and an HR admin each see a different set. A few of them:
Employees
- “What's our policy on parental leave?”Answers from the policies this person is allowed to see, with the passages it used. If it cannot find a grounded answer, the question goes to an HR admin's queue.
- “How is my onboarding going?”Reads their own plan: what is done, what is due, and which gates are still open.
- “Add Team Lead to my career interests.”Records the interest on their own profile. It can be removed the same way.
Managers
- “Which sign-offs are waiting on me?”Lists the onboarding gates waiting on this manager. The manager decides each one in the app; no tool exists that signs for them.
- “Who is overdue on their training?”Shows overdue assignments for the people this manager can see, and nobody else.
- “Dismiss the skill-gap insight for Mateo, we've handled it.”A dismissed insight is never raised again, so this cannot be undone and waits for confirmation unless the area is on autopilot.
HR admins
- “Create a Senior Support Specialist job role.”A job role cannot be deleted once it exists, so unless the area is on autopilot the assistant shows what it will create and waits for a yes.
- “Start Tunde on the Support Specialist onboarding plan.”Tunde can read the plan the moment it exists, so this counts as irreversible and waits for confirmation unless the area is on autopilot.
- “Publish the fraud awareness program.”Publishes only if a person has approved every block. Otherwise the API refuses and the assistant says why.
Four limits that hold whatever the model decides.
- The agent borrows the person's authority. Every action goes through the same API route the app uses, carrying the asking person's own sign-in. There is no agent account and no agent role, so the assistant cannot do anything that person could not do by hand.
- An autonomy dial per area, set by admins at /admin/autonomy. Proposes only, the default for every area, means every action waits for the person to confirm it. Acts, asks first on risky ones, runs actions that can be undone and holds the rest. Acts, tells you after, runs without asking.
- A size limit the dial cannot lift. An action that names more than 10 people waits for confirmation at every level, autopilot included. Reversible is not the same as harmless once forty people have been told.
- Performance, succession and offboarding can never be set to run unattended. The database refuses the setting, not just the screen.
Every change an agent makes is written to an audit log the application cannot edit or delete from. Each row names the accountable person and, when an agent carried the action out, the agent alongside them. Admins can filter the log to agent actions and export it as CSV.
| Capability | Chatbot on your docs | Generic AI agent | Ascend's AI agents |
|---|---|---|---|
| Answers from your documents | Yes, often with links to sources | If it is connected to them | Yes, from the company brain, citing the passages used |
| Takes actions | No, it answers questions | Yes, across whatever tools it is given | Yes, 18 actions inside HR and learning work |
| Respects each user's permissions | Varies by product | Varies; some run on a shared service account | Acts with the asking person's own sign-in |
| Limits on how far it acts | Not needed; it changes nothing | Varies; often set per tool or per workflow | Autonomy dial per area, plus a hold above 10 people |
| Decisions about people | Not applicable | Depends on how it was set up | Cannot approve or sign off for anyone; performance, succession and offboarding never run unattended |
| Audit trail | Usually conversation history | Varies; often tool-call logs | Append-only log naming the accountable person and the agent |
High-stakes work needs a person who answers for it.
Banks, healthcare providers, retailers and support teams all get asked the same question sooner or later, by an auditor, a regulator or a customer: who decided this, and on what basis? An AI agent that approves on someone's behalf turns the answer into “the software”. Ascend's agents do the finding, drafting and preparing, and the decision stays with a named person.
That is why the assistant cannot sign off an onboarding gate, no agent drafts anyone's goals, and a certification conversation can only count as evidence if the simulation held. The time saved is the forty minutes before the decision, not the decision.
What buyers ask about AI agents.
- Can the agents act without permission?
- Never beyond the asking person's own permissions. Whether the assistant asks before acting depends on the dial for that area; every area starts at Proposes only, where nothing changes until a person confirms.
- What can they never do?
- Approve, sign off or attest for anyone. Publish a training program a person has not approved block by block. Draft someone's goals. Act on more than 10 named people without confirmation. Run unattended in performance, succession or offboarding.
- Who is accountable for what an agent does?
- The person who asked for it or confirmed it. The audit log names that person on every row and records the agent alongside them, so the two are never confused.
- Which model provider do the agents use?
- Language model calls go to Anthropic. Embeddings, search reranking and narration audio run on models Ascend hosts itself.
- Can we turn it off?
- You can leave every area at Proposes only, so the assistant can prepare actions but never carry one out without a person confirming. Questions and lookups still work, because they change nothing.
Where the agents connect.
Bring one job role. We'll run the loop on it.
Thirty minutes, your own documents, one job role. You'll see the company brain built, the drafts it produces, the sources behind them, and where the autonomy dial sits before anything publishes.