Services / AI Automation
AI automation delivered as working software
One concrete workflow — report assembly, lead triage, inbox-to-CRM — mapped, built, and tested on your real data, then handed over working. Engineered automations with retries, logging, and fallbacks, not a fragile no-code template and not a slide deck.
Sound familiar?
The problems we build for
Someone assembles the same weekly report by hand from systems that already hold every number in it.
Inbound leads sit unread for hours, and the first reply is written from scratch every time.
A no-code automation held things together until volume grew; now it silently drops records and nobody notices until a customer calls.
You've been pitched "AI transformation" twice this year, and neither pitch named a workflow, a deliverable, or a date.
The team wants to use AI in the back office, but nobody can say which tasks are safe to hand it.
Start small — fixed price
One Workflow, Working
The low-risk way to start: pick the one workflow that eats the most hours. We map it on a call, build it, test it on your real data, and hand it over working — with a fixed price in writing before anything starts.
- A workflow-mapping call, then a written scope and fixed price
- Tested on your real data, handed over with a runbook and video
- Two weeks of post-delivery fixes included
Not ready to commit? Ask for three free quick wins →
Fixed price
quoted in writing · delivered in 7 working days
Our guarantee — If the workflow doesn't run end to end on your real data at handover, you don't pay.
Pick your workflowWhat we deliver
The work, concretely
One workflow, working
A workflow-mapping call, then the automation built end to end — n8n or plain code plus an LLM API, whichever the job actually needs — tested on your real data and handed over working.
- A test run on your real data before handover, not a demo on ours
- A handover video and a written runbook your team can follow
- Two weeks of post-delivery fixes, included
Ops automation sprint
A process audit across the team's recurring work: the workflows worth automating mapped with honest time-cost estimates, the best of them built, and a "what NOT to automate" map for the rest.
- Three to five workflows mapped with time-cost estimates
- The best two or three built to the single-workflow standard
- A team walkthrough, plus what should stay manual and why
Care + build retainer
Automations rot: APIs change, models drift, edge cases pile up. Monitoring and fixes keep everything running, with one build-or-improve slot each month.
- Monitoring, fixes, and an LLM/API usage review
- One new or improved automation each month
- A monthly one-page report in plain English
Guardrails & human review
The model never gets the last word on anything that matters: rules decide, models draft, humans approve. Hostile or malformed input is caught before it ever reaches a model.
- Deterministic rules for decisions; the model writes prose, not verdicts
- Drafts marked for human review before anything is sent
- Prompt-injection and bad-input handling built in from day one
Not on this list?
If it's adjacent to this work, ask — we'll tell you whether it's inside our depth, including when the answer is no.
Ask us →In practice
What agencies and stores actually automate
Not technology for its own sake — the recurring, manual work that quietly eats staff hours. If one of these sounds like your week, it's automatable.
Reports that write themselves
The weekly numbers a team assembles by hand — sales, stock, campaign results — computed from the systems that hold them, with a plain-English narrative drafted for review and every figure traceable to its source.
Leads triaged and answered in minutes
Inbound leads scored by explicit rules, summarized, and given a draft reply a human approves — so the ten minutes per lead becomes thirty seconds of review.
Inbox-to-CRM without the retyping
Supplier and customer email parsed into structured records and pushed where they belong, with the original attached and uncertain cases flagged for a person instead of guessed.
Order-desk email answered with context
When a customer asks about an order, the reply draft already carries their order, shipping status, and history — a human reads, edits, and sends.
How we approach it
AI Automation the way we practice it
Engineered, not glued
Every automation ships with retries, logging, bounded timeouts, and a fallback path, so a provider outage means a delay, never lost work. The same discipline as any production integration.
The model never decides
Scores, categories, and anything money-adjacent come from traceable rules computed before a model runs. The LLM writes summaries and drafts — output a human reviews — so a prompt-injection attempt can't buy itself a result.
Working demo before working claim
Our own internal builds run end to end before we pitch the pattern to anyone, and what we show is labeled for exactly what it is: internal builds on sample data, never client results.
Tool-honest
If Zapier or Make covers your case at your volume, that's the recommendation — it's cheaper. Custom builds are for workflows that need reliability, judgment steps, or logic the templates can't hold.
Access & data
How we handle access to your systems
Automation and integration work means touching business-critical systems. These are the rules we work under — stated plainly so your legal and finance people can say yes.
- Least-privilege access
- We ask for the narrowest API scopes and roles that do the job — read-only wherever possible, staff roles instead of owner access, scoped tokens instead of master keys.
- Your accounts, your credentials
- Integrations run inside accounts you own. Secrets live in your platform settings and vaults — we don't keep parallel copies of your credentials.
- Access ends at handover
- Collaborator access is granted at the start and revoked at handover, with a written list of every system we touched so you can verify it yourself.
- NDA and DPA, no friction
- We'll sign your NDA before seeing anything sensitive, and a data-processing agreement where your data is in scope. GDPR-aware handling is the default, not an add-on.
- Continuity by design
- Runbooks, documentation, and full code ownership mean any competent developer can operate or take over what we build. The work never depends on us staying reachable.
Is the workflow wrapped around a store?
When the manual work is Shopify-shaped — orders, stock, catalog, storefront tickets — it belongs with our Shopify development service, and the two often meet in one build.
Platforms and technologies we build with
- Webhooks
- REST
- PostgreSQL
- Python
- Claude API
- Ollama
Related work
AI Automation in practice
FAQ
AI Automation: frequently asked
Can you show working proof?
Yes — two internal builds, documented end to end on this site: a weekly client report generator and a lead-intake triage pipeline. The triage build also runs as a live demo you can try right now at triage-demo.veldway.com — paste a lead, watch the rules score it, and see a hostile message get quarantined before any model runs. Everything is labeled for exactly what it is — internal builds on sample data, not client projects — with full source at github.com/veldway/internal-builds and github.com/veldway/live-demos.
What happens when the AI gets something wrong?
The system is designed so it can afford to: rules make the decisions, the model writes drafts, and a human approves anything that leaves the building. When a model is unavailable or unusable, the automation degrades to a deterministic path instead of failing — and that behavior is tested, not hoped for.
Which AI models and tools do you use?
Whatever the workflow actually needs: hosted LLM APIs such as Claude for language steps, local models where data shouldn't leave your infrastructure, and n8n or plain code for the plumbing. Our internal builds run the same prose step against a hosted API or a local model interchangeably — the architecture, not the vendor, is the deliverable.
Is our data used to train AI models?
Not by us, and we build against that risk: workflows run inside accounts you own, under your API keys, with providers' no-training settings applied where offered. Where data is too sensitive to leave your infrastructure at all, a local model does the language step instead.
What does AI automation cost?
A single workflow is a fixed price quoted in writing after the mapping call; sprints and retainers are quoted the same way, so you always see the number before committing. API usage runs under your own keys and is billed at cost — we don't mark up tokens.
What's wrong with just using Zapier or Make?
Often nothing — for low volume and simple mappings, we'll actually recommend them. They become the wrong tool when volume grows, when failures need retries and reconciliation, when a step needs judgment and review, or when per-task pricing starts exceeding the cost of owning the automation. That's the point where owning it beats renting it.
Have one workflow that eats the most hours?
Describe the problem in plain language — broken, slow, manual, or missing. We'll tell you whether and how we can help.
- 01We reply within one business day
- 02A short call to understand the problem
- 03A written scope and fixed quote — no obligation