Elias MercerAI engineering and consulting Describe a workflow

I test AI on your real work before you pay to build it.

Pick one task your team does by hand. I test AI on real cases from it and build only what passes. You can stop after any stage and keep what you paid for.

Opens a draft in your email app.

Goes into the ordering system

  • SupplierOrchard Lane Foods
  • NumberF-2291
  • Date15 Oct 2026
  • Line 118 × Olive oil 5 l
  • Line 2Printed 60, changed by hand to 48.
    Needs a person.
  • Line 3200 × Paper sacks
  • Total€716.00

One field needs a person. Nothing is entered until it is checked.

1 entered

1 waiting for a person

0 guessed

An illustration with made-up data, not a client. Point at a row to see where it was read from.

Four stages. Stop after any of them.

Each stage answers one question and ends with something you keep. The example is a wholesale distributor where three people type supplier invoices into the ordering system. It is an illustration, not a client.

  1. What is worth testing, and what should wait?

    Fixed fee, agreed up front2 to 3 weeks

    I sit with the people who do the work. You get a ranked list of ideas, a value estimate with its assumptions, the risks and a 90-day plan. If a plain script beats AI, I say so.

    You can stop here. Your team or another partner can run with the plan.

    Ranked ideas, supplier ordering

    1. Read supplier invoices into the ordering systemPilot thisAbout 400 a week, and each one can be checked against its PDF.
    2. Draft replies to routine customer emailsLaterUseful, if a person approves each reply before it goes out.
    3. A chatbot on the websiteWaitNobody owns it yet, and nobody has said what success looks like.
    4. Predict which quotes will closeNot nowThe sales history has too many gaps to learn from.

    Top idea, rough value. About 15 hours a week, if 9 in 10 invoices are read correctly.

  2. Does it work on your own cases?

    Fixed fee, agreed up front3 to 5 weeks

    We agree the pass mark before I write any code. Then I build a small version and run it on your real cases. You get the design, a working prototype and the results, failures included.

    You can stop here if it fails. You have spent a pilot's budget and you know the answer.

    Pilot results, 30 real invoices

    Pass mark, agreed before any code. At least 27 of 30 read correctly. Anything unreadable is flagged for a person, never guessed.

    Read correctly
    28
    Flagged for a person
    2
    Wrong and not flagged
    0

    The two flagged were a faded scan and a delivery note with handwritten changes.

    Passed28 correct against a pass mark of 27.

  3. How does it get into daily use?

    Fixed fee, agreed up front1 to 3 months

    I build it into the systems you already run. The code lives in your repository and runs in your cloud account. You get tests, a runbook and a handover session.

    You can stop here. Your team runs it from now on.

    What went live

    A supplier emails an invoice

    The reader pulls out supplier, lines and totals

    Sure of every field

    Straight into the ordering system

    Unsure of anything

    A person checks it, then it goes in

    The pilot's 30 test invoices stay with the team, so they can re-run them after any change.

  4. Who looks after it once it is live?

    Monthly fee, or by the hourMonth to month

    Suppliers change layouts and models get replaced. I fix what breaks, retest, and add what you ask for. You get a short note every month.

    You can stop any month with 30 days' notice. Everything built stays with you.

    Monthly note, month 3

    • A supplier changed its invoice layout. Retested on 20 of its invoices and fixed the reader.
    • A newer model came out. Tested it on the pilot invoices and kept the current one, which read dates more reliably.
    • Added credit notes, as the team asked.

What you can hold me to.

These apply to every job, large or small.

  • The price is agreed before work starts.

    You get a written scope and a fixed fee first. Anything outside the scope is quoted before I do it.

  • You own what you paid for.

    The custom work is yours. Where it uses tools I already own, you get a licence to run and maintain it.

  • Your data stays under a written agreement.

    Before I see real data we sign an NDA and a data processing agreement. I use provider terms that exclude training on your data.

  • Every estimate shows its assumptions.

    Estimates help you decide. They are not a promise of savings.

Two more. If I recommend work I could sell you, I say so. And I am on call only if you buy support, with the response times written down.

I test AI on your real work before you pay to build it.

We pull 30 real invoices from last month, the messy ones included.

Before I write any code, we agree the pass mark. Here it is 27 of 30.

The prototype reads each one. What it can't read, it flags for a person. It never guesses.

28 of 30. It passed, so now it is worth building. Had it failed, you would stop here with a pilot's bill and a clear answer.

An illustration of a pilot, not a client result.

What changes for the team.

The same example, before the build and after it.

Today

Three people, about 6 hours a week each, on every invoice.

  1. Open the email
  2. Download the PDF
  3. Type the supplier and number
  4. Type each line
  5. Check the total
  6. File it

About 400 times a week.

About 18 hours a weekOne block is one hour of typing

After

The reader enters what it is sure of. People see only what it flags.

Today's invoices

Arrived
80
Entered
72
Need a person
8

About 3 hours a weekChecking what was flagged

An estimate from the example, not a result. It assumes 9 in 10 invoices are read correctly and volume stays near 400 a week.

Is this a fit?

A quick check before you write.

Probably a fit

  • A task eats hours every week: documents, emails, data entry or reports.
  • Leadership wants an answer on AI before money goes into a build.
  • Someone on your team will own the result once it is live.
  • You want to own what is built, not rent it.

Probably not

  • You need a chatbot live within days, without testing or human review.
  • You need a formal security or compliance audit.
  • You need someone on call around the clock.
  • You want AI to act alone where mistakes are expensive.

Before you sign.

  • You work with me directly, from the first email to the handover.
  • When the work ends, I delete my copies of your data and confirm it in writing.
Do we need to know what to build?

No. A list of ideas is enough, and so is one process everybody complains about. The assessment sorts out what is worth doing.

Does an assessment commit us to more work?

No. It is a complete piece of work. Use the plan with your own team, another partner or me.

What if the pilot fails?

Then you know before paying for a full build. We agree the pass mark before the pilot starts, so there is no arguing about the result afterwards.

What happens to our data?

Before any real data moves, we sign an NDA and a data processing agreement that names every AI provider involved. I prefer processing in the EU and can work entirely inside your own cloud account.

Can you work with the tools we already have?

Usually. I start from the systems you already run, and what they let me access decides what is possible.

Is there a smaller way to start?

Yes. A written second opinion on an AI feature or automation your team already built: what will break, and how to fix it. It takes about a week.

Can you work inside our team?

Yes. I work with your developers on hours agreed each month, invoiced as worked or reserved in advance.

Tell me about one workflow.

A few sentences is enough. I'll tell you whether an assessment is worth doing, or whether you need something else.

studio@eliasmercer.ai
  1. You email. I reply within 2 working days and suggest a 30-minute call.
  2. Within 3 working days of the call, you get a written scope and a fixed fee.
  3. If you accept it, work usually starts within 2 weeks.

Tostudio@eliasmercer.ai

SubjectA workflow we want to improve

Hello,

Today, what happens, step by step.

It is done by who, about how often.

We have already tried anything, or nothing yet.

Four blanks I can't guess. They need a person.

Open this as an email