AI Automation Engineer
AI workflows that run against production data — with a human gate on every action that matters.
I'm Arshad Ansari. I build AI automation that does real work in real operations: document-to-table pipelines, natural-language querying over your warehouse, agents that investigate and propose and then wait for a yes. Every consequential action routes through an approval gate first, and every step is logged and traceable.
Every AI shop can demo a workflow. Almost none of them are running one unattended against production data a year later. I run mine — on my own data, in the open, before any of it touches yours.
The demo is not the hard part
A working prototype of almost any AI workflow is a weekend now. That is why every pilot looks good in the room and why so few of them are still running six months later. What kills them is the boring half: what happens when the API times out, when the model returns something plausible and wrong, when the person who built it moves teams, when someone asks in an audit what the agent did on the 14th of March and why.
So the question worth asking a vendor is not whether they can build the workflow. It is what still runs unattended a year later, who owns it, and what the record looks like when it gets something wrong. That is an engineering answer, not a demo.
The hard part is control, not the model call
An agent that can act on your systems is only as safe as the smallest thing it is allowed to do without asking. So I build the gate first and the intelligence second. Every action with a consequence — spending money, writing to a production system, sending something to a customer, deleting anything — stops and becomes a single decision record you approve or reject.
Everything else runs unattended: reading, drafting, classifying, investigating, preparing the thing for you to say yes to. The point is not to keep a human in every loop. It is to keep the human in the loops that can cost you something.
Gated
Money leaving, writes to production, anything a customer sees, anything irreversible. It waits. The default for a new tool is gated — a tool becomes unattended by an explicit decision, never by omission.
Unattended
Reads, retrieval, classification, drafts, investigation. This is where the hours come back — an agent that has already done the work and just needs a yes is worth far more than one you have to supervise.
Logged
Every step: the inputs, the tool called, what came back, who approved it and when. Durable workflows, so a crash resumes where it stopped instead of half-completing an action twice.
This is not a description of how I would do it. It is how AEGIS — my own automation platform, open source under MIT — already works: 4 named agents running 28 approval-gated workflows over 42 permission-gated tools, against my real tasks, email, money and infrastructure. Each agent is a permission boundary with its own tools. You can read the enforcement code before you talk to me, which is the point of it being public.
Why “on your own data” is the whole difficulty
A workflow that reads a curated demo dataset is a toy. Real automation reads and writes the systems your business actually runs on — and inherits every problem those systems have. If two tables disagree about what a customer owes, an agent will act on whichever one it read, confidently, at three in the morning. If a pipeline lands late, the automation makes yesterday's decision today and nobody notices until it reaches someone outside the company.
That is why I sell one thing rather than two. The data platform is not a separate practice I also happen to offer; it is the substrate that decides whether the AI work is trustworthy. Most AI shops cannot fix the layer their own workflows are standing on, so they build carefully on top of whatever they were handed.
If your inputs are the shaky part, start there: the fractional data engineering side covers the pipelines and the warehouse underneath, and the free platform teardown tells you in an afternoon whether it is shaky.
How we work together
Start with the free checklist. It costs you an afternoon and it is the fastest way to find out whether you have an automation problem or a data problem.
1 · AI Workflow Teardown
Free · no signupThe questions I work through before building anything: what the workflow touches, what happens when it is wrong, what must never happen without a human, and what the record has to show afterwards. Run it against your own pilot with your team — you don't need me for that. Run the teardown →
2 · AI Automation Build
Scoped in weeks1 build slot at a time — currently open
I build the workflow into your operations — the gates, the retries, the permission boundaries, the audit trail — and run it against your data until it holds. If the data underneath needs work first, that is the $3,000 Data Platform Audit and I will say so before you spend anything on automation. More on builds →
3 · Ongoing, part-time
Automation is not a delivery, it is something that runs. I stay on for a set slice of time each month to watch it, widen what it is trusted to do as it earns that, and be the person who owns it working. Priced against a fraction of a full-time hire, never a bare day rate.
What you'd do instead
Every one of these is a reasonable choice for someone. Here is who each one is actually right for, and where it stops.
An AI agency or automation shop
A demo, then a handoverThe demos are genuinely good. Ask how many of their workflows have run unattended against production data for a year, and who owns the thing the week after handover.
Hire an ML engineer
$180k+/yr + months of hiringYou get someone who can train and evaluate models. Almost none of your problem is model training — it is plumbing, permissions, retries and audit trails around a model somebody else already trained.
n8n or Zapier wired to ChatGPT
A weekend, then upkeepThe right way to find out in a week whether the idea works at all. Fine until the first silent failure reaches a customer, because nobody owns the retries, the guardrails or the log of what the model actually did.
Buy a vertical AI SaaS
Per-seat, immediatelyFastest path if your process fits their boxes. The moment it does not — your data model, your approval chain, your compliance rule — you are negotiating with a roadmap instead of changing a workflow.
Me, fractionally
Free checklist to startOne senior engineer who builds the workflows and the data platform they run on, hands them over documented, and stays around to run them. The proof that I work this way is public code, not a case study I wrote about myself.
This fits if
- →You have AI pilots that demo well and have never survived contact with production.
- →Someone has already wired an LLM to something that matters, and everyone is quietly nervous about it.
- →You have a backlog of "we should automate this" that never gets resourced, because it is nobody's job.
- →The work you want automated touches real systems — invoices, customer email, your warehouse — so "mostly right" is not a passing grade.
- →You want one person who builds the workflows and stays around to run them, not an agency that demos and hands off.
It probably doesn't if
- ×You want a model trained or fine-tuned on your corpus. That is an ML engineer's job, not mine — I will say so and point you at one.
- ×You want an agent with free rein over production and no gates. I do not build that; it is the thing that goes wrong.
- ×You are testing whether an idea has legs at all. Wire it up in n8n first — that costs you a week and tells you more than I can.
- ×You need the cheapest possible contractor by the hour. I price against outcomes and the cost of a hire, not a day rate.
Questions I get on the first call
What does "approval-gated" actually mean?
Every action with a consequence — spending money, writing to a production system, sending something to a customer, deleting anything — stops and becomes a decision record: what the agent wants to do, the inputs it used, and an approve or reject with your name and a timestamp on it. Reads, drafts, classifications and investigations run unattended. The gate is the design of the system, not a setting you switch on.
Can I see one running before I hire you?
Yes. AEGIS is my own automation platform and it is open source under MIT: 4 agents running 28 approval-gated workflows over 42 permission-gated tools, against my real data — tasks, email, money, infrastructure alerts. You can read exactly how the gates are enforced before you commit to anything.
Do I need a data platform before any of this works?
You need the parts the workflow touches to be reliable — the tables it reads, the systems it writes to. Sometimes that already exists and we start on the automation. Often the honest first step is a week on the data underneath it, which is what the Data Platform Audit is for. Either way I will tell you which one you are on the first call.
Have a pilot that needs to survive production?
Tell me what you are trying to automate and what it touches. The first step is a free 30-minute call — no pitch, just whether I can help.
Book a free scoping callNot ready to talk? Run the AI workflow teardown first →