Every co-worker owned, checked, stoppable and provable.
You don't buy it, install it or learn it. It's how we run every co-worker and custom build we deliver.
- Approved and sent the rail contractor approachYour owner, approvedApproved
- Prepared a reply to a new enquiryMira, waiting for a personWaiting for a person
- Drafted an approach to a rail contractorVera, waiting for a personWaiting for a person
- Found a new segment to rankAtlas, doneDone
- Re-ranked the weekly mapScout, doneDone
Six promises, today and next.
Today is what runs from launch. Next is designed, and we'll tell you on the call what's ready for your setup.
| Promise | Runs today | Comes next |
|---|---|---|
| Someone owns it | A named person in your firm owns each co-worker and reviews its work. | Every escalation reaches a named owner with a clock running, and moves to a backup if it waits too long. |
| It stays in bounds | Default guardrails stop it inventing facts or making commitments. | Rules your owner changes in plain English, checked before every action. |
| You can stop it | A stop switch for your whole setup. | Pause any single co-worker in one click, or stop all of them and cut their access within minutes. |
| You can see why | We record every run and can rebuild any run in minutes. | A plain-language account of what happened and why, for anyone in your firm. |
| It earns freedom | Co-workers prepare the work. Your people send every message. | It starts by shadowing, moves up one level at a time, and risky actions still ask a person. |
| It proves itself | A person reads a sample of runs before any client goes live, and we track what each setup costs. |
Four parts, and you can read all of them.
Every co-worker runs on the same tested machinery, set up with four parts your owner can read. So each new one starts from everything we've already learned.
| Part | What it is | Runs today | Comes next |
|---|---|---|---|
| Its rules | What it's for, what it may do, which risks need a person, who owns it, and how much freedom it has. Kept as settings a person can read and change, not buried in code. | Default guardrails, set by our team. | Your owner edits the rules in plain English. |
| Its actions | The specific things it may do in your systems, such as reading a CRM record or drafting an email. Each action carries its own level of risk. | A fixed set of actions for account mapping and reply handling. | Risk level checked on every single action. |
| What it can read | Only the knowledge and history its job needs, nothing wider. What one co-worker can read is set separately from the next. | Your account data and the sources we connect for its job. | Separate knowledge areas, set for each co-worker. |
| Its instructions | How it works and writes, in plain language anyone on your team can read. Your corrections and our checks tune them over time. | Tuned by our team from your corrections. | Your owner types a rule and confirms how the co-worker read it. |
Where your data goes, step by step.
Your data sits in a database of its own. Every action passes the same checks, and nothing reaches your clients until a person approves it.
- Back in your systemsBack to your systemsFinished work goes back into your systems through the same connections.
Our services run on Railway. Your hosting option decides where your database runs. Yarlis Hosted keeps it on managed Postgres next to our services. Ring-Fenced moves it to a separate database provider. On Your Servers runs everything inside your own cloud account.
We run the platform, and we don't read what's inside it.
How we look after your dataGuardrails, memory, knowledge and checks.
Open each one to see how it works, and what runs today.
Every action passes the same checks, freedom is earned one level at a time, and a stop switch halts your setup at once.
Same checks for every action. Every outcome recorded.
Nothing reaches your systems or your clients by a side door. Each action goes through the same steps in the same order, and anything the checks can't clear goes to a person.
Freedom changes how often a person is asked, never whether the other checks run.
A co-worker at the highest level still stops and asks when an action crosses a line you set.
Every action visible. Set to your firm's preference.
- What runs on its own
- Research, ranking and reminders.
- What waits for a person
- Every message to a prospect or client. Anything about price, scope or terms. Any reply that needs judgement.
- What you can switch on next
- Acknowledging RFPs and enquiries, and outreach to lower-priority accounts, inside rules you set.
Comes next
- Escalation and stopping
- Anything that needs judgement goes to a named owner, with the context attached. A stop switch halts your setup at once.
At launch the owner gets the review queue and the run records.
Reply to Brindle & Oak Consulting
Waits for a person: message to a prospect
Thanks for writing about support for your two delivery centres.
Before we suggest a scope, could you tell us which teams you would want covered first? Our Head of Delivery can walk you through how similar engagements have run.
Context attached: the enquiry, the CRM record, the last two emails.
Reply on pricing to Fenwick Hale Advisory
Price, scope or terms
Follow-up to Marlowe Street Studio
Message to a client
It starts by watching. Freedom is earned one level at a time.
Every co-worker starts in Shadow, showing what it would do while nothing runs. It moves up only when its record shows it's ready, and a person always confirms the move.
Risk still decides.
Oversight is worked out again for every action.
Even at "On its own, within rules", an action over a threshold you set waits for a person.
When something goes wrong, it stops rather than guesses.
Pause isn't stop
Runs today. A stop switch for your whole setup.
Comes next. Pause or stop one co-worker, one kind, or your whole setup.
Pause holds the work and keeps its access, and one click resumes it.
Stop cancels the work and anything waiting for approval, and cuts its access within minutes.
Show the safeguards
If it can't confirm, it doesn't act
When a permission check can't be completed, the action doesn't run. It treats silence as a no.
Comes next. Every action checked against permissions as it runs.
Stale data stops work
It checks data for freshness each time it reads it. If a connection to your CRM breaks, the work that depends on it pauses, and it tells a named person.
Comes next. Freshness checked on every read.
Retries never double up
If a step has to be retried, the action in your systems still happens once, so nobody gets the same email twice.
Comes next. Safe retries on every action.
Choose an action to send through the checks.
Allowed, within its rules, no person needed. It reads the record, and the run is recorded.
Proposes an action, passed
Using only the knowledge it's allowed to read.
Is it allowed?, passed
Permission and the stop switch, checked before every action.
Is it within its rules?, passed
Its job, and the actions it may take.
Does it need a person?, passed
Judged on the kind of action, money, confidence and how far it reaches. At launch, every message to a prospect or client needs a person.
Acts in your systems, passed
Through the connections you approved, and nowhere else.
Shadow. It shows what it would do. Nothing runs.
Runs today
Each approval, edit and rejection teaches it, and every change is recorded, so you can read, check and undo it.
It learns from your corrections, and you can see every change.
Learning happens in the instructions, knowledge and rules around the model, never inside the model itself. So you can read, check and undo every change.
Your decisions are the lesson
Each approval, edit and rejection teaches it. Your team does no separate labelling.
Runs today. Our team reads your edits and rejections and applies them.
Comes next. It applies corrections itself and tells the owner what changed.
Fix the fact, not a ticket
If it gets a fact wrong, your owner corrects it once, and the next run uses the corrected version.
Runs today. Tell your named contact and we correct it.
Comes next. Your owner edits the fact directly.
One record of everything.
Every approval, rejection, block, escalation and action lands in the same record. That record feeds its memory and our checks.
Facts are dated and sourced, old facts are kept and marked as replaced, and anything it isn't sure of waits for a person.
What it can read
Only the knowledge and history its job needs, nothing wider. What one co-worker can read is set separately from the next.
Runs today. Your account data and the sources we connect for its job.
Comes next. Separate knowledge areas, set for each co-worker.
Facts are dated, never overwritten
When something changes, it keeps the old fact and marks it as replaced, so you can see what it believed and when.
Comes next. Every fact carries its date and source.
Unconfirmed facts wait
A fact it isn't sure of waits for a person before it's used. It treats anything it reads as information, never as an instruction to follow.
Comes next. Low-confidence facts held back until a person confirms them.
A person reads 100 of its runs before your setup goes live. It goes live only with no critical failures and at least 95% judged acceptable.
Checked on every run. Read by a person.
Why one check isn't enough
If each step of a task is right 95% of the time, a five-step task comes out right only 77% of the time, and a ten-step task 60%. That's why work reaching your clients passes a person, and why we layer the checks, starting with the simplest.
Before you go live
A person reads 100 of its runs before your setup goes live. It goes live only with no critical failures and at least 95% of runs judged acceptable.
Runs today
Show the full checks
Runs today
Guardrails
Fixed rules that run on every action and block anything that breaks them. Is there a source for every fact? Has a person approved it? Is it under the cost limit?
Run records
Every step of every run is recorded and can be searched.
A person reads the work
A named person reads a sample of runs every week, and your team's edits and rejections show us what to improve.
Comes next
A second opinion
A different model scores samples after each run, to show people where to look. It never stops live work, and we check its scores against people's judgement.
Five measures
Did it run properly, was the work good, was it used, did your people accept it, and did it move the number you agreed?
The monthly sheet
One page for each co-worker: what it did, what you approved, what it moved, what it cost, and what changes next month. It ends in one verdict: continue, tune, pause or stop.
How we report results
A result counts only when it traces to a person acting on the work. We report it as associated with the work, not caused by it, so our numbers run low on purpose.
We measure the team, not the person
We measure what your people and the co-worker achieve together, never one against the other.
Three ways to deploy. The same co-workers in each.
What changes is who owns the environment. Your people work the same way in all three. Nothing to rip out. Your CRM stays the source of truth.
| Option | What it means | Who runs it | Availability |
|---|---|---|---|
| Yarlis Hosted | The fastest start, on our infrastructure, with your own database. | We host it. Your team operates the co-workers. | Standard |
| Ring-Fenced | Nothing shared: a separate environment, kept apart and run by us. | We run it, apart from other clients. | Scoped with you |
| On Your Servers | It runs inside your environment, including your own cloud account, under your control. | Set in your contract. | Scoped with you |
True in every option.
Model calls leave your environment.
Even On Your Servers, prompts go to an outside model provider, because we don't run local models.
Central controls stay with us.
They carry settings and health data, never your business data.
Where your data sits is agreed up front.
If it has to stay in one country, tell us and we'll be straight about what's possible.
Show the architecture
Our controls never hold your business data.
That's what lets us set up, update and watch over your co-workers without keeping your records. The three options differ only in where the lower half runs.
Runs today
Our controls
Run by Yarlis, shared across clients, and holding none of your business data.
- Users and roles
- Co-worker settings
- Permissions and the stop switch
- Health and cost monitoring
Only settings and instructions cross down. Health figures come back up, never your content.
Your environment
Kept apart for each client: Yarlis Hosted, Ring-Fenced or On Your Servers.
- Co-workers at work
- Knowledge and memory
- Run records
- Checks and scoring
- Connections to your systems
- A key of your own
Your systems
CRM, email and calendar, reached only through the connections you approve.
Model providers
The one way out to a model. Prompts go to an outside provider, in every option.
Your data stays yours.
| How it's kept | We keep each firm's data separate from every other client's. What we learn across clients is patterns, never data. |
|---|---|
| Who can access it | Only the people you name, and the Yarlis team members who look after your setup. |
| We look after it without reading it | Our team sees volume, health and how long items have waited. If one of us opens a client's content, that is logged as its own event, which you can see. |
| Every action logged | Every action is logged and auditable, in a record that is only ever added to. |
| Data rules | Designed to support your DPDP or GDPR obligations. |
| Certification | Compliance-grade by design. Built to SOC 2 controls, certification on a defined path. |
Looked after once it's live.
We monitor it, fix what breaks, update it as your systems or market change, and improve it from your team's corrections.
On by default; you can opt out.
What our team does.
- Reads your runs.
- A sample every week, and more before any change goes live.
- Checks the checker.
- We compare automated scores with people's judgement, every fortnight for work that reaches your clients.
- Answers a stop request at once.
- A named person picks it up.
One step at a time.
Demo on your market.
We look your firm up first, then show it on your own market.
The Account Map.
Your market mapped, ranked and explained, in a working session with your team.
Co-workers running.
Set up for your firm. Your team runs them; we maintain them.
Custom co-workers.
When you need them, in any function.
Clear lines, from day one.
| Co-workers | Research, rank, draft, and keep priorities in front of each owner. |
|---|---|
| Your people | Decide, lead the accounts that matter, negotiate, close, and set the guardrails. |
| Our team | Set everything up for your firm, maintain it, and keep improving it. A named person stays with you. |
What we need from you.
- One named owner for each co-worker
- A person, not a team inbox, with about 30 minutes a week set aside to review it with us.
- Access to what it works with
- The CRM and inbox it needs, agreed at the start. Nothing wider.
- Your starting point
- Before a co-worker acts, we record where you start. Without that, nobody can show what moved.
- Decisions at the agreed points
- Who approves what, and who signs off going live, agreed before the work begins.
Questions
- A tool leaves the outcome with you. We take on a defined outcome and run the co-worker that delivers it. It comes with a named owner, checks on every action, a full record and a stop switch, and we look after it once it's live.
- No. Corrections change the instructions, knowledge and rules for your setup, which your team can read. No model is retrained on your data.
- That's set out in your contract before work begins. In the full design each client's data has its own key, and deleting the key makes the data unreadable.
- Yes, a short order form for each step. You commit one step at a time, never to a long contract up front, and each step has a defined scope before it starts. That way you decide whether to go further on what you've already seen.
- It will, sometimes, which is why there are checks at every step. A named owner reviews its work, and anything touching price, scope or terms waits for a person. We record every run, so you can see what it did, correct it, and stop it at once. Each correction sharpens what it does next.
- Each setup is made to measure, so we scope the price with you and quote it on the call. No per-seat fees.