AIAgents
Agents that do the work.
An agent reads what comes in, decides what to do, uses your tools to do it, and checks its own result. We design, build and deploy them for one person's week or a whole company's operations, with a person always able to see, pause and overrule what it is doing.
01What an agent is
An assistant answers.
An agent acts.
A chatbot waits for a question and replies. An agent is given a job and a set of tools, and works through it: it reads the ticket, looks up the order, checks what the policy says, creates the shipment, writes the reply, and confirms that it all went through. When it is unsure, it asks a person instead of guessing.
-
Chatbot
Answers questions
Knows what it was trained on and nothing about your systems. You read the answer, then you do the work.
-
Copilot
Suggests and drafts
Lives inside one tool, summarises and proposes. Useful, but you still press every button yourself.
-
Agent
Is given a job, and does it
Uses your tools to get from the request to the result, records every step, and reports back. You set the limits and can stop it at any moment.
02Anatomy
What's inside one we deploy
A model on its own is not an agent. The parts around it are what make one dependable enough to hand real work to.
- 01
A model
The language model that reads, plans and writes. We choose the right one for the job, and build so it can be swapped later for a better or cheaper one without rebuilding the rest.
- 02
Tools
The actions it may take: search the CRM, create a shipment, send an email, run a query, open a pull request. Each one scoped to exactly what it needs, each one logged.
- 03
Knowledge
Your documents, policies and history, retrieved when they are relevant, so every answer is grounded in what your company actually knows and can cite where it came from.
- 04
Memory
What it learned from earlier runs: this customer prefers a call, this supplier is always a week late, this rule changed in March.
- 05
Guardrails
What it may do alone, what needs an approval, what it must never do. Spending budgets, rate limits, and a hard stop a person can pull.
- 06
Evaluation
A set of real cases it has to handle correctly before any change ships, so an improvement in one place never quietly breaks another.
- 07
Observability
Every step of every run recorded and readable by a person: what it saw, what it decided, what it did, how long it took and what it cost.
- 08
Hand-off
A clean path to a person, with the context attached, whenever it is unsure, the stakes are high, or the policy says so.
03Use cases
Jobs agents are doing well today
Nothing speculative. Each of these is built around one specific job, with the tools it needs and the limits it must respect, and a person who can see every run.
For a businessNine jobs across the company
-
01
Support desk, resolved
Reads every incoming ticket, finds the customer's order and history, answers the routine ones end to end and resolves what the policy allows: a replacement, a refund below a limit, a corrected address. Everything else reaches a person with a drafted reply and the full context attached.
Takes the routine two-thirds. Hands over the rest, better prepared.
-
02
Sales research and the first draft
For every new lead: reads the company's website, filings and news, scores fit against your ideal customer, writes a one-page briefing and a first email in the rep's voice, and proposes meeting times from the calendar. The rep reviews and sends.
Reps walk into every call already briefed.
-
03
Receivables and the month-end
Matches incoming payments to invoices, chases what is overdue with reminders that escalate politely, flags disputes to a person, and prepares the reconciliation pack for the accountant to sign off.
Fewer overdue invoices, and a close that is ready to review rather than build.
-
04
The on-call partner
When an alert fires, it reads the logs, correlates them with recent deploys, proposes the likely cause, opens the incident with a timeline, drafts the status update and, where the fix is small and obvious, opens a pull request for an engineer to review.
The engineer wakes up to a diagnosis, not a dashboard.
-
05
Contract review
Reads inbound NDAs, supplier terms and leases against your playbook, marks up what deviates, explains the risk in plain language, and drafts the counter-position. Your lawyer decides; the reading is already done.
Every contract reviewed the same way, the day it arrives.
-
06
The dispatcher
Watches orders, stock levels and carrier feeds. Reorders below a threshold, reroutes a delayed shipment, tells the customer before they ask, and raises anything unusual with the operations lead.
Problems get handled while they are still small.
-
07
The hiring screener
Reads every application against the actual role, ranks candidates with written reasons, drafts questions specific to each one, schedules first calls, and keeps a fair, auditable trail of every decision it took part in.
A shortlist with reasons by the end of the day, not the end of the week.
-
08
The knowledge concierge
Answers your team's questions in Slack or Teams from the handbook, past tickets, decisions and documentation, with the source cited every time, and files the questions nobody could answer as gaps to fix.
The same answer whoever asks, with the source attached.
-
09
The Monday report
Every Monday it pulls the week's numbers from analytics, advertising, the CRM and the finance system, writes the note in your format, explains what moved and why, and proposes three things to try next.
The report writes itself. The meeting is about what to do.
For one personSix jobs for your own week
-
01
A chief of staff for your inbox
Triages your email overnight, drafts replies in your voice for you to approve in the morning, unsubscribes from what you never read, protects time for real work, and reschedules with the people on the other side when the day changes.
You open a sorted inbox with the replies already written.
-
02
The morning brief
Before you wake: today's calendar with what each meeting is really about, the messages that need you, the weather and the commute, the news that matters for your work, and the one thing you meant to do. Two minutes to read.
Two minutes, and you know your day.
-
03
A research analyst
Ask a real question, "which heat pump suits a 1930s semi under £12k?", and it reads dozens of sources, checks the claims against each other, and comes back with a ranked, sourced brief and the questions to ask an installer.
An afternoon of reading, done properly, in minutes.
-
04
The admin steward
Categorises your spending, spots the subscriptions you forgot, flags an unusual charge, files receipts for tax, tracks the renewals (insurance, passport, the car), and drafts the tedious emails to the customer service desks that keep you on hold.
The admin that eats your Sundays, handled.
-
05
Travel that reroutes itself
Plans the trip inside your constraints, books it, watches the flights and trains, rebooks when something moves, and keeps the people expecting you informed without you having to type a word at the gate.
You hear about the delay from your agent before the airline.
-
06
A coach that adapts
Reads your training, sleep and diary, adjusts the week's plan when life gets in the way, books the appointment you keep postponing, and prepares the questions to ask the doctor.
A plan that changes with your week instead of breaking on it.
A job of your own that isn't here?
Describe it to us04Deployment
Where they run, and
how much they may do
Three ways an agent can live, and one rule for all of them: it starts with the least autonomy and earns more only as the log and the evaluation results say it should.
-
For one person
A private agent
Runs on your own machine or a small server only you can reach. It connects to your mail, calendar, files and the apps you already use, and you talk to it where you already are: a chat app, email, or the terminal. Your data stays in your accounts, and it stops the moment you say so.
-
For a business
In your cloud, on your systems
Runs in your cloud or ours, behind single sign-on, connected to the systems of record: the CRM, the ERP, the helpdesk, the code, the chat. Every tool is scoped to exactly what the agent needs, every action is logged, spending is capped, and a person approves anything above the line you set.
-
Inside your product
A feature for your customers
Agents as part of your own software: a support agent inside your app, an assistant that sets things up, a workflow that runs itself for each customer. Metered, monitored and evaluated like any other part of the product.
The four rungs of autonomy
Every agent we deploy starts on the first. It moves up only when the numbers say so, and it can be moved back down in a minute.
- 1
Suggest
It drafts, summarises and recommends. A person does everything that has an effect on the world.
Every agent starts here - 2
Act with approval
It prepares the action and waits. One click from a person lets it through, or sends it back.
- 3
Act within limits
It acts alone below a threshold you set (a refund under €200, a reorder under a week's stock) and asks above it.
- 4
Run on its own
It handles the whole job and records every step. A person reviews the log each week rather than each case.
The work gets done. You stay in charge.
Every agent we ship can be watched, paused and overruled by a person. That is the whole design.
05How we build one
From the job as it is
to an agent you trust
The same road we walk on every project, with the steps that matter for an agent: the limits written down first, real cases as the exam, and autonomy earned in the open.
- 01
Shadow the job
We sit with the people who do the work now and write down every decision they make, every system they touch, and every way it goes wrong.
You get: a map of the job, and where an agent helps - 02
Draw the limits
What the agent may do alone, what needs approval, what it must never do, and what it may cost. Agreed in writing before anything is built.
You get: a one-page operating policy - 03
Build against real cases
We gather a set of real, anonymised cases, and the agent has to handle them correctly before it meets a customer. That set keeps growing for as long as the agent runs.
You get: an agent that passed the exam - 04
Launch on the first rung
It goes live suggesting, then acting with approval, with the log, the budget and the hand-off in place from the first day.
You get: a calm launch with a person in the loop - 05
Read the log, earn autonomy
Every week we read what it did, fix what it got wrong, extend the test set, and move it up a rung when the numbers say so.
You get: an agent that is better, and more trusted, every month
06Straight answers
What we tell everyone
before they sign
- Will it make mistakes?
- Yes. Every agent does, which is why we design for it: limits, approvals, a log a person can read, and a test set it has to pass before every change. The question is never whether it can be wrong. It is whether a mistake is caught, contained and cheap.
- Do we need an agent, or an automation?
- If every case follows the same steps, you need an automation, and we will build that instead. An agent earns its place when the work needs reading, judgement and a choice between options.
- Where does our data go?
- It stays in your systems. The model sees what it needs for the case in front of it, under your contract with the model provider or on a model you host, and nothing is used to train anything.
- What does it cost to run?
- Pennies to a few pounds per run, depending on what it reads and how much it does. Every agent has a budget, and the cost per case sits next to the outcome in the log, so the sums are always visible.
- How long until it works?
- A first agent on the first rung usually ships in weeks. Earning full autonomy takes as long as your data, and your comfort, need it to take.
- Which model do you use?
- The one that fits the job, and it can change. Agents are built so the model can be swapped when a better or cheaper one arrives, without rebuilding the tools, the guardrails or the tests around it.
07Start a project
Which job first?
Tell us about the work you would hand to an agent. We'll tell you honestly whether it is the right tool, and what the first rung would look like.
hello@emmauslabs.com