When To Build - Sana Agent Builder & Custom Patterns
Sometimes the right answer is a custom agent. This module teaches how to tell, how to scope, and how Kainos delivers it.
Welcome
Most AI conversations with customers end with use or adapt. A small, important subset ends with build. This module is about that subset - when a custom agent is genuinely the right answer, and how Kainos scopes, delivers and sustains one.
"Use, Adapt, or Build" is the Kainos AI Navigator. Build is the smallest leg of the three on purpose. By the end of this module you should be able to look a customer in the eye and tell them whether build is the honest recommendation - and if it is, what that actually commits them to.
Learning objectives
Build criteria - the filter
A custom agent is justified only when all four of these hold. If any one is shaky, the honest answer is not-yet, or not-at-all. This filter is non-negotiable - use it before you agree to a scoping call, not after.
(a) No GA or near-term agent fits
Map the need against the current Workday roadmap. If a GA agent covers the job - or one is landing in the next two releases - use it and adapt around the gap. Custom build is only for genuine whitespace.
(b) Value justifies lifecycle cost
Not just build cost - evaluate, design, deploy, monitor, maintain, evolve, and the credit consumption the agent drives in production. If the value case only works against a one-off build number, it doesn't work.
(c) Data and process are ready
The underlying Workday objects are clean, the process is well-defined, and the organisation already runs it consistently. Custom agents amplify the process underneath them. They do not fix it.
(d) The customer has an executive sponsor
A named, accountable exec who owns the value case and will defend the agent through its lifecycle. Without a sponsor, a custom agent becomes a project that nobody updates and nobody decommissions. That is how trust erodes.
All four criteria, or we walk. We would rather lose a build engagement than sign up to a custom agent that will quietly degrade for eighteen months. The CoE AI will back any architect who applies this filter honestly.
Worked example - applying the filter
A walk-through of the four-criteria filter against a realistic scenario. Illustrative only - numbers and specifics will differ per account, but the shape of the reasoning is the one we want architects using in the room.
A FTSE-250 CFO asks Kainos to build a custom "close agent" that drafts variance commentary, chases late journals from controllers, and auto-generates the month-end close pack. They want an MVP inside the quarter.
(a) No GA or near-term agent fits. Partial fail. Financial Test Suite is GA today and covers anomaly checks; Financial Audit lands H1 2026 and overlaps with variance commentary. Roughly 60-70% of the stated need is inside the roadmap. Custom build would rebuild ground Workday is already shipping.
(b) Value justifies lifecycle cost. Fails under scrutiny. Finance quotes two FTE of controller time saved - credible. But lifecycle cost (monitor, maintain, evolve, release alignment) plus credit consumption at close-cycle volume erodes the case to break-even by year two. The business case only works against the build number in isolation.
(c) Data and process are ready. Fails. Discovery reveals three ledger instances post-acquisition, inconsistent journal-approval thresholds across regions, and a close calendar that slips most months. A custom agent amplifies that inconsistency at speed - exactly the failure mode in the "When NOT to build" section.
(d) Executive sponsor. Passes. The CFO is named, accountable, and credible. Not enough on its own.
Verdict: no-go - for now. Three of four criteria fail. The honest recommendation is: adopt Financial Test Suite today, prepare for Financial Audit, run a ledger-consolidation and close-calendar remediation in parallel, and revisit build in twelve months against a narrower job-to-be-done. The CFO keeps the sponsor; Kainos keeps the trust.
Sana Agent Builder
Sana Agent Builder is Workday's platform for customer-built agents - the tooling that lets us construct, test and deploy bespoke agents against a customer's tenant data, inside Workday's trust and security model. It is the "build" leg of the Kainos AI Navigator, not a separate product motion.
What it is
A configurable agent-authoring environment: the scaffolding to define an agent's capability, wire it to tenant data with the right permissions, attach guardrails, and deploy it where the relevant users work.
How it's different from a GA agent
A GA agent is Workday-built, Workday-maintained, Workday-evolved. A Sana-built agent is customer-built, customer-owned, and - when delivered by Kainos - Kainos-maintained on the customer's behalf. Same trust model, same tenant security, but the roadmap belongs to the customer.
Where it fits in the tenant
Inside the same Workday tenant, using the same data model, the same business process framework, and inheriting the same tenant security that every GA agent inherits. A custom agent is not a bolt-on. If you find yourself describing it as one, you are probably describing the wrong solution.
Sana Agent Builder is how Kainos delivers custom capabilities for customers. When we are in front of a customer, the motion is "Use, Adapt, or Build" - Sana is the how for build, not the headline.
Reference patterns for custom agents
Four recurring patterns from Kainos engagements. Most credible custom-agent conversations land in one of these shapes. If a customer's idea doesn't resemble any of them, slow down and revisit the filter.
These are starting shapes, not finished products. Every one lands differently per customer - the point is to recognise the pattern early, then apply the build filter to it before scoping.
Lifecycle: build is the small part
The most common mistake customers make is treating a custom agent as a project. It isn't. It is a capability with a lifecycle. The build phase is a fraction of the effort, and it is not where value lives or dies.
Evaluate
Apply the four-criteria filter. Qualify value, confirm sponsor. Decide whether to proceed honestly.
Design
Define capability, data access, guardrails, human-in-the-loop, and the measurement frame up front.
Build
Author the agent in Sana. This is the smallest, shortest phase - a few weeks on a well-scoped MVP.
Deploy
Roll out with the relevant users, in the right parts of the tenant, with adoption runway.
Monitor
Instrument from day one - accuracy, usage, credit consumption, business outcome. No monitoring, no managed agent.
Maintain & evolve
Ongoing. Workday releases move, tenant data shifts, customer priorities change. A custom agent needs stewardship for its whole life.
"Build is a few weeks. Maintenance is forever." If a customer cannot commit to the maintain-and-evolve phase - their own capacity or a managed service from us - we should not build the agent. We have declined engagements on exactly this point. Do it again if you have to.
Kainos delivery model for custom agents
We deliver custom agents in three sequential, de-risked steps. Each step has a clear decision point - either party can stop. That discipline is what makes the commercial model defensible.
Discovery workshop
Fixed-fee, short engagement. Applies the four-criteria filter, confirms sponsor, shapes the MVP, and produces a honest go / no-go. We walk away gladly if the criteria don't hold.
Fixed-price MVP
Scoped, time-boxed build of a single agent against a single job to be done. Fixed price, fixed outcomes, monitored from day one. MVP lands live - not a prototype.
Managed evolution
Ongoing service: monitor, maintain, evolve the agent against the customer's shifting context and the Workday release cadence. This is where most of the value - and most of the work - lives.
The commercial shape matters. Discovery protects the customer from a bad build; the fixed-price MVP protects the budget; managed evolution protects the value. All three, or none.
Cost & ROI framework
Custom-agent economics have three cost components and one value component. Customers who only look at the first cost lose faith in the AI programme at month nine. Surface all three up front.
Build cost
One-off - discovery plus the fixed-price MVP. The visible number. Typically the smallest of the three over the life of the agent.
Lifecycle cost
Recurring - monitoring, maintenance, evolution, tenant-release alignment, change management. Budget for this explicitly or it will surprise the sponsor.
Credit consumption
Every agent invocation consumes flex credits at runtime. A successful agent drives volume - that is a good problem, but it is a real line in the cost model. See Module 5 for the architecture and Module 7 for forecasting.
Value - measured, not claimed
Time saved, error rate, decision quality, or revenue / cost moved. Agreed with the sponsor at design time, instrumented at deploy, reported in managed evolution. No measurement, no ROI conversation.
Every custom-agent proposal from Kainos must show all three costs and the measured value. If a proposal only shows build cost, send it back. We do not sell agents - we sell outcomes, honestly priced.
When NOT to build
More custom-agent conversations should end here than anywhere else. Walking away is a credibility-building act - customers remember the architect who said no for the right reason.
70%+ GA coverage
If a GA or near-term agent covers most of the need, use it and adapt around the gap. The 30% you would gain from a custom build rarely justifies the lifecycle cost.
Data isn't clean
If the underlying Workday data or process is inconsistent, a custom agent will amplify the inconsistency at speed. Fix the foundation first.
No executive sponsor
No sponsor means no-one owns the value case, no-one defends the investment at year-two, and no-one signs off the evolution. The agent decays silently. Don't start.
"AI for AI's sake"
Board-level pressure to "do something in AI" is not a job to be done. If the customer cannot describe the specific outcome, go back to discovery - or back to adapt.
Walking away from an ill-founded build is not lost revenue - it is protected trust. The customer who hears an honest "not yet" from us is the customer who buys the right build next year.
Check your understanding
Three questions. Each explains why every answer is right or wrong - the reasoning matters more than the score.
Next steps
- Architects: Module 5 - Flex credits & AI architecture. Essential for understanding the runtime economics of any custom agent you scope.
- Commercial architects & account leads: Module 7 - Predicting credit consumption. How to model the third cost line honestly before signing.
- All learners: Module 12 - Case studies. Real Kainos-delivered custom agents - what worked, what didn't, and why.
- CoE AI core: book a working session with the CoE AI Lead to walk a live build-filter conversation from a current account.