Only four things run on their own.

NearSync AI acts without a person in exactly four places: answering customers, writing up meetings, agents on a schedule, and workflows you have switched on. Each one has firm limits, each one turns off in a click, and everything else waits for you.

People HQLeave over five days5 steps, activeDry run
PeopleAsked, approvedno chasing
AttendanceWrittennot re-entered
WhatsAppTold themno email
BoardPostedno report run
One database, so one workflow reaches all of it.
One workflowStarts in People, finishes in three other departments

A leave request over five days asks a skip-level manager, writes the answer to attendance, tells the requester on WhatsApp and posts the month to a board. One workflow, four departments.

The whole list

What each one does, and what stops it.

Pick one to see what it does and the limits that hold it.

NearSyncNearSync AIAgentsSearch, plan and askKAll hubs

What it does on its own

  • Answers an inbound message from what is in your knowledge base
  • Cites the article the answer came from
  • Books nothing it cannot confirm, and sends your booking link instead
  • Hands over to a person and opens a ticket when it should

What stops it

  • Off until you turn it onAnd it cannot be turned on until you have said where escalations go.
  • It defers when unsure"I do not have a confident answer to that, let me connect you with someone."
  • A person takes over and it stopsThe moment somebody claims the conversation, it stands down.
  • It cannot inventNo prices, discounts, dates or commitments that are not in your own articles.

This is the only one that talks to your customers. It is also the most heavily fenced, and it ships switched off.

AI replies to customers are off until you turn them on.

Talking to customers

It hands over to a person when unsure.

When it is not confident in an answer, it stops and brings a person into the conversation.

  • Answer only from your own knowledge base, and show which article it used
  • Say it does not know, instead of guessing at a price or a date
  • Open a ticket when it escalates, so nothing is dropped
  • Step back the moment a person picks up the conversation
  • Stay inside the WhatsApp reply window, so you never have to track it
Customer questionHanded over
ConfidenceBelow the threshold
AnsweringStops
Goes toA person
HistoryGoes with it
What the customer sees
Starts againNo
Repeats themselvesNo
Knowing the edgebeats answering everything
Agents and workflows

Pick what it is allowed to do.

An agent can only take actions you ticked, and only so many in one run.

  • Point an agent at anything you track, without a developer
  • Let it check several related areas in one pass
  • Cap how many actions a single run can take
  • Test the whole thing with a dry run that changes nothing
  • See a record of every run: what it looked at and what it did
Describe itBeta
"When an invoice is overdue, email the client and ping the deal owner."
BecomesA live workflow
InSeconds
You thenEdit it like any other
Or start from
Starter5
Intermediate5
Advanced5
Expert5
One click loads the builderso a template is a starting point, not a black box
The model

Unattended work runs on the model you chose.

An agent is not stuck with a model we picked for it. Every run uses whatever your organisation set in NearSync AI, including a model running on your own hardware. Change that setting and the next run uses the new one, with nothing to reinstall and nobody to tell.

Every run
Model usedThe one you set
Fixed by usNo
A change takes effectNext run
One setting covers your people and your agents alike
What it sees

An agent shows the model only the fields it is checking.

Before a run asks anything, NearSync cuts each record down to the handful of details the check is about: a due date, an amount, who it belongs to. Fifteen at a time from the main list, and fifteen from each related area it looks at. Everything else stays where it is.

One run
Records handed overFifteen at a time
Related areasThree at most
Fields sentOnly the ones checked
A whole customer record never goes out to answer one question
Whose records

An agent can only ever read your own company's records.

Everything an agent looks at is narrowed to your organisation before a single record is read, and the related areas it checks alongside are narrowed the same way. A run started against an agent belonging to somebody else is refused outright.

Every read
Narrowed toYour organisation
Related areasNarrowed the same way
Somebody else's agentRefused
Checked where the run happens, not in the browser

Worth knowing

Nobody signs off each action.

A live agent makes its changes without asking first. That is what makes it useful, and it is why the limits matter: you choose the actions it can take, you cap how many it can take at once, you can watch a full dry run first, and one switch turns it off.

The loop guard

It will not get stuck talking to another bot.

An auto-reply on the other end once pulled the AI into a conversation neither side could stop. There is now a hard limit on how often it will reply to the same number, and on repeating itself, so a bot on the far end cannot drag it into a loop.

Access

Who can do what.

Building an agent is a department decision. Letting the AI answer customers is a company one, and it cannot be switched on until someone has chosen where escalations go.

See

See what every agent and workflow did, run by run.

Do

Dry run an agent and see what it would have done.

Manage

Build agents and workflows, and switch them on for your department.

Admin

Let the AI answer customers, and decide where escalations go.

Admin is the highest role and is not limited by these permissions, so keep it to the smallest group that genuinely needs it.

NearSync AI guides

See NearSync AI on your own data.

Half an hour on your own material, routed to whichever model you want answering.