Connect and workflow automation
Workflow automation executes a graph you drew: this trigger, that condition, this action, every time. Connect reads something written by a human and works out what it needs. Use automation wherever the correct response is knowable in advance — it will be faster, cheaper and exactly repeatable. Use Connect where the input is prose and the right answer depends on who sent it and what happened before.
The dividing line#
Ask one question about the work: *could I write down, today, the complete set of inputs and the correct output for each?* If yes, that is automation, and building it as judgement makes it slower, dearer and less predictable for no gain. If no — if the input is a paragraph somebody typed and the right reply depends on their history, their stage and what your Knowledge says — a rule engine can only approximate it with an ever-growing pile of conditions.
Most businesses need both, and the mistake is rarely choosing wrong at the extremes. It is in the middle: an automation that has grown to two hundred branches because the real rule was never expressible, or a model asked to decide something that was a lookup all along.
What workflow automation does better#
- Determinism. The same input produces the same output, forever, and you can prove it by reading the graph.
- Connector breadth. Automation platforms exist to move data between hundreds of applications; that catalogue is their product.
- Cost and latency. No model call. A condition and an HTTP request are microseconds and effectively free.
- Testability. A branch can be unit-tested. A judgement can only be sampled.
- Exact scheduling and data plumbing. Cron-accurate timing, field mapping, format conversion and idempotent retries against a named endpoint.
- Regulatory comfort. A fixed path is far easier to explain to an auditor than a decision that was reasoned.
Where Connect is the right shape#
Unstructured input, and continuity. A customer writes three sentences that contain a complaint, a re-order and a question about an invoice. There is no trigger for that. Connect resolves the sender to a Person, reads the thread, grounds the answer in Knowledge and memory, and drafts one reply that deals with all three — or refuses the part it cannot support and escalates it.
Continuity is the second half. An automation fires and forgets; the next message is a new event with no idea that the previous one existed. Connect's memory resolves at four tiers, so a direction like *always copy the accounts address for this customer* set once against a contact outranks the channel default without anyone editing a graph.
Connect's own deterministic parts#
It would be misleading to present Connect as judgement all the way down. The parts that must be predictable are not left to a model.
| Gate | What it does | Not negotiable by |
|---|---|---|
| Autonomy mode and scope | Decides whether an outbound may go at all, narrowest scope winning | Anything the model concludes |
| Suppression, unsubscribe and do-not-contact | Checked in one place before any outreach | The Assistant, which cannot clear a do-not-contact entry |
| The daily allowance | Bounds how much is processed; work is held, not dropped | Retry pressure |
| The send boundary | Reports sent only on a provider acknowledgement, with uncertain as a real third state | Optimism |
| Follow-up duplicate protection | The same *when* for the same person is a duplicate | A second reasoning pass |
So the honest description is a deterministic frame around a reasoning core — and where the frame cannot reach, Connect says so. On the realtime voice path, for example, the no-inventing-prices rule can only live in the model's instructions, so a breach is detected rather than prevented, and the detection uses the same expression the turn-based path checks each sentence with.
Putting them together#
Automate the plumbing: form to record, record to accounting system, file to storage, scheduled export.
Result Fast, cheap and testable work stays that way.
Let Connect own anything whose input is prose from a human.
Result Nobody maintains a branch for every way a sentence can be phrased.
Where automation needs to reach into Connect, use the public API or the MCP server rather than driving a screen.
Result The call goes through the same permission and workspace checks a person's does, and lands in the audit trail.
Questions#
Can I write if-this-then-that rules for Connect?
You can set standing instructions and the behaviour on Connect Rules, and autonomy per channel, endpoint or contact. What you cannot do is author an arbitrary branching graph — Connect is not a workflow builder, and a direction is honoured as a direction rather than compiled into a path.
Is a model in the loop slower?
Yes, and measurably so on voice: the floor on the realtime path is the model's first token plus its end-of-turn detection, 3.3 seconds median on the best measured call. On email nobody notices, because the comparison is against a human replying in hours.
What if my automation and Connect both act on the same message?
They will both act. Give each one an unambiguous entry point — a mailbox, a channel or a label — rather than letting two systems watch one stream. Duplicate outbound is the failure customers actually notice.