AI that does a job, not AI that does a demo
AI automation means putting a language model inside a business process where the output is checked, logged and measurable: document extraction, support triage, research, classification or drafting. The engineering that matters is not the prompt. It is retrieval, evaluation, fallbacks and knowing which steps must stay deterministic.
Who this is for
You probably need this if…
If none of these sound familiar, you may not need us yet, and we would rather say so now.
Someone spends a day a week reading documents and retyping fields out of them.
A pilot worked beautifully in a notebook and fell apart against real inputs.
Support tickets get triaged by whoever opens the inbox first.
Your team pastes the same context into a chat window twenty times a day.
Nobody can tell you whether the AI feature is accurate, only that it feels good.
Scope
What's included in ai automation
Every item below is a deliverable with an owner and a date, not a capability we happen to have.
- 01
Process selection
We pick the workflow where automation pays: high volume, clear inputs, a checkable output. Most AI projects fail because they started on the wrong process.
- 02
Evaluation set
A labelled set of real cases with a pass threshold agreed before the build, so accuracy is measured rather than asserted.
- 03
Retrieval layer
Where your knowledge lives, how it is chunked, embedded and ranked, and how the system behaves when the answer genuinely is not there.
- 04
Agent or pipeline
Tool calling, structured outputs and the deterministic steps that stay deterministic. Not everything should be a model call.
- 05
Guardrails and human review
Confidence thresholds, escalation paths and an audit log, so a wrong answer is caught and traceable instead of silently shipped.
- 06
Cost and latency controls
Model routing, caching and token budgets with a per run cost you can actually forecast.
Process
How a ai automation project runs
6 stages over 4 to 8 weeks. You get a written plan before stage one, and a named owner on our side throughout.
Process audit
Week 1We time the manual work, count the volume and calculate what automation is worth. If the number is small we say so.
Evaluation first
Week 1Real cases are labelled and a pass threshold agreed. This exists before any prompt, or there is no way to know it worked.
Prototype against real data
Weeks 2 to 3Built and measured on your actual inputs, including the messy ones that break demos.
Harden
Week 4Retries, fallbacks, rate limits, cost controls and the behaviour when the model is wrong or the API is down.
Shadow run
Week 5Runs alongside the humans for a week. Disagreements get reviewed and fed back before anything is trusted.
Cutover and monitoring
Week 6Live behind alerting, with accuracy and cost on a dashboard your team owns.
Commercials
Timeline and cost
AI Automation and Agent Development starts at $12,000 and typically takes 4 to 8 weeks. Includes the evaluation set and a one week shadow run.
That figure is a starting point, not an estimate we revise upward later. The first week of every engagement is a scoping week that ends with a fixed fee and a fixed date. If the scope turns out to be smaller than we assumed, the fee goes down.
We do not bill by the hour for project work. Hourly billing pays us to be slow, and it makes your budget a function of our efficiency rather than of your requirements.
FAQ
AI automation questions
Direct answers. If yours is not here, ask us and we will add it.
What AI processes are actually worth automating?
High volume work with clear inputs and a checkable output: extracting fields from documents, triaging and routing tickets, classifying records, summarising long inputs, drafting first versions of repetitive text. Low volume judgement calls and anything where a wrong answer is expensive and hard to detect are poor candidates.
How do you know the AI is accurate?
An evaluation set built from real cases, with a pass threshold agreed before anything is built. The system is scored against it on every change, and it runs in shadow mode alongside your team for a week before cutover. Vibes are not a measurement strategy.
Which models do you use?
Whichever fits the task, cost and latency profile, usually Claude or GPT class models with smaller models routed the cheap steps. The architecture keeps the provider swappable, because the price and capability landscape changes faster than any twelve month contract.
What does AI automation cost to run?
Build starts at $12,000. Running costs depend entirely on volume and model choice, typically a few cents to a few dollars per run, and we instrument that from day one so you have a per run figure rather than a surprise invoice. Caching and model routing usually cut the naive cost by half or more.
Will this replace my team?
In our experience it removes the part of the job people already dislike and moves them to review and exceptions. We design for human review on anything consequential, because a system with no escalation path is a system whose first bad day is expensive.
Related
Often scoped together
Clay and GTM Engineering
Clay tables, enrichment waterfalls and signal based outbound wired into the CRM that has to act on it.
Clay GTM engineering →Custom Integrations and API Development
Middleware, private apps and file based pipelines for the platforms that never made it to anyone’s marketplace.
custom API integrations →Full Stack Web Application Development
Customer portals, internal tools and SaaS products on a stack your team can hire for.
full stack web application development →
Next step
Start with a conversation about ai automation.
A short call, no deck. We will tell you what we would do, roughly what it costs, and whether you actually need us, sometimes the answer is that you do not.