Service
AI agents and automation
An AI agent does a job in several steps rather than one answer: it reads data, calls tools, checks its own result and asks a human when it gets stuck. Zero-Shot builds these agents, traces every step and ships them with an evaluation set that proves the agent does what it claims.
What we build
- 01Agents that call your own APIs, database and internal tools
- 02Approval gates — anything touching money, contracts or customers stops for a human
- 03A trace layer that records the input, output and cost of every step
- 04Recovery flows so the agent notices its own failure and retries
- 05A regression evaluation set that blocks a release when the new version is worse
How it runs
01
Map the job first
For a week we write down who does the work today, on which screen, in how many minutes. The process becomes explicit before anything is automated.
02
Evaluation set before code
We collect 50-200 real cases and label the correct outcome by hand. The definition of success is written down before the agent is.
03
Narrowest possible v1
One flow, one tool set, limited permissions. It goes live with real users at a small volume.
04
Widen the mandate
As long as the evaluation set stays green, the agent gets more tools and more decisions. When it goes red, the mandate is pulled back.
What we measure
- Task completion rate — how often the agent finishes without a handover
- Handover rate and the reasons behind it
- Cost per step and cost per completed task
- End-to-end latency, compared against a human doing the same task
- Number of reverted actions
When this is the wrong fit
If the job is a single step, you do not need an agent — a plain call is cheaper and more reliable. Processes with no written rules, run differently by every person, should be settled on the human side first; automating an undefined process only makes it fail faster.
By sector
The same work runs differently in every sector. The detail is here.
Common questions
How is an AI agent different from a chatbot?
A chatbot produces answers; an agent does work. It calls tools, writes data, triggers other systems and makes decisions between steps. That is why an agent needs permission limits, traces and a rollback path, and a chatbot does not.
How do you stop the agent from doing the wrong thing?
Three layers: the list of callable tools is closed, any step involving money or a customer requires human approval, and every release has to pass the evaluation set before deployment. On top of that, actions are designed to be reversible.
Which model do you use?
The task picks the model. Inside one agent different steps can go to different models — a cheap model for routine steps, a stronger one for decisions. We keep the provider swappable rather than locking you to one vendor.
Other services
What we've written on this
- Why AI agent projects failAgents that impress in a demo stall in production. Four observations on tool design, error handling, handing off to a human, and the cost ceiling.
- Where to start with enterprise AIHow to choose a first AI project: which tasks qualify, in what order to proceed, and what to expect in the first eighty days.
- Writing prompts for productionWhy a prompt that works in a demo breaks in production. Treating prompts like code: structure, versioning, boundaries and measuring every change.
Last updated: 2026-09-05
Let's talk about this one
Tell us what you want to build and within two days we'll tell you who takes it and how we'd measure it.