Workflow automation and AI
Automate with purpose and keep people in control. Rules for predictable work, AI where it adds value, and a defined answer for what each workflow can do.
Delivered from Vancouver for teams in Canada and the United States · records can be stored in Canada
Updated
What you get
- Human and automated steps in one process, not two parallel systems
- What each workflow can access, what it can do, and when a person takes over
- Grounding, confidence thresholds and escalation set per workflow
- A clear answer on what should stay a person’s job
Where this sits in the work
- Discovery
- Design
- Develop
- Deploy
- Continuous improvement
Predictable work wants rules. Work that needs judgment about unstructured input may want AI. Most processes want both, and the design question is where the line falls — not which technology to buy.
Rules first, where rules work. A routing decision with six known cases is a routing rule, and dressing it up as an agent adds an operating cost and an evaluation burden for nothing, and we will tell you when that is the case.
AI where it earns the complexity. Reading unstructured input, summarizing a history, drafting from context, triaging a queue nobody has capacity for. We recommend it when its value justifies the added complexity and running cost, and not otherwise.
One process, not two. Most organizations end up running the processes people follow, and AI experiments bolted alongside. The seam is where value leaks — work gets re-keyed, status lives in two places, and nobody can say what state a case is in. A step that needs a person is a person’s step; a step automation can own is automation’s; the transition between them is designed, logged and reversible.
That design work is most of the job. Deciding which steps genuinely need judgment, where a workflow should stop and ask, and what an escalation looks like at 2:00 in the morning is what determines whether the deployment survives its first bad week. We do it with the people who run the process today.
Two shapes of AI
Where AI is involved, there are two shapes and they solve different problems. Assistive sits with your people and makes them faster — draft this reply, summarize this account, find the contract term. Autonomous is given an outcome and works toward it on its own schedule, with oversight at points you choose. The comparison further down this page sets the two side by side.
Where it runs, and what stays yours
Where the work lives in CRM we configure it in Creatio AI Studio, where lifecycle, policy, approvals and monitoring have been native since the 10x release (why we build on Creatio). Where it runs outside Creatio we build on Google Vertex AI and integrate back. That call is made during Discovery, on your constraints, not on what we prefer to build.
Two decisions are worth separating. The runtime is where the work lives, and we are opinionated about it — that is where our delivery and support experience is. The model stays your choice: OpenAI, Anthropic, Google, or one you bring. Grounding, approval gates and the audit trail sit in the runtime rather than the model, which is what makes swapping the model a configuration change. It is not free — a material model change warrants revalidation on representative tasks rather than one successful demo. See our approach to controls for how that holds up under procurement review.
Assistive or autonomous?
They are peers, not rungs on a ladder. Most organizations should be running both, for different work — and deploying the wrong one is the most common way an AI program quietly fails.
Assistive agents
Productivity for the people you already have. Broad adoption, quick payback, and usually the right place to start.
Autonomous agents
Goal-driven digital employees for work nobody has capacity to do. Higher ceiling, and more design required before you switch one on.
| Aspect | Assistive | Autonomous |
|---|---|---|
| What it is | AssistiveA copilot working beside your team | AutonomousA digital employee owning an outcome |
| Who does the work | AssistiveThe person — faster | AutonomousThe agent — you review it |
| Where it lives | AssistiveInside Creatio, Outlook and Microsoft Teams | AutonomousOn its own queue and schedule |
| How it starts | AssistiveSomeone asks it something | AutonomousA goal, a trigger, or a clock |
| Human involvement | AssistiveContinuous — it never acts alone | AutonomousAt approval gates and confidence thresholds you set |
| Measured on | AssistiveTime saved per person | AutonomousOutcomes produced — revenue found, queue cleared |
| Typical first win | AssistiveDrafting, summarizing, finding the record | AutonomousIntake triage, referral spotting, document validation |
From the thing that arrived to the thing that got done
Four steps, and a person is in the loop wherever you decided one should be.
- 1Work arrivesAn email, a document, a call, a form, a record pushed from an ERP.
- 2An agent reads itGrounded against your own records — not the model’s training data.
- 3The process runsInside the CRM, under the approvals and audit trail you already have.
- 4A person gets the outcomeOr the approval, at the point you decided a person should see it.
One task, grounded, with a person at the end
The same loop we build, drawn in BPMN 2.0. Pick a scenario and watch the token take its route.
- Message start eventEmail with an attached scope document lands in the monitored inbox. The process is triggered by the message, not by a schedule.
- Service taskSix line items, a 48-hour target response and an indicative budget read out of the PDF.
- Business rule taskEvery SKU resolved against the live price book through a structured tool call. Nothing is priced from the model’s memory.
- Exclusive gatewayConfidence clears the threshold for this workflow, so the process continues on the upper branch.
- Service taskOpportunity created and a proposal drafted against the resolved figures.
- Send taskThe account owner is told there is a prepared proposal waiting, with the figures and their sources.
- Message end eventThe runtime’s own process ends here, acknowledging the requester. Nothing has reached the customer yet — the agent has finished, and a person has not started.
- User taskThe account owner’s process starts on that message. A person reads the proposal and releases it. Anything a customer sees passes through here.
- End eventReleased by a named person, with the whole run in the audit log.
A worked example of one agent loop in BPMN 2.0, not a recording of a customer system. Sequence flows stay inside the runtime; message flows are the only thing that crosses a participant boundary.
What buyers ask us
Both, and the split matters. Creatio AI Studio, which arrived with the 10x release in July 2026, manages the agent lifecycle inside the platform: instructions in the Prompt Agent Designer, multi-step processes in the Workflow Agent Designer, plus policy, approvals, execution monitoring and an audit trail. If the work is CRM work it belongs there, and we configure it for you.
Where the work runs outside Creatio we build on Google Vertex AI and connect back over Model Context Protocol (MCP) or REST. Either way the agent is scoped to one task and grounded in your live records — a generalist chatbot pointed at a CRM schema misreads it and exposes data nobody meant to share.
One caveat: Creatio AI Studio ships with the 10x release. On an earlier version, adopting it means planning an upgrade, and we scope that in Discovery rather than leaving you to find it later.
Predictable work gets a rule: if the inputs and the decision are the same every time, a rule is cheaper, faster and auditable by reading it. AI earns a step when the input varies — free text, documents, signals across systems — and when a wrong answer can be checked after the fact.
Each workflow gets a written answer for what it can access, what it can do, and when a person takes over. That document is the design, not an afterthought.
Anything where the judgment is real and nobody has written it down, where a mistake is expensive and surfaces late, where the volume is too low to repay the design work, or where the human contact is the product. We say it in the recommendation, per workflow.
An implementer who never says “leave this one alone” is selling automation rather than choosing it. The long version: when a workflow should stay human.
Not answered here? Ask us directly — we reply by the end of the next business day.
Bring us the part that isn’t working
A thirty-minute call with a principal consultant — not a sales qualification script. If we’re not the right fit, we’ll tell you.