meta
Vatsal Shah

AI Workflow Automation

Multi-step AI agents that replace the repetitive internal workflows your team runs every day — built into your existing stack, evaluated like software, monitored in production.

Talk to Vatsal

A workflow agent that actually closes the loop — intake, classification, routing, enrichment, follow-up — running inside the tools your team already uses. We build agents that handle real volume, not orchestrations that fall over the first time a vendor sends a weird PDF. Most engagements ship a working pilot by week four and reach production by week six.

Fit for

  • Canadian businesses with a workflow that runs 100+ times a month and currently eats analyst, ops, or coordinator time
  • Teams that can describe the workflow step-by-step and have examples of the inputs (tickets, emails, invoices, forms, CRM records)
  • Operators who want a human-in-the-loop checkpoint on the first version, not a fully autonomous black box

Not a fit for

  • One-off automations that run a few times a quarter — a Zapier or n8n recipe is the right tool
  • Workflows where the inputs are not yet digitized or structured anywhere
  • Teams expecting an agent to decide policy without a defined approval path

What you walk away with

  • A working multi-step agent deployed into your production environment, triggered by your existing systems (email, webhook, CRM event, queue)
  • Integrations with HubSpot, Salesforce, Microsoft Dynamics, Slack, Teams, Gmail, Outlook, and the line-of-business tools the workflow touches
  • An eval harness covering 50+ real historical cases with expected outputs, so regressions are caught before they hit production
  • Human-in-the-loop checkpoints — approval queues, confidence thresholds, and clean handoffs when the agent declines a case
  • Monitoring for run volume, success rate, error patterns, average handle time, and cost per run
  • A runbook for your team: when to retrain, how to add new cases to the eval, and how to extend the agent to adjacent workflows

How the engagement runs

  1. 1

    Discovery · Week 1

    We map the workflow end to end with the people who run it today. Pull historical examples, document the decision points, define what success looks like in numbers.

  2. 2

    Architecture · Week 2

    Agent design, model selection, integration plan, approval and escalation policy. You see the system shape — and the cost-per-run estimate — before we build.

  3. 3

    Pilot build · Weeks 3–4

    Working agent in a staging environment, running against real historical cases. Weekly demo, weekly course-correct. Eval results shared every Friday.

  4. 4

    Production deployment · Weeks 5–6

    Phased rollout into your live environment with monitoring already wired. Start with one workflow or one team, then expand once the metrics hold.

  5. 5

    Optimization · Weeks 7–10

    Tune against real production data, expand the eval set, and hand off the runbook so your team owns the agent end to end.

By industry

For Canadian Financial Services

Workflow agents for credit unions, MGAs, and brokerages — application intake and pre-qualification, claim triage and routing, KYC document review, vendor onboarding, broker commission reconciliation. Integrates with your AMS, CRM, and core systems, with human review on every payout-affecting decision.

For Canadian Healthcare

Workflow agents for multi-site clinics, dental groups, and physiotherapy networks — referral intake and routing, prior-auth submission, lab result triage, recall and follow-up scheduling, insurance claim resubmission. Tuned for your provincial framework and connected to your EMR or practice management system.

For Canadian Customer Operations

Workflow agents for support and revenue operations across SaaS, e-commerce, and B2B services — ticket triage and routing, lead enrichment and scoring, refund and RMA processing, churn-risk flagging, contract-renewal prep. Plugs into HubSpot, Salesforce, Zendesk, Front, and the CRM and helpdesk your team already lives in.

Selected work

Recent workflow automation work.

“Vatsal is an excellent full stack developer and highly skilled project manager. He identified our business needs quickly and established a very strong framework. His incredible speed should be noted, this is a developer who doesn't waste time and hit every target date we threw at him.”
Josiah Liesemer
Josiah Liesemer
IT Specialist and Developer, Zucora Home
Read more in our engineering log →

Frequently asked

How do you keep a human in the loop without slowing the workflow down?
Every agent ships with confidence thresholds and an approval queue. High-confidence cases run end to end; low-confidence cases route to a reviewer with the agent's reasoning attached, so the human is approving in seconds rather than redoing the work. You set the threshold and you can change it any time.
How do you eval a workflow agent?
We pull 50–200 real historical cases from your systems with the known correct outputs. The agent runs against that set after every change. We track exact-match accuracy where the output is structured, and rubric scoring where it isn't. Eval results show up in CI — a regression blocks the deploy.
Which CRMs and tools do you integrate with?
HubSpot, Salesforce, Microsoft Dynamics, Monday.com, Zendesk, Front, Intercom, Slack, Teams, Gmail, Outlook, Google Drive, SharePoint, Notion, and most line-of-business systems with an API. For systems without an API, we build adapters or use platforms like Workato or n8n as the integration layer.
What happens when the agent hits an error in production?
Errors are instrumented. Anything outside the eval distribution triggers an alert, the case is parked in a review queue, and your team handles it manually while we add the case to the eval set. Production errors become training data for the next iteration rather than a fire drill.
Do we own the code?
Yes. Everything we build ships into your repository from commit one. No vendor lock-in, no recurring license tax on top of inference costs.
What does it cost?
Pricing is custom per engagement and depends on workflow volume, integration surface area, and the scope of the production rollout. We share pricing on the first discovery call once we understand what the workflow actually looks like.
Where does the agent run?
Default is your existing cloud — AWS, Azure, or GCP. For Canadian data residency, we deploy on AWS ca-central-1, Azure Canada Central, or Bedrock Canada. Inference can run through Anthropic, OpenAI, Google, or open-weights models depending on cost, latency, and residency requirements.
Who does the work?
Two to three engineers from our Toronto-based team, led by Vatsal. The people who scope the engagement are the people who write the code. The full team is named on our team page — you can see and talk to them before we start.