AI & Business Process Automation.
We build AI and automation where it replaces a task a human shouldn't be doing — lead triage, invoice OCR, content drafting, CRM hygiene, routine replies. Nothing flashy, no dashboards that win awards and gather dust. Boring tools that save hours every week and keep working when nobody's watching.
OpenAI
ClaudeAnthropic
n8n
Supabase
Cloudflare
Node.js
PostgresOpenAIClaudeAnthropicn8nSupabaseCloudflareNode.jsPostgresSix places automation
tends to pay back first.
Not every workflow is worth automating, and nobody needs a chatbot for the sake of having one. These are the shapes that consistently return more hours than they cost — across the operators and founders we've worked with.
Lead triage and routing
Inbound leads scored, deduplicated against your CRM, routed to the right owner, and enriched with firmographic data before anyone looks at them. The sales team sees a cleaner pipeline and loses fewer leads in the first 24 hours.
Document and invoice OCR
PDFs, scanned invoices, contracts, purchase orders — parsed, validated against your rules, and dropped into the right system. Replaces a person manually retyping fields from email attachments into Zoho, Tally, or Xero.
Content drafting with guardrails
First-draft long-form, SEO briefs, social captions, internal comms — with house-style constraints, banned-phrase lists, and source-grounding. Your team still edits; they stop starting from a blank page.
CRM and data hygiene
Deduplication, field-standardisation, lapsed-contact re-engagement, stale-deal surfacing. A weekly cron that quietly keeps the CRM in a state your reps will actually trust.
Internal Q&A over your docs
A retrieval-augmented assistant over SOPs, handbooks, contracts, or support tickets. Private, cited, and restricted to the documents you feed it. Useful for onboarding, support deflection, and ops.
Routine reports and summaries
Monday-morning digests from your data — bookings, spend, support load, pipeline — written in plain English, delivered to Slack / email. Replaces the weekly 'can someone pull these numbers' Slack ping.
Automation as utility,
not demo bait.
Most AI pilots fail because they were built to impress an exec, not to replace a task. We work the other way around: start with the task, add the minimum amount of model, and only add cleverness if it earns its place.
- What usually happens
A chatbot that opens with 'Hi! I'm your AI assistant' and can't do anything specific.
What we do insteadA workflow scoped to one job — answer these questions, update this CRM field, route this lead — and bounded on what it can do outside that.
- What usually happens
A dashboard that needed six meetings to spec and gets opened twice a quarter.
What we do insteadA weekly digest pushed to Slack / email, written as a paragraph a person would actually read.
- What usually happens
An LLM hallucinates a customer number, a price, or a policy — quietly.
What we do insteadRetrieval-grounded, with citations, with a confidence floor. When it doesn't know, it says so, not makes it up.
- What usually happens
A growing monthly OpenAI bill and nobody can tell what it's actually doing.
What we do insteadLogged calls, per-workflow cost tracking, and a monthly report in plain English. If a workflow's cost outruns its value, we retire it.
- What usually happens
A prototype that breaks the first time someone uploads a slightly different CSV.
What we do insteadInput validation, failure modes, and a human-in-the-loop fallback before anything goes near production.
- What usually happens
A black-box SaaS that handles sensitive customer data on servers you don't control.
What we do insteadWe deploy on your cloud, your APIs, your data boundary. Vendor models are fine; vendor data stores are not.
Four principles.
No slogans.
Replace a task, not a person
We don't build automation to justify firing someone. We build it to return hours to people who are currently spending them retyping data or answering the same email for the 40th time.
Cheap, boring models first
GPT-4-class is overkill for most workflows. We start with the smallest, cheapest model that gets the job done. Upgrade only when we can point at a specific task where the cheaper one fails.
Retrieval beats fine-tuning, almost always
If the problem is 'the model doesn't know about our company', the answer is retrieval-grounding, not a fine-tune. Faster to ship, easier to update, cheaper to run.
A log, a rollback, an off-switch
Every workflow we ship has a kill switch, a log of every call, and a way to diff a decision if it goes wrong. This is table stakes, not polish.
How we work, start to finish.
- 01Workflow mapping
We sit with the team doing the task today. Watch them do it for an hour, document every step, find the actual bottleneck. Output: a list of candidate workflows ranked by hours-saved vs. build complexity.
- 02Smallest useful slice
We build the smallest version of the top workflow end-to-end — the one that could theoretically replace the task. Deploy it, have the team use it for a week, keep the human in the loop for review.
- 03Harden & hand over
Based on that week, we tighten the edges: validation, error handling, cost caps, logs, rollback. Once the team trusts it, we reduce the human-in-the-loop step and the workflow runs on its own.
- 04Iterate
Every month we look at cost, call volume, and errors. Retire workflows that aren't earning their keep, expand the ones that are, add the next one on the list. AI is a programme, not a project.
Before you
ask us.
Find the workflow that's worth replacing.
Tell us how your team spends its week. You'll get a short list of automation candidates ranked by what we'd build first — and an honest read on where AI wouldn't help.
or email hello@genvoid.com