From AI Pilots to Operating Reality

Client: Xceptor‍

Industry: Financial Technology / Data Automation
Client presenter: Michael Kinloch, SVP Engineering, Xceptor 
Forte Group lead: Alex Lukashevich, Chief AI Officer
Presented publicly at: CTO Craft Conference, 10 March 2026
Watch the full on-stage recording

‍

‍

In Brief

‍Most organisations talk about AI in software delivery. Xceptor is doing it across every phase of their product lifecycle, in production, with numbers to back it up. Working with Forte Group, they moved from ad-hoc AI tool adoption to a fully embedded, three-stage AI delivery model in under twelve months. This is not a pilot. It is not a proof of concept. It is how Xceptor builds software today.

‍

‍

The Challenge

Xceptor builds a data automation platform that financial institutions use for pre- and post-trade workflows, reconciliations, tax processing, and regulatory reporting. The platform connects upstream and downstream data sources, extracts files of any format or structure, and automates manual processing journeys across the full trade lifecycle. It provides a live web interface for business users to manage exceptions, with purpose-built products spanning reconciliations, post-trade operations, tax processing, and client lifecycle management. The platform runs across 170+ SaaS instances serving enterprise clients including Fortune 500 companies. The engineering team is 39 people.

The team had rolled out GitHub Copilot and was experimenting with Claude. Individual developers were writing code faster, but team-level delivery metrics had not moved. Tool usage was scattered, unmeasured, and disconnected from the development process. Engineers used AI for code completion. Nobody was using it for requirements, architecture, test strategy, or release documentation.

Three problems stood out in the initial assessment:

• Sequential handoffs were the bottleneck: features moved through product, architecture, development, QA, and DevOps as a relay, and wait time between stages often exceeded the work time within them.
• The team put their rework rate at 30%, with requirements gaps surfacing during build or QA rather than during planning.
• 18,064 AI events in six months had no connection to outcomes: the team could not answer whether AI was making them better or just busier.

The root cause was a “tool-first” approach — implementing tools into processes not yet prepared for them. To unlock real gains, Xceptor needed to understand their processes first and then apply AI to workflows ready to receive them.

‍

The Approach

The programme moved through three stages:

• Augment: AI as a tool for individual practitioners.
• Automate: AI embedded into the process itself.
• Agent: AI agents executing pipeline stages with human approval at every gate.

The first two stages are complete. The agentic stage is in production, deployed as a governed plugin (v1.13, 46 installs org-wide) with slash commands for every PDLC stage.

The pipeline covers nine stages: feature requirements, architecture and design, test strategy, story breakdown, implementation planning, build, test, PR review, and release. Two roles govern the pipeline. Product works with AI to define and shape requirements, ensuring the right problem is solved before build begins. Builder orchestrates AI through design, implementation, testing, and release, accountable for quality of everything that ships. AI proposes at every stage. Humans approve the gates and the outcome. No artefact moves forward without sign-off.

‍

How Roles Transform

The AI-native PDLC changes what people spend their time on. Roles do not disappear: they elevate:

• Product Owners now steer agent-built requirements rather than writing them from scratch.
• Developers focus on architecture and reviewing PRs drafted by AI.
• QA and Ops approve AI-generated test suites and insights, focusing on high-level risk and strategy.

‍

“We were trying to get AI to build a process on our platform. When we gave it too much context, it did a terrible job. When we gave it just the bare minimum — a codebase link, the documentation, one good working example — within a day it was building an end-to-end process like our business user would.”

— Michael Kinloch, SVP Engineering, Xceptor

‍

Experiment 1: Connector Build (MS Graph and Databricks)

The first proof point was connectors, the most isolated unit of work in Xceptor's codebase: self-contained, clear API boundaries, independently deployable. Two connectors (MS Graph and Databricks) that the team estimated at two weeks through the traditional process were delivered in two days : an 83% reduction in time and a 4.85× efficiency gain.
‍

Experiment 2: Config Builder (Loan Notices)

The second experiment extended the pipeline beyond code into platform configuration. Xceptor built an AI Configuration Builder that translates plain-language business process descriptions into structured platform configuration.

The first use case, Loan Notices, was delivered in 6 days against a 26-day estimate, a 4.3× improvement in delivery speed compared to traditional methods.

The Loan Notices solution was demonstrated to a customer at PoC stage. Customer feedback on quality and user experience: excellent. This was the first time Xceptor demonstrated an AI-configured, end-to-end solution to a live customer. The pipeline now works for both code and platform configuration from the same process.

‍

AI Agent Products

Alongside the internal PDLC work, Forte built agent products running on the Xceptor platform:

• SSI Extraction Agent (launched October 2025): extracts structured settlement data from unstructured emails and attachments (PDFs, Excel, CSVs, images), with source attribution and audit trails for regulatory compliance. Production accuracy target is 95%+; initial test cases hit 100%.
• Operations Agent: monitors all 170+ SaaS instances, detecting exceptions, diagnosing root causes, and routing issues. Exception triage that previously took days now completes in hours. No P1 incidents have been missed during the monitoring period.

The agent framework is architected as Xceptor IP: custom orchestration, prompt governance, multi-tenant deployment with usage metering, and security controls built for financial services. It now serves as the foundation for future agents, cutting delivery time from months to weeks.

‍

Results
‍

‍

Additional delivery improvements:

• Design prototyping: AI now generates all UX prototypes, enabling more variations and better solution validation before engineering begins. This returned 32 days per designer, per year.
• Engineer adoption: 85% of engineers (39 of 49 across the broader team) adopted the governed plugin, exceeding the 50–60% target, driven by a champions programme and role-specific playbooks.
• Test scripting: by switching to Claude Opus and the AI-native Playwright MCP framework, one QA engineer now produces the equivalent output of four.

‍

What Actually Changed

The implementation plan became the primary review artefact. Before any code is written, the agent produces a detailed plan: every file it will create or modify, what changes it will make, what tests it will write. The team found a clear pattern: when the plan is correct, execution is consistently accurate. Course-correcting a plan takes minutes. Course-correcting generated code takes hours. Developers stopped correcting errors and started confirming decisions.

Somewhere around the third connector, the cognitive shift happened. The builder was no longer reading every line of generated code looking for errors to fix. They were reading to confirm the approach, occasionally adjusting, mostly approving. The cognitive load moved from production to verification.

The AI asks the same questions, runs the same security checks, applies the same structure every time. No skipped steps because the standup ran over. That consistency across every artefact is what reduced rework from 30% to under 10%.

Documentation debt got paid as a side effect. When the team built skills that reference existing docs (coding standards, architecture patterns, functional specs), they discovered gaps, contradictions, and outdated content. Fixing documentation for the agents also fixed it for human onboarding and cross-team collaboration.

‍

What Was Hard

Three challenges are worth calling out for teams embarking on a similar journey:

• User story granularity was wrong on the first pass. The agent initially broke features into stories the way a human developer would: small, independent, individually testable. For an AI agent, each story requires reloading the full context, and a small story costs nearly as much as a large one. A single larger story per connector worked better. The right approach for humans was the wrong approach for agents — it took a failed iteration to discover that.
• Artefact detail is a cost lever. The agent generates more detail than humans need to review. Early design documents created review overhead that slowed the pipeline. The team iterated on prompt design to calibrate output. Expect two to three iterations per skill before output consistently matches how the team actually works.
• Token consumption was higher than planned. The agent sometimes loaded entire documents when it only needed specific sections. The team is now evaluating whether non-coding stages should run on smaller, cheaper models, reserving the expensive model for the engineering stage where code quality matters most.

‍

What’s Next

Three streams are scaling in parallel. The Connectors AI PDLC is rolling out to the full delivery team. Features AI PDLC and Config Builder AI PDLC are each on a 6-week sprint to MVP. The Config Builder then phases across Solutions, Pre-Sales, and Delivery teams over the following months. Each cohort validates before the next begins.

The Code PDLC and Config PDLC are merging into a single pipeline: one lifecycle, two output types. If AI can build platform configuration in 6 days instead of 26, that capability can extend to customers — compressing project delivery timelines and opening self-service options that change the commercial model for professional services.

‍

Where This Led

Michael Kinloch, Xceptor’s SVP Engineering, co-presented this work with Forte at CTO Craft London (March 2026) and CTO Craft Toronto (May 2026), with updated results at each conference.

The engagement expanded from a single agent build into a board-approved, firm-wide AI programme. Through Xceptor, Forte was introduced to Astorg, the PE firm behind the investment, which led to a broader relationship covering AI adoption across Astorg portfolio companies.

‍

Technology Stack

Anthropic Claude (Opus for engineering stages, Sonnet for non-coding stages), GitHub Copilot, Claude Code with custom skill architecture (v1.13, 46 org-wide installs), Playwright MCP (test automation), Azure DevOps, Azure OpenAI (agent products), Azure Document Intelligence (OCR and document extraction), Kubernetes on Azure (agent deployment), custom orchestration engine, SOC 2 and ISO 27001 compliance frameworks.

‍

Download the full case study to see how AI-native delivery drives 100× velocity and real production impact.