Process engineering
Mapping, measuring and redesigning work processes end to end.
- BPMN 2.0
- SIPOC
- Value Stream Mapping
- Lean
- Process Mining
Forward Deployed Engineering
Forward deployed engineering: process engineering with AI and ML.
KO AI engineers and consultants rebuild your organization's work processes into leaner, AI-based processes, inside your organization, on your data, and measured against a business KPI.
40+
AI / ML systems in production
500+
New AI-based processes deployed at clients
250+
AI agents live in production
10K+
Hours of AI consulting for companies
Last updated
Definition
What is forward deployed engineering?
Forward deployed engineering (FDE) is advanced process engineering that turns legacy business processes into AI-based processes and deploys them inside the organization. A team of engineers and consultants works with the people who do the work, on the real data and systems, and owns both the business case and the build.
KO AI is a pioneer of FDE in Israel: as of September 2026 our team has put 40+ AI / ML systems into production and deployed 500+ new AI-based processes at clients. We combine process engineering with data, AI and ML, and we work on-site with clients in Israel and Europe from our offices in Tel Aviv, Prague and Porto.
The same team that maps the process designs it, builds it, connects it to your core systems and stays with it in production. Success is a business KPI agreed before the build starts, not a number of features.
A discipline, not a tool
What does a forward deployed engineer do?
A forward deployed engineer works where four disciplines meet. It is a rare combination: senior business judgement and deep technical skill in the same team, sitting with the people who run the process.
Operations management
Capacity, bottlenecks, cost and service levels, as they look on the floor.
- Theory of Constraints
- Little's Law
- SLA
- Cost-to-serve
- KPI trees
Project management
Leading change in a large organization: stakeholders, risk, timelines and adoption.
- Agile
- RACI
- Risk register
- Change management
- Stakeholder mapping
AI and ML
Models, data and engineering that hold up in production, not only in a demo.
- LLM
- RAG
- Classification
- Anomaly detection
- Evals
- MLOps
The method
How does FDE work? Five levels of AI process re-engineering.
FDE starts with the business process, not the technology. We measure the process as it runs today, redesign it around what AI and ML can do, and deploy it until the numbers move.
01
Process discovery
Map and measure the current process.
We learn the process as it actually happens, not as the procedure says it does.
- SIPOC
- BPMN 2.0
- Value Stream Mapping
- Process Mining
- Time & Motion
With the people on the floor
Interviews and observation with the people who do the work, mapping every step, decision and hand-off between teams.
From the systems
Logs and data from CRM, ERP and operational systems, to find the real path every case takes.
A numeric baseline
Cycle time, cost per case, error and rework rates, waiting time and SLA compliance.
02
Diagnosis
Find where the time and money go.
We rank the improvement points by business value and by how realistic they are to deliver.
- Theory of Constraints
- Little's Law
- Pareto
- Root cause (5 Whys)
- Cost-to-serve
Bottlenecks and rework
Steps where cases wait, get keyed in twice or come back for correction. Cycle time is set by the bottleneck, not the average.
Decision points
Recurring decisions that follow rules or patterns in the data, and fit a business rule, an ML model or a language model.
Value against effort
A priority map that ties every improvement to ROI, risk and time to deliver.
03
Re-engineering
Redesign the process around AI and ML.
We don't add a tool to an old process. We rebuild the process.
- BPR
- Lean (8 wastes)
- DMN decision tables
- RACI
- Human-in-the-loop
- Exception paths
What to drop, what to automate
Steps that disappear, steps that move to deterministic automation, and steps where AI adds real value.
Decision-node engineering
For every node: who decides (a rule, a model or a person), on what data, at what confidence threshold, and what happens below it.
Human in the loop
Exactly where human judgement stays, and the controls around it.
04
Build & deploy
Build and deploy inside the organization.
KO AI engineers build the new process in your environment.
- Orchestration
- API integration
- RAG
- Evals
- Guardrails
- ADKAR
Process, models and integration
An orchestrated workflow, AI and ML models, and connections to your core systems, under your security policies and permissions.
Adoption by the people
Training for the teams, new working routines, and support until the new process becomes habit.
05
Measure & improve
Measure, then keep improving.
The new process is measured against the baseline and improves over time.
- SPC control charts
- DMAIC
- Model drift
- LLMOps
- Unit economics
Metrics and monitoring
Comparison with the baseline, model performance and drift monitoring, and regular reporting to management.
Token economics
The right model for each task and control of usage costs, so AI spend doesn't eat the savings.
Decision-node engineering
Anatomy of an engineered AI process.
A good process is a chain of steps and decision nodes. At each node we decide who decides: a business rule, an ML model, a language model or a person, and at what confidence threshold. Select a node to see how it works.
Supplier invoice processing: a deterministic process with AI components at the nodes
- Deterministic automation
- Language model (LLM)
- Business rule (DMN)
- ML model
- Human decision
- Input
- Email, supplier portal or scan.
- Logic
- Intake, duplicate check by invoice number and supplier, and a new record.
- Control
- Every invoice gets an ID and is written to the log.
01 Invoice intake · Deterministic automation
- Input
- A PDF or image, in any supplier's format.
- Logic
- A language model extracts supplier, amounts, line items and VAT into a fixed structure (JSON).
- Control
- A confidence score for every field. Below the threshold, it goes to a person.
02 Data extraction · Language model (LLM)
- Input
- Purchase order, delivery note and invoice.
- Logic
- A decision table (DMN): quantities and prices match within tolerance.
- Control
- Fully explainable: every decision points to the rule that fired.
03 Three-way match · Business rule (DMN)
- Input
- Supplier history, prices and payment patterns.
- Logic
- An ML model scores risk: unusual price, changed bank account, duplicate invoice.
- Control
- A score threshold set against the cost of an error. Monthly drift monitoring.
04 Anomaly detection · ML model
- Input
- The match result plus the risk score.
- Logic
- Clean and low risk: to payment. Otherwise: to the approver for that amount (RACI).
- Control
- Routing thresholds are managed as parameters, not in code.
05 Threshold routing · Business rule (DMN)
- Input
- A summarised case with the reason it was flagged.
- Logic
- A person decides only on the exceptions, with all the information in front of them.
- Control
- The decision and its reason are stored and used to improve the model.
06 Exception approval · Human decision
- Input
- An approved invoice.
- Logic
- Posting to the ERP and release for payment on the supplier's terms.
- Control
- A full audit trail, end to end.
07 Posting and payment · Deterministic automation
An illustration, not a client result. In every project the nodes, thresholds and controls are set from the organization's own data and risks.
Which component fits which decision?
| Component | When it fits | Cost per decision | Explainability | Speed |
|---|---|---|---|---|
| Business rule (DMN) | Clear, stable rules: thresholds, matching, permissions | Close to zero | Full | Instant |
| ML model | Patterns in historical data: forecasting, risk scoring, anomalies | Low | Partial (feature importance) | Milliseconds |
| Language model (LLM) | Unstructured text and documents: reading, extraction, summarising, classification | Variable, by tokens | Needs source citation and controls | Seconds |
| Human decision | Judgement, exceptions, accountability and regulation | The highest | Full, when documented | Minutes to days |
Agents or workflows
Autonomous AI agents vs engineered AI workflows.
Autonomous agents attract a lot of productivity fantasies. In practice, most of the efficiency comes from placing AI components inside well-designed, deterministic processes.
| Criterion | Autonomous agent | Engineered AI workflow |
|---|---|---|
| Efficiency | A path that changes on every run, with extra steps and retries. | A fixed, short path. AI runs only at the node where it adds value. |
| Cost | Hard to predict: every step is another model call. | Predictable and measurable per case, including token cost. |
| Control | Hard to explain why a decision was made, or what happened on the way. | Every node has a threshold, a fallback and a log. You know what happened and why. |
| Scale | A small error is multiplied across thousands of runs. | The same quality at ten cases and at ten thousand. |
| Value | Impressive in a demo, hard to prove in production. | Measured against a business KPI from day one. |
We use autonomy where it earns its place: in an open, well-defined step, inside a controlled process, with limits, permissions and tests.
Anthropic: Building effective agents Gartner: over 40% of agentic AI projects predicted to be cancelled by end of 2027
AI process re-engineering
What changes in the process itself.
FDE doesn't bolt a chatbot onto an existing process. It rebuilds the process: AI components do the repetitive work, and people decide where judgement is needed.
Insurance claim intake, today: manual process
- AgentReceive email and documents12 minAn agent opens every email, downloads the files and identifies the request type.
- AgentOpen the case and key in data18 minManual entry of policyholder and incident details into the core system. Most data-entry errors start here.
- AdjusterCheck policy coverage25 minSearching policy terms, exclusions and deductibles against the incident documents.
- AgentRequest missing documents2–3 daysA follow-up email to the policyholder, then waiting. The case sits in the queue.
- AdjusterDecide and pay20 minSummary, decision, payment entry and documentation.
75 minHuman work time per case
3 daysFrom receipt to decision
Insurance claim intake, with AI agents
- AI agentAutomatic intake and classificationSecondsAn agent reads the email and files, identifies claim type and urgency, and routes it.
- AI agentExtract data and open the case1 minData is extracted from the documents and written to the core system via API, with a confidence score per field.
- AI agent + adjusterCoverage check with the clause cited4 minThe agent checks the policy terms and cites the clause. The adjuster approves or corrects (human in the loop).
- AI agentRequest missing documentsSame dayA precise, automatic request to the policyholder, with a reminder. The case doesn't stall in the queue.
- AdjusterDecide and pay8 minThe adjuster gets a ready, summarised case and decides. The decision stays with a person.
13 minHuman work time per case
1 dayFrom receipt to decision
An illustration, not a client result. In a real project we measure the actual times in the first week and calculate the savings on the organization's own numbers.
What improves
Which KPIs does AI process optimization improve?
Every FDE project is measured against business metrics agreed before the build, not against a count of features.
Cycle time
From receiving a request or case to closing it, including waiting time between teams.
Cost per case
Work hours and direct cost per unit of work, including the cost of model usage.
Quality and accuracy
Error rate, rework, and the exceptions that reach a control step.
Throughput and growth
The volume a team handles without adding headcount.
Response times (SLA)
Meeting service targets for internal and external customers.
Decision-making
Faster, data-based decisions, each documented and explainable.
Why now
Which companies have forward deployed engineers?
In 2026 the largest AI and cloud companies set up dedicated FDE units, because a good model turned out not to be enough. Someone has to put it into the real work of the organization.
OpenAI: The Deployment Company
OpenAI launched a dedicated company for embedding forward deployed engineers in enterprises, with an investment of about $4 billion and about 150 FDEs from day one.
AWS: a $1 billion FDE unit
AWS invested $1 billion in a unit that embeds engineering teams with customers, works on their data and infrastructure, and leaves them self-sufficient at the end.
When everyone offers FDE, the question is who sits across the table from you.
Why KO AI
Why our FDE services fit your company.
AWS, Anthropic, OpenAI and Microsoft all offer FDE today, around their own platforms. Our FDE is multi-system and vendor-agnostic: we bring the most effective solution for each process, with no tie to any one vendor.
Vendor-agnostic
We are an official partner of AWS and Microsoft and part of Anthropic's partner network. Partnership means deep knowledge and the ability to combine technologies, not a preference for any vendor.
Deep business process engineering
Real experience designing, measuring and optimizing processes. We start from the process, not the tool.
Operators, not only service providers
We have run units inside large organizations and made the decisions ourselves. We know how a process looks to the person accountable for it.
Technical depth in AI and ML
Hands-on experience selecting, adapting and evaluating AI and ML models, not only using tools.
International applied research
We work in international applied research in the field, and bring methods into projects that have been tested outside the organization too.
A boutique team in Tel Aviv, Prague and Porto
On-site with clients in Israel and Europe. We know the systems and the regulation you work under, from GDPR and the EU AI Act to DORA in financial services.
ANTHROPIC
Comparison
FDE vs vendor FDE vs traditional consulting vs a software house.
Four ways to bring AI into an organization. The difference is where you start, who owns the result, and when you see it.
| Criterion | FDE (KO AI) | Cloud / model vendor FDE | Traditional consulting | Software house |
|---|---|---|---|---|
| What you get | A process that runs in production and measures a result | A production solution, on the vendor's platform | A recommendations document and a roadmap | Software built to a specification |
| Vendor preference | Agnostic: the best-fitting solution, from any vendor | Built on the vendor's own technology | Varies | The house stack |
| Where the team works | Inside the organization, with the people who do the work | Inside the organization, around the platform | In workshops and interviews | At the supplier, through a contact person |
| Starting point | The business process and its metrics | The platform's capabilities | Strategy and organization design | A specification the organization wrote |
| Who owns the result | The same team, from mapping to production | The vendor's team, with a focus on platform adoption | The organization, after the engagement | The organization, after hand-over |
| How success is measured | A business KPI set on day one | A business KPI plus platform adoption | Satisfaction with the deliverable | Meeting the spec and the schedule |
| Pace | A process in a week, a decision system in a month, a unit in 3 months | Depends on the engagement | Months before anything is built | Depends on the spec, usually long |
Pace
A week, a month, three months: what you can build.
One engineered process in a week. From there the pace compounds: a decision-making system in a month, and a whole unit in three months.
1week
In a week: one process replaced
A day for diagnosis, about three days for design, a few days to build, then experiments on real data until the target value is reached.
- One old process out, one engineered process in.
- One numeric target, measured from day one.
1month
In a month: a decision-making system
Several processes joined into one system: decision nodes with ML and language models, confidence thresholds, exceptions routed to the right people, and full control.
- Decision nodes with thresholds and controls.
- Exceptions reach the right person, with all the information.
3months
In three months: a whole unit
Dozens of new processes replacing many dozens of old ones, with monitoring, model-cost management and an internal team running the unit.
- Dozens of new processes in production.
- Monitoring, cost control and a team that runs it.
The process sprint: what one week looks like
Day 1
Diagnosis
Map the process on the floor and in the data, measure the current state and set one numeric target.
Days 2–4
Design
The optimal process, its decision nodes, the right component for each node, and the controls.
A few days
Build
The workflow, AI and ML components and system connections, in the organization's environment.
Until the target
Experiments
Runs on real data, measured against the target and tuned until the value is reached.
A complex process? We split it into sub-processes, and each one goes through the same sprint.
Who it fits
Which organizations and processes is FDE right for?
Mid-size and large companies with a high-volume process, repetitive manual work, and data that already sits in their systems, even if it isn't tidy.
Operations and back office
Intake, checking and processing of requests, documents and cases that repeat hundreds of times a day.
Customer service and contact centres
Handling repeat requests, summarising calls and routing to the right agent.
Finance
Reconciliations, quotes, controls and reports that come back every month.
Legal, regulation and compliance
Contract review, regulatory monitoring (EU AI Act, DORA, GDPR) and documentation, with the source cited.
Data and decision-making
ML and forecasting models wired into the process instead of staying on a dashboard.
HR and internal knowledge
Recruiting, onboarding and internal knowledge assistants that answer from your own procedures.
Proof
What already runs in production.
KO AI figures as of September 2026. Industries: Financial Services · Insurance · Software & SaaS · Mobility & Automotive · Healthcare & Life Sciences · Manufacturing & Industrial.
40+
AI / ML systems in production
500+
New AI-based processes deployed at clients
100+
Projects delivered
10K+
Hours of AI consulting for companies
Questions, answered
Forward deployed engineering FAQ.
What is forward deployed engineering (FDE)?
Forward deployed engineering (FDE) is advanced process engineering that turns legacy business processes into AI-based processes and deploys them inside the organization. A KO AI FDE team of engineers and consultants works inside your organization, on your data, systems and processes, and owns both the business understanding and the build, until the new process runs and produces measurable value.
What does a forward deployed engineer do?
A forward deployed engineer sits with the people who run a process, maps and measures it, redesigns it around AI and ML, builds the new workflow in the client's environment and stays with it in production. The role combines process engineering, operations management, project management and AI/ML engineering.
What is FDE in consulting, and how is it different from AI consulting?
Traditional consulting usually ends with a document or a deck, and delivery moves to another team. In FDE the same people who mapped the process also build the system, connect it to your existing systems and support it in production. Success is a business KPI agreed in advance, not the number of slides. See also our AI consulting service.
How is FDE different from a software house?
A software house expects a finished specification and builds to it. An FDE team starts before the specification: it sits with the people who do the work, finds where AI really saves time or money, and adjusts the solution as it learns what works on the floor. At the end the knowledge stays with you, with documentation and an internal team that can maintain it.
How long does an FDE project take?
One process is usually replaced within a week: a day for diagnosis and measuring the current state, about three days to design the process and its decision nodes, a few days to build, then experiments on real data until the target value is reached. Complex processes are split into sub-processes, each going through the same sprint. In a month you can build a full decision-making system, and in three months a unit with dozens of new processes replacing many dozens of old ones.
Which processes are right for FDE?
Processes with high volume, fairly clear rules and repetitive manual work: document and claims intake, customer service requests, financial reconciliations, quotes, classification and documentation, and recurring data analysis. Our rule: if you can measure the work time today, you can set a target and prove the saving.
What is the difference between an autonomous AI agent and an AI workflow?
An autonomous agent decides its own path on every run: which steps to take and in what order. An AI workflow is a designed process in which AI components run at defined nodes, with confidence thresholds, fallback paths and logging. For most business processes an engineered workflow is more efficient, cheaper and more predictable, easier to control and scales without losing quality. We use autonomy only in open, well-defined steps inside a controlled process.
What is decision-node engineering?
Every business process contains points where a decision is made: approve, reject, route or ask for more information. In decision-node engineering we set, for each node, who decides (a business rule, an ML model, a language model or a person), on what data, at what confidence threshold, and what happens when the threshold isn't met. The rule: the simplest, cheapest component that meets the target, and AI only where it really adds value.
Do you recommend a specific AI vendor?
Not in advance. Our FDE is multi-system and vendor-agnostic: we choose the most effective solution for each process by cost, performance, data security and the systems you already run. We are an official partner of AWS and Microsoft and part of Anthropic's partner network; the partnership means deep knowledge and the ability to combine technologies, not a commercial preference.
Do your engineers work on our data, and how is data security handled?
Yes, that is the core of the model. We work inside your environment, under your security policies and permissions, with the option of models that run on-premises or in a private cloud (VPC) when the data is sensitive. For European clients we design for GDPR and data residency from the start, and we plan the EU AI Act obligations that apply to the process. An NDA, permission mapping and security checks are part of the kick-off, not an add-on.
Do you work on-site with clients in Europe?
Yes. KO AI has offices in Tel Aviv, Prague and Porto, and our teams work on-site with clients in Israel and across Europe. Discovery and the first sprint usually happen at your office, next to the people who run the process.
What happens once the process works?
The goal is for you to be self-sufficient. Every process is handed over with monitoring, controls, documentation and training for the team that runs it. You can continue with us to the next process, keep a small ongoing support arrangement, or take maintenance fully in-house.
How do we start an FDE project?
In a 30-minute call we pick one process that occupies a lot of people or costs a lot of money. Then you get a proposal with a measurable target, a timeline and the team size. You don't need tidy data or a ready specification: we do that on day one.
Keep exploring
Explore our other services.
01
AI Consulting
Where AI creates value, the roadmap and governance, from the boardroom to the model layer.
02
AI Development
Enterprise AI agents, custom AI/ML and the infrastructure to run them in production.
04
AI Training
Training as part of adoption, by role and department, so new processes become habit.
→
AI Automation
Automating the processes you already have, and knowing when a process needs FDE instead.
Pick one process. We'll come back with a target, a timeline and a scope.
A 30-minute call with an engineer who will work on the project, not a salesperson.