AI with a
human in
the loop.
Sodelia is a certified Relay.app partner based in Toronto. We build AI-assisted workflows where the model handles the grunt work — drafting, classifying, summarizing — and your team gets a clean approval checkpoint on the calls that actually matter.
Native
Built-in
What we build
with Relay.
A good AI workflow is not a chatbot. It's a pipeline where the model handles what it's good at, and a human approves what matters — with an audit trail for both.
AI-drafted workflows
Email replies, support responses, summaries, and reports drafted by AI — reviewed by a human before they go out.
Human-in-the-loop checkpoints
Explicit approval steps where your team sees the AI's output, can edit or reject, and the audit trail tracks every decision.
Connected to your stack
Relay tied into Gmail, Slack, HubSpot, Notion, and the rest — so AI-assisted output lands in the same systems your team already runs.
The AI workflows
we ship most.
These are the Relay builds that consistently free up the most time without handing decisions to the model that shouldn't be automated.
AI email triage + drafting
Inbound email classified, drafted, and queued for one-click human approval.
Support ticket summarization
Long threads summarized into the facts your team needs before they respond.
Meeting-to-CRM automation
Call transcripts turned into CRM notes, next steps, and follow-up drafts.
Contract review assists
AI flags clauses, drafts redlines, human approves — faster turnaround, same oversight.
Content repurposing pipelines
Long-form content summarized and reformatted for email, social, and docs with human sign-off.
Internal ops copilots
AI-assisted internal requests — IT, HR, ops — where the model does the grunt and the owner approves.
Sample workflow — Inbound → AI → Human → Send.
AI is a teammate,
not a replacement.
Most AI rollouts fail because they tried to automate the wrong steps. We start by finding where AI actually saves time — and where a human still has to sign off.
Find the right step
We walk through your workflows and flag the steps that are repetitive, high volume, and safe for AI to draft.
Design the checkpoint
We agree, per step, what a human has to approve — and how they see the AI output before it goes out.
Build and calibrate
We run the workflow in shadow mode first, so your team can see what the AI would have done before it's live.
Stay on to tune
AI workflows drift as the model, your prompts, and your team change. We stay on to keep them sharp.
Questions we
get all the time.
Straight answers on scope, timelines, pricing, and what a Relay engagement with Sodelia actually looks like.
01What does it mean that Sodelia is a Relay partner?
We're a Relay Certified Partner — vetted on the platform, shipping production AI workflows across multiple industries. That means we know where Relay is genuinely better than rolling your own LangChain stack, where human-in-the-loop checkpoints actually matter, and how to keep AI-assisted flows maintainable six months after launch.
02How long does a Relay engagement take?
A focused AI workflow — one or two flows with LLM steps, approval checkpoints, and integrations — typically takes 4 to 6 weeks from scoping to production. More ambitious programs with multiple flows, custom models, and data warehousing run 8 to 14 weeks. We always start with one high-value workflow before expanding.
03When should we use Relay vs. building on OpenAI, Claude, or LangChain directly?
Relay is the right pick when you need LLM-powered workflows that a business team can observe, approve, and iterate on without a full engineering rebuild. It's not a replacement for a custom RAG system or an agent framework with heavy custom tooling — but for 80% of 'automate this workflow with AI' asks, Relay ships in weeks instead of months.
04What does pricing look like?
Initial builds are fixed-fee against a written SOW. Typical projects land in the $15–60k CAD range depending on flow complexity, custom integrations, and the evaluation/QA work required. Retainers ($4–10k CAD/month) cover iteration, new flows, model-drift monitoring, and expansion. LLM API costs are passed through at cost.
05How do you handle AI getting things wrong — hallucinations, bad outputs?
Every Relay workflow we ship has explicit human-in-the-loop checkpoints for anything customer-facing or irreversible. We build evaluation sets before going to production (a minimum of 50 to 200 gold-standard examples per flow), and continuous monitoring post-launch. 'AI did it' is never an acceptable excuse to your customers, so we design for the failure case first.
06Which models do you use — GPT, Claude, Gemini, or open-source?
We match the model to the task. Claude Sonnet and GPT-4-class models for most reasoning and drafting work; smaller/faster models (Haiku, GPT-4o-mini) for classification and extraction; open-source via Relay's model connectors when cost or data-sovereignty demands it. Most production flows route between 2 to 3 models in the same workflow.
07Can Relay integrate with our CRM, support desk, and data warehouse?
Yes. Native integrations with HubSpot, Salesforce, Zendesk, Intercom, and Slack cover most workflows. For data teams we ship events into Segment, Snowflake, and BigQuery, and we write custom connectors where Relay's native ones fall short. Every integration is auth'd per-environment so staging never talks to production systems.
08What do you need from us to get started?
For the first call — nothing. Before the second call: a concrete description of one workflow you'd like AI-assisted (not 'use AI somehow'), rough volume estimates, and the approval/review path that workflow currently follows. Book a call and we'll scope from there.
Let's see
if it's
a fit.
Book a 30-minute call. We'll walk through where your team is losing time today and sketch what we'd build in the first 60 days. No sales deck.
Schedule a call →