top of page

Graph Engineering: The New Paradigm

Nov 5
9 min read

AI often disappoints small businesses for a simple reason: it is treated like magic software instead of like a smart new employee.


A smart employee can do good work, but only if the task is clear, the right files are nearby, the tools work, and someone checks the output before it affects customers or money. The same idea explains four terms you may hear more often now: prompt engineering, context engineering, harness engineering, and graph engineering.


These terms are still new and informal. Different vendors may use them in slightly different ways. The useful part is not the label. The useful part is knowing what each one does, so you can make better decisions about AI in your business.


Overhead view of paper invoices, coloured folders, and a calculator on a wooden workbench
AI works better when the task and supporting documents are clear.

Think of AI as a smart new employee


Imagine hiring a bright new assistant.


You would not say, “Handle finance,” then walk away. You would explain the job, give them the right records, show them the tools, set rules for approvals, and tell them when to ask for help.


That gives us a simple way to understand the four concepts.


  • The prompt is the task you give the employee.

  • The context is the files on their desk.

  • The harness is the office, tools, forms, rules, and manager reviewing their work.

  • The graph is the team process, where each person handles a defined step.


This is also where AI workflow automation becomes more realistic. The goal is not to “replace everything.” The goal is to help routine work move from one step to the next with fewer manual handoffs, while keeping people involved where judgement matters.


Prompt engineering means giving a better task


Prompt engineering is the practice of writing instructions that help an AI tool give a useful answer.


For a small business owner, this does not need to be complex. It usually means being more specific about:


  • What you want done

  • What format you want back

  • What rules the AI should follow

  • What it should avoid

  • What to do when information is missing


Think of the prompt as the instruction you give the smart new employee.


A weak prompt might be:


Check this invoice.

A better prompt would be:


Read this invoice and extract the supplier name, invoice number, invoice date, due date, subtotal, VAT, total amount, and bank account details. Return the result as a simple table. If any field is unclear or missing, write “Needs review” in that field.

That second version gives the AI a clear job. It also tells the AI what to do when it is unsure.


Practical tips for better prompts


Start with the outcome.


Say what a good answer looks like. If you want a checklist, ask for a checklist. If you want a table, ask for a table. If you want a short customer reply, say how long it should be.


Add rules.


For invoice work, useful rules might include:


  • Use the invoice total exactly as shown.

  • Do not guess missing values.

  • Mark unclear dates for review.

  • Keep currency as shown, such as ฿ or THB.


Give examples when the task repeats.


If you process many similar invoices, show a correct example. A good example can save long explanations.


Ask for uncertainty to be visible.


This is one of the most useful habits. Tell the AI to flag anything it is unsure about. For business tasks, a clear “Needs review” is safer than a confident guess.


Context engineering means giving the right files


Context engineering is about giving the AI the right background information at the right time.


If the prompt is the task, context is the pile of files on the employee’s desk. A smart person still needs the purchase order, supplier list, payment terms, and company policy to do the job well.


AI is the same. Without context, it may produce a polished answer that misses your real rules.


For example, if you ask AI to check an invoice, it may need:


  • The invoice itself

  • The matching purchase order

  • Your approved supplier list

  • VAT rules used by your business

  • Payment terms, such as 30 days after invoice date

  • A list of people allowed to approve larger payments


This is why context engineering matters. It helps the AI use the information that belongs to the task, rather than relying only on general knowledge.


Practical tips for better context


Keep source files clean.


AI works better when documents are named clearly and stored in predictable places. A file called `Supplier ABC invoice 2025-01-15.pdf` is easier to handle than `scan_2398_final_new.pdf`.


Separate current rules from old rules.


If payment policies change, remove or archive the old version. Giving AI five versions of a policy can cause confusion.


Give only what is needed.


More context is not always better. If the task is to verify an invoice total, the AI may not need your full customer database. Good context is relevant context.


Use trusted sources.


For business processes, connect AI to approved documents, not random copied notes. If the source is messy, the output often will be too.


Harness engineering means building the working environment


Harness engineering is less familiar, but the idea is simple.


If the prompt is the task and the context is the files, the harness is the office setup around the employee. It includes the tools they can use, the forms they fill out, the permissions they have, the checks they must pass, and the manager who reviews their work.


In AI terms, the harness is the system that surrounds the AI model.


It may include:


  • A form where staff upload invoices

  • A connection to accounting software

  • A rule that checks whether totals match

  • A list of exception types

  • A human approval step

  • A log of what happened

  • A limit on what the AI is allowed to change


This is where many business AI projects become practical. A chatbot alone may answer questions. A harness can help turn an answer into a controlled business process.


Practical tips for a good harness


Limit what the AI can do.


For example, let AI read invoices and prepare draft entries, but do not let it pay suppliers without approval.


Add checks before action.


If the invoice total does not match the purchase order, the system should stop and flag the issue.


Keep an audit trail.


For money, customer records, or compliance-related work, you should be able to see what the AI extracted, what it changed, and who approved it.


Make human review easy.


The best review screen shows the original document next to the extracted fields. A person should not need to search through folders to check the work.


Close-up view of a paper checklist beside an invoice and a red review stamp
Clear review steps help keep AI-assisted work under control.

Graph engineering means designing the process


Graph engineering sounds technical, but the business idea is familiar. It means designing a workflow where different steps happen in a defined order.


Using our employee analogy, a graph is a team with a process.


One person extracts invoice details. Another checks the supplier. Another compares the invoice to the purchase order. Another reviews exceptions. The work moves step by step.


In an AI system, those “people” may be AI calls, software checks, database lookups, and human approvals. The graph defines what happens first, what happens next, and what happens when something goes wrong.


For an invoice, the process might look like this:


  1. Receive an invoice by email or upload.

  2. Extract key fields.

  3. Validate the supplier against an approved list.

  4. Match the invoice to a purchase order.

  5. Check the totals, VAT, currency, and due date.

  6. Flag exceptions.

  7. Send exceptions to a human reviewer.

  8. Prepare an accounting entry only after approval.


This is powerful because each step has a clear job. It is also safer than asking one AI prompt to “process this invoice” and hoping it handles everything correctly.


A key point: multi-step workflows are not “self-improving AI.” They do not magically teach themselves to run your business better. They are structured processes. They can be monitored, adjusted, and improved by people, but they are not automatically becoming smarter on their own.


Invoice processing shows how the four ideas work together


Let’s put the four concepts into one practical example.


A small business receives supplier invoices by email. Staff spend time opening PDFs, copying details, checking totals, and asking managers to approve anything unusual.


A basic AI-supported process could work like this.


Extract the invoice data


The prompt tells the AI what to pull out:


  • Supplier name

  • Invoice number

  • Invoice date

  • Due date

  • Line items

  • VAT

  • Total amount

  • Currency

  • Bank details


The context includes the invoice PDF and any linked purchase order.


Validate the details


The harness checks the extracted fields against business rules.


For example:


  • Is the supplier on the approved list?

  • Does the invoice number already exist?

  • Does the total match the line items?

  • Does the VAT look consistent?

  • Does the bank account match the supplier record?


Some checks may use AI. Others may be simple software rules. For totals and duplicates, simple rules are often better.


Flag exceptions


The graph decides what happens next.


If the supplier is approved and the totals match, the invoice can move forward. If the supplier is new, the invoice total is higher than expected, or the bank details changed, the system marks it for review.


Useful exception labels might include:


  • New supplier

  • Duplicate invoice number

  • Amount mismatch

  • Missing purchase order

  • Changed bank account

  • Low confidence extraction


Send to human review


A person reviews exceptions before anything is posted or paid.


This is not a failure of AI. It is good process design. AI can handle repetitive reading and checking, while people handle judgement, approval, and risk.


How the four concepts compare


Concept

Employee analogy

Main question it answers

Practical small business use

Prompt engineering

The task you give the employee

Did we ask clearly?

Tell AI exactly what to extract, write, classify, or check

Context engineering

The files on their desk

Does AI have the right information?

Provide invoices, policies, supplier records, purchase orders, and examples

Harness engineering

The office, tools, rules, and manager

Is the work controlled and checked?

Add forms, software connections, approval steps, logs, and limits

Graph engineering

A team with a defined process

What happens step by step?

Move work through extraction, validation, exception handling, and review


The progression matters. Better prompts help. Better context helps more. A good harness makes the work safer. A clear graph turns scattered AI tasks into a process people can trust.


Where to start without making it complicated


Start with one repetitive task that has clear inputs and clear outputs.


Good first candidates often include:


  • Invoice intake

  • Customer email drafting

  • Quote preparation

  • Job report summaries

  • Document classification

  • Internal knowledge search


Pick a task where the current process is already understood. If nobody agrees how the work should be done, adding AI will not fix that.


Next, write down the current process in plain language.


For invoice processing, that might be:


  1. Receive invoice.

  2. Check supplier.

  3. Match purchase order.

  4. Check amounts.

  5. Ask for approval if needed.

  6. Enter into accounting software.


Then decide where AI can help.


AI may be useful for reading messy documents, summarising notes, drafting replies, or finding information. Simple rules may be better for maths, duplicate checks, approval limits, and dates.


Build a small version first.


Use a sample set of real but non-sensitive documents if possible. Test the process. Look at mistakes. Improve the prompt, clean up the context, add checks, and adjust the workflow.


Keep people in the loop, especially at the start.


For money-related work, let AI prepare. Let people approve.


Common mistakes to avoid


The biggest mistake is expecting one prompt to run a whole business process.


A prompt can help with a task, but business work usually needs files, rules, tools, and review. If the process handles payments, customer data, or legal documents, control matters.


Another mistake is giving AI too much context.


Dumping every document into a tool can create noise. Give the AI the documents it needs for the job at hand.


A third mistake is letting AI guess.


For internal notes, a rough draft may be fine. For invoices, contracts, pricing, or customer commitments, guessing is risky. Tell the AI to mark missing or unclear information.


Businesses also forget to separate AI work from rule-based checks.


AI can read an invoice. A calculator should check the totals. A database should check for duplicate invoice numbers. Use the right tool for each step.


The last mistake is skipping review because the demo looked good.


A smooth demo does not prove a process is ready for daily use. Test edge cases, unclear documents, duplicate files, changed bank details, and missing purchase orders.


A practical way to think about the next step


If the terms feel new, keep the employee analogy in mind.


Give the employee a clear task. Put the right files on their desk. Give them the correct tools and rules. Create a defined process for the team. Check important work before it goes out.


That is the practical path from simple prompt engineering to useful automation.


If you are exploring AI workflow automation for invoices, admin, customer support, or internal operations, contact us for a free consultation. We can help you choose a realistic first workflow, identify the right checks, and design a process that fits how your business already works.


bottom of page