How we work
How AI and automation get from pilot to daily use.
We start with the problem and today's numbers, build in small steps, and keep people in charge of the decisions that matter. Your team sees results early, and keeps using them.
We start with the problem and today's numbers, build in small steps, and keep people in charge of the decisions that matter. Your team sees results early, and keeps using them.
When not to automate
Some problems don't need a robot or an AI. We'll say so.
If your pricing is wrong, fix the pricing. If nobody owns a process, a machine won't fix that. If the defect is made upstream, a faster camera just counts the scrap sooner.
We're also the wrong firm if the equipment is already chosen and you only need someone to install it, if you want a chatbot built this week, or if the work involves controlled defence technology. You'll hear that on the first call, along with who might suit you better.
What we believe
New software won't fix an old process. We change both, then test them in real work.
Nothing moves to the next step until it has proved itself, and that goes for the new way of working as much as for the technology.
-
The process and the machine are one job.
A camera, a robot or an AI agent changes who decides, who checks and who answers for the result. Leave the roles and measures as they were, and the new system turns into an expensive side project. Between them, the two founders cover both halves.
-
A pilot is a step. Daily use is the result.
We judge our work by whether real people use it, on real parts or real cases, against numbers agreed at the start. So before anything gets built, we write down today's numbers.
-
Safety checks built in early get you live sooner.
Test sets, checks before release, human sign-off and a record of every decision go in from week one. Added early, they help a system launch. Bolted on at the end, they block it.
From first call to daily use
Nothing moves to the next stage until both founders approve it.
-
Stage 1
Find the problem
In the business: How the work runs today, where it gets stuck, the numbers and the data. In the technology: Where AI and automation can help, whether it will work, how good the data is, what could go wrong. The stage ends with a ranked list of what's worth doing, and today's numbers.
-
Stage 2
Design
In the business: How the work will run, who owns what, who decides. In the technology: How the system is built, how it gets tested, what counts as a pass, the checks before release. The stage ends with a design both founders stand behind.
-
Stage 3
Build
In the business: New workflows, dashboards, reports that build themselves. In the technology: The AI or vision system and its data, running on devices or in the cloud, including the engineering on the robot cell itself. The stage ends with a working system on your real data, in your own environment.
-
Stage 4
Run
In the business: Regular reviews, people using it, return on the money against today's numbers. In the technology: Watching it in use, learning from anything that goes wrong, deciding when to retrain. The stage ends with a system in daily use, with an owner in your business.
Every EliteWorks project runs in four stages: find the problem, design, build, then run it in daily use. A stage only ends when both the business lead and the technology lead are satisfied with the results.
Rahul Nath signs off the business side: who owns the process, what gets measured, revenue numbers that add up and people actually using it. Arjun Murthy signs off the technology: data, models, testing and release.
| Stage | Business side (Rahul) | Technology side (Arjun) | What you have at the end |
|---|---|---|---|
| Find the problem | How the work runs today, where it gets stuck, the numbers and the data | Where AI and automation can help, whether it will work, how good the data is, what could go wrong | A ranked list of what's worth doing, and today's numbers |
| Design | How the work will run, who owns what, who decides | How the system is built, how it gets tested, what counts as a pass, the checks before release | A design both founders stand behind |
| Build | New workflows, dashboards, reports that build themselves | The AI or vision system and its data, running on devices or in the cloud, including the engineering on the robot cell itself | A working system on your real data, in your own environment |
| Run | Regular reviews, people using it, return on the money against today's numbers | Watching it in use, learning from anything that goes wrong, deciding when to retrain | A system in daily use, with an owner in your business |
Is it worth automating?
Before anyone prices a machine, we check the job is worth automating.
The Automation Decision Method checks whether a task is worth automating before anyone quotes for a machine. It pins down what the automation should save, finds the real bottleneck, fixes the process first, tests the payback three ways, and works out whether a paid trial is worth it before you commit money.
Built on well-known management methods, including the Theory of Constraints.
-
01
Name the buffer it should remove
Every operation hides a buffer: extra stock, spare capacity or waiting time. We name the one the automation should cut, and what it costs you today.
-
02
Find the real bottleneck
We put every step's capacity and workload in one table. Automate a step that isn't the bottleneck, and you add cost without adding a single extra unit.
-
03
Fix the process first
Automate a broken process and it just breaks faster. We cut the waste, the rework and the steps nobody owns before anything is bought or built.
-
04
Test the payback three ways
Does it pay? Will the volumes and prices behind the numbers stay true? And will the edge last once a competitor buys the same machine?
-
05
Know when a trial is worth paying for
Before you commit money, we work out the most it's worth spending to find out more. Sometimes a paid trial earns its fee. Sometimes you already know enough to decide.
Accuracy you can trust
High accuracy can still mean a bin full of good parts.
When defects are rare, a camera that catches most of them can still throw out more good parts than bad ones. A demo on a sample packed with defects hides that.
So every vision result we give you answers two plain questions at the defect rate your line really runs: of the parts it flags, how many are really bad, and how many good parts get thrown out.
See the worked exampleChecks before go-live
Four checks before any AI reaches your people.
Every AI system passes four checks on its way to your users: development, testing, a small release and full production. Each check needs proof on paper, and an owner in your business approves before the system moves on.
Everything we build passes all four, under two rules. The tests are built separately from the system, since a system that tests itself proves very little. And a failed check is useful: it caught a problem before your people did.
Check 1: Development
What has to be true
- The job the AI does is clearly limited. What counts as a pass is written down. Tests are built from your real cases. Risks are listed.
Who approves
Your business owner and our technology lead
Check 2: Testing
What has to be true
- It meets the agreed pass marks. Safety and attack tests are run, including attempts to trick the AI where that applies. Links to your systems are tested. What the AI may do alone is set for each action.
Who approves
Our technology lead and your security contact
Check 3: Small release
What has to be true
- Instructions for running it and rolling it back are written. Monitoring and alerts are live. The first users are briefed. Everyone knows who to call if something goes wrong.
Who approves
Your business owner
Check 4: Production
What has to be true
- Results from the small release are checked against today's numbers. Every decision is recorded. Monitoring and reviews are scheduled. The call to roll out, extend or stop is recorded with the reasons.
Who approves
Your business owner and both EliteWorks founders
What AI may do on its own
You decide what the AI may do on its own, one action at a time.
The Agent Autonomy Ladder sets how much an AI agent may do on its own, action by action, across four levels: Inform, Recommend, Act with approval and Act. Anything that affects an important decision, or sends a message to a customer, stops at Act with approval.
The same AI can be trusted to summarise a document alone and still need a person's okay before it emails a customer. We set a level for every action, and the software enforces it. A policy document alone is not enough.
- Level 1
Inform
The AI gathers and reports. A person does the rest.
- Level 2
Recommend
The AI suggests what to do and why. A person decides.
- Level 3
Act with approval
The AI gets the action fully ready. A person approves it before it runs, and the approval is logged.
- Level 4
Act
The AI acts on its own within set limits, and every action is logged for review.
Summarise a document
- Level
- Act
- Why
- Nothing leaves the company, and a person reads the summary before relying on it.
Recommend a refund
- Level
- Recommend
- Why
- A refund is money. A person decides.
Email a customer
- Level
- Act with approval
- Why
- Anything that goes to a customer needs a person's approval first.
Is your business ready for AI?
Six things decide whether AI makes it into daily work.
The EliteWorks Readiness Model scores how ready a company is to get real value from AI in daily work, across six areas: strategy and value, data and visibility, process and workflow, technology and delivery, governance and risk, and people. It scores what people actually do, not what they plan to do.
-
Strategy and value
How AI projects get picked, owned and funded, and whether anyone tracks the payoff.
-
Data and visibility
How fast you get an answer you trust, whether key measures have owners, and whether the data for your top idea is ready.
-
Process and workflow
Whether core processes are written down and measured, how much routine work is still manual, and who watches the automation.
-
Technology and delivery
How far your AI has got, how its answers are checked, and how changes are released.
-
Governance and risk
Whether there's an AI policy people follow, how risks are checked before launch, and who can see what the AI did.
-
People and operating model
Where AI skills sit, how people are prepared, and how leaders review progress.
The same six areas power our AI readiness check and shape the Business and AI Diagnostic. A weak score on data or governance pulls the whole result down, because AI built on shaky data or loose rules doesn't last.
How we measure
Today's numbers first. You set the pass mark.
Before we build, we agree three to five numbers with the person who owns the result, and measure where they stand today. Without that starting point, any improvement is just an opinion.
You write, or help write, the test the system has to pass: what it must do, on which cases and how well, before more people get it. Pass, and more people get it. Fail, and we fix it or stop, with the results on record.
Independence
You see every option we looked at, including doing nothing.
Each option comes with the reason we passed on it. If a founder has any link to a vendor you're considering, we tell you in writing before you decide.
How we charge
No surprises on the invoice.
- The price is in a short written proposal you see before anything begins.
- Your first conversation is with a founder, not a salesperson.
- You help decide what success looks like, before we start.
- We build inside your own systems from week one, so everything we make is yours as it's made.
- We don't cut our price in exchange for using your name.
- The written proposal is the whole deal. Any change is agreed in writing before either side acts on it.
- Want a mutual non-disclosure agreement (NDA) first? Ask, and we'll sign one before you share details.
- A data processing agreement is available on request, before work begins.
- Before we touch your data, we tell you in writing where it will be processed and stored, and by which tools.
- Work is built inside your environment wherever your policies need it.
- We never name a client, or describe their work, without written permission.
Send us your security questionnaire. We'll answer every question, including the ones where today's answer is no.