Technology
Technology
Perception, intelligence, orchestration, action, verification. The layers every PTAS AI workflow runs through.
What this section is for
How PTAS AI is built, for the people who have to sign off on it
AI Automation says what we can do. Products say what we built. This section says how it works underneath, so an architect can check it before a CIO approves it.
Most automation pitches stop at the demo: a clean invoice, a confident answer, a posted entry. The questions that follow in a real evaluation are different. What read the document? What decided? What posted? Who can see why, and how long after the fact? These pages answer those questions in plain language without turning into developer documentation. There are no endpoint references here and no model names. Models change. The architecture around them is what you are buying.
Each page covers one technology the way we would cover it in an architecture review. What the thing is. Where it sits in the workflow. What it does inside DocPro and FinSuite. What it cannot do, and which control sits around that gap. Every page keeps the same distinctions explicit: what a model does, what a rule does and what a person does. When those blur, you end up with automation nobody can audit, and an auditor will eventually ask.
Architecture
Perception to verification, in order
Every workflow we build runs through the same five layers. A document or a record enters at the top. A verified, recorded outcome leaves at the bottom. Each technology page tells you which layer it belongs to.
The five layers, as one invoice passes through them
- 01 Perception Pixels, scans and files become text, regions and an image quality score
- 02 Intelligence Text becomes typed fields with a confidence on each, and one proposed next step
- 03 Orchestration Agents, rules and people are sequenced into a single workflow that holds state
- 04 Action The workflow posts, routes, notifies and updates the system of record through declared integrations
- 05 Verification The result is read back, recorded with its reasoning, and reviewed wherever a rule says so
Nothing in the action layer runs on model output alone. Between intelligence and action there is always a rule, and above a threshold there is always a person.
The pages
The eleven technology pages
One technology per page, each with the same five questions answered: what it is, where it sits, what it does in the products, what it cannot do, and the control around that.
Agentic AI Architecture
Intelligence layer. An agent is a role with context, tools, rules and a verifier. How one is built, bounded and audited, and why that is not a chatbot.
Document AI
Perception and intelligence. OCR, layout, extraction, confidence and validation, and the exact point where each one stops.
Generative AI in Business Workflows
Intelligence layer. What language models do here (normalise, classify, explain, draft) and what they are never allowed to do.
Machine Learning
Intelligence layer. Classification, field confidence, matching and anomaly detection, measured on your documents, with the feedback loop that retrains them.
Computer Vision
Perception layer. Image quality, deskew, regions, tables, stamps and signatures: what is detected before a single character is read.
RAG and Knowledge Systems
Intelligence layer. How agents consult policies, master data and history before deciding, with every source traceable.
AI Orchestration
Orchestration layer. How agents, models and rules are coordinated in one workflow: handoffs, state, retries and escalation.
Workflow Orchestration
Orchestration and action. The engine underneath: steps, queues, approvals, SLAs and integrations, modelled, run and audited.
Business Rules
Orchestration layer. Thresholds, policies, validation and routing kept as explicit rules, so decisions can be read and changed.
Human-in-the-Loop
Verification layer. Review designed as a control: thresholds, queues, escalation and the feedback that trains the next run.
AI Security and Governance
Verification layer. Permissions, data boundaries, audit logs and behaviour controls for models and agents in production.
Why the bottom layers matter
Three ways to read a document, and what each hands you
The perception and intelligence layers are where most document projects quietly fail. The figure shows why. Three generations of document reading return three different things, and only one of them can be acted on without a person re-keying it.
What each layer hands you
A block of text in reading order. Correct characters, no idea which of them is the invoice total.
Named fields, as long as the layout matches a template someone built. A redesign breaks it silently.
Typed fields found by context, validated against your masters and arithmetic, each carrying a confidence score and a route to review when it falls short.
The difference is not accuracy on a clean sample. It is what happens on the document nobody anticipated.
The technologies on the first five pages, vision, document AI, machine learning and generative AI under constraint, are what move a workflow from the first row to the third. The orchestration, rules and review pages are what let you act on the third row without a person re-checking every field. If you are evaluating DocPro, read them in that order. If you are evaluating FinSuite, start with the orchestration pages, because the hard part of a finance workflow is rarely the reading.
Where to start
Three reading paths, depending on what you have to decide
You do not need all eleven pages to make a decision. Pick the path that matches the question on your desk.
If you are evaluating agents
Start with the architecture
Read Agentic AI Architecture, then AI Orchestration and Business Rules. Together they answer what an agent may decide alone and who can change that. The agentic AI vs RPA guide is the shorter version for a steering committee.
If your problem is documents
Start with perception
Read Document AI and Computer Vision, then Machine Learning for how confidence is measured. Then test one of your own documents rather than taking our word for it.
If audit or security has to approve
Start with the controls
Read Human-in-the-Loop and AI Security and Governance, then the security page for deployment options, data handling and the controls list in precise terms.
Keep reading
Where this technology shows up
AI Automation
What PTAS AI does with these layers, capability by capability.
Read more →DocPro
The document intelligence product built on the perception and intelligence layers.
Read more →FinSuite
The finance workflows the orchestration and action layers run every day.
Read more →Security at PTAS AI
Deployment options, data handling and access control, stated as controls.
Read more →Next step
Check the architecture against your own workflow
Bring one process. We will map it to the five layers, say what is model, rule and person at each step, and tell you where we would not automate yet.