Platform & Engineering

How we build Accrual.

An 8-person platform team and a 22-person automation COE support a 14 million customer bank. We keep it running by betting on orchestration, resilient automation, and case-based workflows where linear diagrams break down.

The estate

Automation at Accrual, by the numbers.

We have been serious about RPA since 2016 and serious about orchestration since 2024. Here is what that looks like today.

420
Bots in production
+47 YoY
9
Orchestration processes live
+6 YoY
8
Platform engineers
steady team
22
Automation COE members
+4 YoY
14M
Customers the platform serves
+8.2% YoY
2.8M
Bot transactions per day
+31% YoY
99.94%
Platform uptime (P1 processes)
vs 99.91% target
14.2M
Labour hours returned since 2016
cumulative
How the platform fits together

Three layers. One runtime.

Our platform team rebuilt the bank's automation stack in 2023 around a single orchestration runtime. Different teams use different front doors into the same product.

Layer 1 · Orchestration

Process and flow orchestration

The backbone. Business analysts model in a BPMN-style designer. Developers author the same runtime in a flow-based, code-near experience. Same engine. Same deployment pipeline. Same observability.

How we chose two front doors
Layer 2 · Case management

Structured stages, adaptive tasks

For the workflows that cannot be drawn as diagrams. Commercial lending, claims, internal investigations. Stages hold structure, tasks adapt on the fly.

Commercial lending case study
Layer 3 · Bots and agents

Resilient RPA and code-generating agents

Our 420 bots sit in the execution layer. Integrations prefer APIs and fall back to UI automation when upstream systems drift. Code-generating agents help our COE author and maintain bots at scale.

Why we fall back to UI
How we work

Ten principles on the platform wall.

Printed, laminated, signed by everyone. In every Accrual platform team room. We update them about once every two years when someone has a better idea.

1. Same product, many front doors

One orchestration runtime. Different authoring experiences for different audiences. No second product for developers. No second product for analysts.

2. Diagrams where diagrams work

Linear processes get BPMN. Everything else gets case management. Forcing chaos into a diagram is how you get a 47-step diagram that nobody maintains.

3. APIs first. UI automation always.

APIs are faster when they exist and stable. UI automation is the safety net when vendors deprecate without warning. Every integration ships with both paths.

4. The coding agent is a colleague

Our code-generating agent for bot authoring is a pair programmer, not a magic button. Our engineers review every change. Our agent ships faster than either alone.

5. Tests exist because bots exist

Every bot ships with contract tests for its APIs and regression tests for its UI targets. The 420th test is as important as the 1st. We test Delegate-class changes end-to-end.

6. One deploy pipeline

BPMN processes, flows, and bots all ship through the same CI. Same gates. Same rollback strategy. Same observability surface.

7. Stages over steps

When work is inherently non-linear, give it stages and tasks. Commercial lending is a case. Claims is a case. Internal investigations are a case.

8. No hero deploys

If only one person can ship a change safely, that change should not ship. Our 8 platform engineers are interchangeable, on purpose.

9. Platform serves the bank

Not the other way around. When a business team needs something that is not on the roadmap and it is reasonable, we move the roadmap.

10. Boring is a feature

Reliability is the brief. Prestige is a nice-to-have. Our incidents are rare and our pages are boring. Both are on purpose. The moment a platform starts optimizing for interesting, it stops being trustworthy.

Our journey

From 5 bots to 420.

It did not happen overnight and we did not plan it this way. We just kept saying yes to the next thing and yes to writing down why.

2016
Automation COE established. First 5 bots go live in Finance. Payback measured in weeks.
2018
COE grows to 14. 120 bots in production. First bot-on-bot dependency causes us to start writing tests.
2020
Platform team formed out of the COE. 8 people. Brief to modernize the runtime underneath all the bots.
2022
First BPMN orchestration process goes live. KYC refresh for retail. Cycle time drops 60 percent.
2024
9 BPMN processes live across Retail, Capital, and Services. First case-based workflow for commercial lending.
2025
Introduced flow-based designer for developers alongside BPMN for analysts. Same runtime. Board asks about AI.
2026
420 bots. Code-generating agent rolled out for bot authoring. Commercial lending cycle time down 73 percent.
Latest from the blog

How we think. In public.

Posts from our platform engineers and COE leads. No sponsored content. No thought leadership. Just working notes from people building this stuff.

All posts

We are hiring platform engineers.

If you like working on small teams that ship things used by millions of people the next morning, get in touch.

Open roles