Solution — Workflow & Approval Engine
The rules engine that runs the chain: request to invoice, without re-typing
Request → approval → work order → dispatch → invoice → report, as one state machine rather than six systems joined by email. Every transition has an owner, a deadline and an audit record.
KodDelta builds the workflow engine underneath operational processes: a rule-based state machine that routes requests by amount and department, escalates when an approver does not respond, enforces margin and spending limits server-side, and posts the result to your accounting system. A single focused workflow costs $3,000–6,000 and reaches production in 2–4 weeks, source code delivered.
This page covers the engine layer in depth. For the wider question — which back-office and outsourced work should move into software at all — start with business process automation.
The bottleneck is the handshake, not the department
In most mid-sized companies each department works acceptably on its own. The delay accumulates at the points where work changes hands:
- Requests waiting on an approver. A purchase requisition or a discount request sits in an inbox with no deadline attached to it and no visibility of how long it has been there. Nothing is wrong; nothing is moving.
- The same record typed three times. Sales enters an order in a spreadsheet, operations re-types it into the production schedule, finance re-keys it into accounting. Each pass adds a chance of a wrong figure and none of them adds information.
- No answer to "where did this stop". When an order is late, establishing which approval was held and which step ran long means asking four people and reading an email thread.
- Rules that live in someone's head. The discount floor, the approval threshold, the deputy when the manager is away — known by the people who have been there longest, and unenforced when they are not in the room.
Five workflow chains we automate
| Workflow chain | Manual flow today | Rules engine |
|---|---|---|
| Procurement and expense | Paper or email requests, signatures chased in person | Threshold-based routing, tokenised mobile approval, budget line checked before submission |
| Quotes and discounting | Spreadsheet pricing, discount agreed by phone | Server-side margin floors, branded PDF generated from the priced record, exceptions routed for sign-off |
| Order to work order | Morning briefing and printed job sheets | Approved order creates the sequenced production route with materials reserved |
| Pick, pack and dispatch | Printed picking lists, shelves searched by eye | Barcode-guided picking on a handheld, dispatch note and e-waybill raised on completion |
| Customer status updates | Phone calls answering "where is my order" | Event-triggered SMS and email at each milestone, plus a status page the customer can check |
The same engine underpins production systems we have built: Elevatora SAHA for lift maintenance and field service dispatch, Elevatora İMALAT for lift-component manufacturing tracking, SEMP Group as a group-wide operations platform, and PAP for pharmaceutical import/export process tracking. What each does is described on cases.
Governance: delegation, escalation and audit
The engine is built so that policy changes are configuration, not a release:
-
Hierarchical limits and delegation
Thresholds are set per department and per cost category — for example, under $2,000 approved by the department head, over $20,000 routed to the finance director and managing director in sequence. Delegation rules cover leave and role changes without editing every open request.
-
SLA timers and escalation
Each step carries a target duration. When an approver does not act within it, the request escalates to a named deputy and the delay is recorded against the step, so cycle-time reporting reflects where the time actually went rather than an estimate.
-
Tamper-evident audit trail
Every state transition stores the user, timestamp, IP address, the values before and after, and any comment. Rule changes are versioned the same way, so an auditor can reconstruct which policy was in force on the date a given request was approved.
-
Separation of operations and accounting
Approved transactions post to your existing ledger — SAP, Dynamics, NetSuite, or Logo, Mikro and Netsis — through the integration layer, without export and re-import. Statutory accounting stays where your accountants expect it. See integrations.
The AI agent layer on top of the engine
A rules engine needs structured input. In practice a large share of work arrives unstructured: an email in a shared inbox, a PDF from a supplier, a message describing a fault in a sentence. The AI layer is what turns that into something the engine can route.
Free text becomes a record
An inbound email or message is read, the relevant fields extracted — customer, product, quantity, date, site — and a draft request created in the workflow. A person confirms it rather than typing it. Fields the model is unsure of are flagged rather than guessed.
Document classification and form filling
Incoming invoices, delivery notes and certificates are classified by type, their reference numbers extracted, and matched to the open order or work order they belong to. Mismatches go to a review queue; matches complete without a touch.
Summaries for the approver
A long request, a supplier contract or a service history is summarised to the points the approver needs to decide, with a link to the source. The decision stays with the person; the reading is done before they open the message.
Natural-language answers from your own data
Staff and customers ask questions in plain language and get answers from your records and documents through retrieval-augmented generation (RAG) — the model answers only from your indexed content, with the source cited, rather than from general training data. Detail on enterprise digital assistants and in what a corporate AI assistant is.
The AI layer is priced separately from the workflow engine: an enterprise AI assistant or RAG agent over your document set is $5,000–12,000 over 2–4 weeks. It is added to an existing engine, not a prerequisite for one. Full price list on pricing.
What we do not automate
Exception handling is where automation projects overreach. A workflow with more exceptions than standard cases will not straighten out by adding branches — it will accumulate rules nobody can audit. Negotiated pricing, escalated complaints, credit decisions and bespoke specifications stay with a person; the engine gathers the data, drafts the record, and puts the decision in front of them with the evidence attached. We say the same on the business process automation page, where the hire-or-automate comparison sits in full.
Frequently Asked Questions
How long does a workflow automation project take to deploy?
A single focused workflow — purchase requisition approvals or automated quote generation, for example — reaches production in 2–4 weeks and costs $3,000–6,000, with a clickable prototype in 2–4 weeks. A multi-department operations system covering several linked workflows is $8,000–15,000 over 4–8 weeks.
Can managers approve requests from email or mobile without logging into a portal?
Yes. Approvals are sent as tokenised single-use links by email or WhatsApp. The approver sees the request summary, the amount, the budget line and the requester, and approves or rejects from the message. The token is bound to one request and one approver, and every action is written to the audit trail with a timestamp.
Can we move our existing spreadsheet calculation logic into the engine?
Yes. We read the spreadsheet, document the pricing logic, discount tiers, margin limits and validation rules it encodes, and reimplement them as server-side rules with version history. The rules then apply to every user, rather than to whoever holds the current copy of the file.
What can the AI layer do inside a workflow?
It handles the unstructured input a rules engine cannot parse: reading an inbound email or PDF and creating the request record from it, classifying documents, summarising a long request for the approver, and answering process-status questions in natural language from your own records using retrieval-augmented generation. An enterprise AI assistant of this kind is $5,000–12,000 over 2–4 weeks.
Can administrators change approval rules without a developer?
Yes. Spending thresholds, approver assignments, deputy rules, SLA durations, notification templates and additional form fields are configured in an admin screen. Changes take effect on new requests and are recorded with the user and timestamp that made them. Structural changes — a new workflow type, a new integration — are development work.
How do we measure whether it worked?
Record the baseline before launch: average time from request to approval, manual hours per order, and the error and rework rate. The engine timestamps every transition, so the same measures are available continuously afterwards from your own data. We publish no percentage improvements of our own, because the figure depends entirely on the process you started with. The method for putting a number on the manual baseline is in the real cost of manual work and our software ROI guide.
Send us the workflow that stalls most often.
Describe the steps, the approvers and the systems it touches. We return a scope, a timeline and a fixed price for the first module.