Yuvraj Raulji | AI Automation
Automation earns its money on work that is high in volume and low in judgement. The mistake is reaching for a model first: most of what businesses want automated is deterministic, and the parts that genuinely need language understanding are narrower than the pitch suggests. Fix the process, automate what remains, and put a person where it matters.
Quick answer
AI automation is the use of language models and workflow tooling to handle business processes that were previously manual. It covers deterministic workflow, retrieval over a company knowledge base, and agents that plan across several steps and call tools. It suits processes that are repeatable, frequent and low in judgement per instance: quotes, approvals, data reconciliation, exception handling and supplier data processing. The business problem AI automation addresses is capacity. What it cannot do is repair a process that should not exist, and automating one of those makes it permanent.
Best suited for
- Operations with repeatable, high-volume steps
- Approval and quote processes run through email
- Data reconciliation done to a schedule
- Teams whose capacity is the growth constraint
Business problems
The symptom as the person with the problem describes it, what it is costing, and where the technology actually helps.
People are the integration between two systems
Someone exports from one system, reformats it, and pastes it into another. It works, it is invisible in any budget, and it consumes a person indefinitely.
OpportunityThis is the highest-return automation available in most businesses and it rarely needs a model. A workflow with an explicit contract on both sides replaces the person, and the person moves to work that needs judgement.
Approvals run through email
Quotes and authorisations chased through inboxes, with no record of who approved what or how long it took. The cost is cycle time, and cycle time is the thing customers actually feel.
OpportunityModelling the approval chain explicitly. On a B2B commerce platform, five-level chains covering development, orders, quotes and vendor management took 90% of order and quote processing off people and cut approval cycle time by 40%.
The knowledge is in people, and they are the bottleneck
Policy, contract terms and product knowledge held in documents nobody can search and heads that are in meetings.
OpportunityRetrieval over the real documents, answering with citations rather than from memory. This is where language models genuinely earn their place, because the input is unstructured and the output is a reference rather than an action.
Every exception becomes a person
The happy path is automated and everything else lands on a human, so volume growth still means headcount growth.
OpportunityThis is the narrow case where agents are warranted: multi-step handling where the sequence varies per case. With explicit scope, defined tool access and a checkpoint before anything commits.
Approach
Five stages, in this order, on every engagement. Skipping the first one is the most expensive habit in this industry.
Understand
Business modelCustomersProductsOperations
Map the process as it is actually performed, not as it is documented. Count the volume and the time per instance. This step regularly ends the project early by revealing that the work should simply stop, which is the cheapest possible outcome.
Architect
SystemsIntegrationsCustomer journey
Decide what is deterministic and what genuinely needs a model, then design the human checkpoints before anything else. Where the output touches money, stock or a customer commitment, the checkpoint is part of the architecture rather than a phase-two addition.
Build
TechnologyExperienceFunctionality
Workflow tooling for the deterministic majority, retrieval where the input is unstructured, and agents only where the sequence genuinely varies. Wired into the real systems through their APIs and permission models rather than running alongside them.
Optimize
PerformanceCROSEOCustomer experience
Review the failure cases on a schedule, not on complaint. Track how often the checkpoint catches something, because that number is the honest measure of how much trust the automation has earned.
Scale
AutomationAIAnalyticsPersonalization
Extend to the adjacent process only after the first one has run unattended for a quarter. Automation that is expanded before it is stable compounds its own errors, and in operations those errors reach customers.
Capabilities
Workflow
- Approval chains
- Quote processing
- Order exceptions
- Routing and escalation
- Scheduled reconciliation
Language models
- Document extraction
- Classification
- Drafting with review
- Summarisation
Retrieval
- RAG over policy and contracts
- Knowledge base search
- Citation and provenance
- Index freshness
Agents and tools
- Scoped tool access
- MCP integrations
- Multi-step planning
- Action checkpoints
Integration
- Platform APIs
- ERP and CRM
- Queues and webhooks
- Permission models
AI and technology
What each application does, how it works, where the value is, where a person still reviews it, and what it will not do.
Retrieval over company knowledge
- What it does
- Answering internal questions from the documents the business already has, with a reference back to the source.
- How it works
- Documents are chunked and embedded, a question retrieves the relevant passages, and the model answers from those passages rather than from training. The citation is the deliverable as much as the answer.
- Where the value is
- Time recovered from the people who currently are the knowledge base, and consistency in what the organisation tells itself.
- Where a human reviews
- Whoever owns the source document owns the answer. When the retrieval is wrong, the fix is usually the document rather than the model.
- What it will not do
- It cannot answer what is not written down, and it will answer anyway unless it is built to say it does not know. That behaviour has to be engineered and tested.
Document extraction
- What it does
- Turning unstructured inbound documents, such as supplier data and purchase orders, into structured records.
- How it works
- A model reads the document, populates a fixed schema, and the result goes into a review queue where confidence is low.
- Where the value is
- This is the classic case where models beat rules. Twenty suppliers with twenty formats defeat a parser and are unremarkable for an extraction pipeline.
- Where a human reviews
- Low-confidence extractions go to a person by design, and the confidence threshold is tuned against real cost rather than set at a default.
- What it will not do
- It is confidently wrong on edge cases that look ordinary. Without a review queue the errors are invisible until they are expensive.
Agents with scoped tools
- What it does
- Multi-step handling where the sequence varies by case rather than following a fixed path.
- How it works
- An explicit tool list, a bounded set of actions, a step budget, and a checkpoint before anything commits. MCP is one way to define that tool boundary cleanly.
- Where the value is
- Exception handling, which is where automation projects usually stop and headcount usually starts.
- Where a human reviews
- A person approves any action touching money, stock or a customer commitment. This is a permanent design constraint, not a launch precaution.
- What it will not do
- Reliability compounds downward. A step that is 95% reliable is about 60% reliable across ten of them, and operations notice long before that.
Deterministic workflow, which is usually the answer
- What it does
- Rules, routing and approvals implemented as software, with no model involved at all.
- How it works
- Workflow tooling wired into the platform and business systems through their APIs, with the approval model expressed as data.
- Where the value is
- The 90% automation of B2B order and quote processing came from exactly this, alongside a 40% cut in approval cycle time. No model was involved, and none was needed.
- Where a human reviews
- Exceptions route to a named person rather than to a queue nobody owns.
- What it will not do
- It cannot handle genuinely unstructured input. That is the boundary where a model starts to earn its place, and it is further out than most proposals assume.
Architecture
How the pieces sit together, top to bottom. Only technologies actually used appear here.
Trigger
What starts the process: an order, a document, a schedule, a threshold.
WebhooksQueuesSchedulers
Orchestration
The workflow itself, where the sequence and the branching live.
n8nWorkflow engineState machine
Reasoning
Used only where the input is genuinely unstructured. Most steps skip this tier.
LLM APIsClassificationExtraction
Retrieval
Grounding answers in real documents, with provenance attached.
Vector indexRAGCitations
Tools
What the automation is permitted to do, enumerated rather than implied.
MCPPlatform APIsScoped credentials
Checkpoint
Where a person approves before anything commits. Designed in, never added later.
Review queueApproval model
Systems of record
The platforms that hold the truth after the process finishes.
ERPCRMCommerce platform
When to use it
Good fit
- A process performed the same way many times a week
- Clear inputs and a checkable output
- A named person who can own the exceptions
- Volume high enough that the review time is worth it
- A measurable baseline: hours, cycle time, error rate
Think twice
- A process that should be removed rather than automated
- Judgement required on every instance
- No appetite for a human checkpoint on financial actions
- Low volume, where a checklist outperforms a pipeline
- A brief that specifies agents before the process is understood
Comparison
Read this table left to right and stop at the first column that fits. The industry sells it right to left. The measured 90% automation figure on this site came from the leftmost column, and the projects that disappoint are almost always the ones that started at the right.
| Criterion | Deterministic workflow | LLM in the loop | Agents |
|---|---|---|---|
| Best at | Fixed rules at volume | Unstructured input | Varying multi-step cases |
| Predictability | Exact | Statistical | Lowest |
| Setup cost | Low to moderate | Moderate | High |
| Running cost | Low | Per call plus review | Per call, multiplied by steps |
| Fails by | Refusing to proceed | Being confidently wrong | Compounding across steps |
| Review needed | At exceptions | On low confidence | Before every action |
| Debuggable | Fully | Partly | With difficulty |
| Try this first | Almost always | When rules fail | Rarely, and last |
Relevant work
Automation and integration work from the record. The B2B platform is where the approval and quote automation was built, and it was workflow engineering rather than AI.
Custom platform · B2BProcurement and approvals platform
Scalable B2B procurement platform streamlining purchase requests and approvals.
B2B workflowsApprovalsIntegrations
Web platform · ManufacturingEngineering manufacturer platform
A modern engineering brand platform built for discovery and enquiry.
Web platformTechnical SEOContent
3x traffic growth and 45% better engagement
Business outcomes
90%
Of order and quote processing automated
Five-level approval chains covering development, B2B orders, quotes and vendor management, modelled as a data structure rather than assembled from extensions.
Where this came fromFrom a high-traffic B2B commerce platform on Magento 2. This was deterministic workflow automation with no model involved. It appears on an AI automation page because it is the honest benchmark: this is what the simpler technology achieved, and it is the bar anything more complicated has to clear.
40%
Cut in approval cycle time
The number customers and colleagues actually feel, as distinct from the volume of work removed.
Where this came fromMeasured on the same B2B platform, alongside the automation figure above.
30%
Cut in development cycle time
From introducing structured workflows and approval processes to the delivery process itself, rather than to the commerce operation.
Where this came fromRecorded as part of the same period of work. It required no new platform, which is the part worth noticing.
Checkpoint catch rate
How often the human review step finds something wrong. It is the honest measure of how much trust an automation has earned, and it belongs on the dashboard from day one.
All three figures come from workflow and approval automation, not from language models or agents. That distinction is kept deliberately: the AI layer is where this practice is now going, and it does not yet have a delivered engagement with a published measurement. Presenting workflow results as AI results would make this page an example of the thing it warns about.
FAQ
- What is AI automation?
- Using language models and workflow tooling to handle processes that were previously manual. In practice it spans three things: deterministic workflow, retrieval over documents, and agents that plan across steps and call tools. Most useful implementations are mostly the first and a little of the second.
- What should be automated first?
- The process where a person is acting as the integration between two systems. It is high volume, it is invisible in every budget, it needs no model, and it is usually the single highest return available. Start there rather than with the most interesting technology.
- Do I need AI, or just automation?
- Usually just automation. A model earns its place when the input is genuinely unstructured: documents in twenty formats, free-text queries, mixed supplier data. If the input has a shape, rules will be cheaper, faster, exactly predictable and far easier to debug at two in the morning.
- What is RAG?
- Retrieval-augmented generation: the model answers from documents retrieved at question time rather than from what it memorised in training. It is the right pattern for company knowledge because the answer can cite its source and the source can be corrected without retraining anything.
- What is MCP?
- The Model Context Protocol, a standard way to give a model access to a defined set of tools and data sources. Its value in automation is the boundary: the tool list is explicit and enumerable, which is exactly what you want when deciding what an automated system is permitted to do.
- When are AI agents actually the right choice?
- When the sequence of steps genuinely varies per case and cannot be expressed as branching logic. Exception handling is the honest example. If you can draw the flowchart, build the flowchart, because it will be cheaper and it will not surprise you.
- How reliable are agents in production?
- Less reliable than the demo, and the maths is not intuitive. Reliability compounds across steps, so 95% per step is roughly 60% across ten. That is why scope, step budgets and a checkpoint before committing are structural requirements rather than caution.
- Where should a human stay in the loop?
- Anywhere the output touches money, stock, credit or a customer commitment. Design the checkpoint in from the start, because a review step added after an incident is much harder to place well than one that was there from the beginning.
- How do I measure whether automation worked?
- Hours removed, cycle time, error rate and checkpoint catch rate. Record all four before starting. Automation projects without a baseline are impossible to defend in a budget review and impossible to improve.
- What does this cost to run?
- Per-call model cost is usually the smaller half. The larger half is index freshness, review time and the engineering to keep integrations working as the systems either side change. Budget it as an operating cost rather than a project cost.
- Can you automate a process that is not documented?
- You can, and it is the most common way these projects go wrong. Undocumented processes contain steps nobody can justify, and automating them makes them permanent. Mapping first is not bureaucracy, it is where most of the saving turns out to be.
Next step
Let us look at your business model, the systems you already run and where you are trying to get to, before choosing the technology. That conversation is usually shorter than people expect, and it occasionally ends with me saying you do not need the project.