AI PMO CoE, RPA/GenAI in IT Shared Services

Automation and AI are the modern levers of shared services value, but they do not adopt themselves. A Centre of Excellence (CoE) — often structured within an AI-enabled PMO — is the organisational capability that governs automation and AI adoption: setting standards, running the opportunity pipeline, building reusable components, and ensuring the automation delivers governed, measurable value rather than a scatter of pilots.

The CoE matters because automation without governance fragments. Every team that builds its own bots and prompts produces automation that is inconsistent, unmonitored, and unmaintainable. The CoE is what turns automation from a team-level experiment into an enterprise-scale capability — much as platform engineering turns cloud-native from team choice into enterprise capability.

This page covers the AI/automation CoE, RPA in shared services, GenAI use cases, the automation pipeline, governance of automated decisions, and the PMO CoE's role in delivery discipline.

The CoE is the central team that owns automation and AI adoption: the standards, the platforms, the reusable components, the pipeline of opportunities, and the governance. It does not build every automation itself; it enables teams to build within a governed frame, and it builds the high-value, cross-cutting automations that no single team would build.

The CoE's components are consistent: a platform (RPA tool, GenAI service, orchestration), a standards library (approved patterns, reusable components), an opportunity pipeline (intake, assessment, prioritisation), a build capability (the CoE team and enabled teams), and a governance function (monitoring, control, review). Each is necessary; the CoE that has a platform but no pipeline builds what is interesting, not what is valuable.

The CoE's success measure is scaled value, not activity. 'Bots built' or 'prompts written' are vanity metrics; 'FTE-hours automated', 'cycle time reduced', 'error rate reduced' are value metrics. A CoE that reports activity without value is not governing adoption; it is enabling it to drift. The discipline is to tie every automation to a measurable outcome from intake to verification.

RPA automates rules-driven, repetitive tasks — the data entry, the system-to-system handoffs, the copy-paste between applications that dominate L1 and back-office work. In shared services, RPA's sweet spot is high-volume, standardised, rules-based work where API integration is infeasible or too slow to deliver. The value is real where these conditions hold; it is not where they do not.

RPA's limits are well known. Bots are brittle on UI change (a screen redesign breaks them), they create a shadow integration layer (the bot is an undocumented integration), and they cannot handle judgment (rules only). These limits do not make RPA wrong; they make it a tool for a specific job. The CoE's role is to apply RPA where it fits and to resist it where a proper integration or a human is the better answer.

The standardise-then-automate principle applies directly: RPA on a standardised process is reliable and valuable; RPA on a variable process is brittle and disappointing. The CoE gates RPA opportunities on standardisation — a process not standardised is not ready for RPA, and the CoE's first job is often to help standardise it.

GenAI extends automation into cognitive work — drafting, summarising, classifying, extracting — that RPA cannot touch. In shared services, the high-value GenAI use cases cluster around knowledge work: L1 support assistance (drafting responses, summarising tickets), document processing (extracting structured data from unstructured documents), knowledge management (generating and retrieving knowledge articles), and content drafting (communications, reports).

The value is significant because these are exactly the tasks that dominate shared services cost and that resisted automation before GenAI. A shared services organisation that automates L1 drafting, document extraction, and knowledge retrieval removes a class of work that RPA could not. This is reshaping the shared services business case — knowledge work is now automatable where only rules work was before.

The risks are equally real. GenAI can be wrong (hallucination), can leak data (if not properly bounded), and can produce inconsistent outputs. The governance — human-in-the-loop for consequential outputs, retrieval-bounding to prevent hallucination, data-class-aware use — is non-negotiable. An ungoverned GenAI deployment is a liability; a governed one is a capability. The CoE's job is to make it the latter.

The pipeline is how the CoE turns opportunities into value. It has stages: intake (opportunities submitted by the business or identified by the CoE), assessment (is the process standardised, high-volume, and valuable enough to automate?), prioritisation (which opportunities deliver the most value?), build (the CoE or enabled team builds the automation), deploy (into production with monitoring), and measure (did the automation deliver the expected value?).

The assessment stage is where most pipelines fail. An opportunity that is not standardised, not high-volume, or not valuable gets built anyway — because someone champions it — and delivers little. The CoE's discipline is to apply the assessment criteria consistently and to defer or reject opportunities that do not meet them. A pipeline that builds everything builds a lot of low-value automation.

The measure stage closes the loop. Each automation is tied to a value hypothesis at intake ('this RPA will save X FTE-hours and reduce error by Y'), and the hypothesis is verified after deployment. Automations that deliver are scaled; those that do not are diagnosed and either fixed or retired. The pipeline that does not measure produces automation whose value is asserted, never proven.

Automated decisions — whether an RPA bot approving an access request or a GenAI agent drafting a customer response — need governance. The principles are consistent with zero-trust AI: never trust an automated output unconditionally, verify before consequential action, and least-privilege the automation's authority. An automation with unrestricted authority is an incident waiting to happen.

Human-in-the-loop is the primary control for consequential decisions. A bot that processes 1000 standard access requests autonomously and escalates the 10 non-standard ones to a human is governed; one that processes all 1010 autonomously is not. The design choice is where the human sits — and the answer is at the consequential decision boundary, not as an afterthought.

Monitoring is the second control. Every automation produces logs of what it did; the CoE monitors these for anomalies (the bot that started failing silently, the GenAI that started producing off-brand responses). Automation that is not monitored is assumed to work until it is not, and the gap between failure and discovery is the damage. The CoE's monitoring is the early-warning system.

Structuring the automation CoE within an AI-enabled PMO brings delivery discipline to automation adoption. The PMO disciplines — portfolio prioritisation, value tracking, stage gates, dependency management — are exactly what automation adoption needs and what ad-hoc adoption lacks. The PMO CoE applies these disciplines to the automation pipeline.

The PMO CoE also brings the cross-functional view. Automation opportunities often span functions (a process that crosses finance and HR), and a CoE embedded in one function misses them. A PMO CoE, cross-functional by design, sees and prioritises the cross-cutting opportunities that deliver the most value. This is why the AI PMO CoE is displacing the single-function automation team.

The discipline the PMO CoE most enforces is value realisation. Automation, like any change investment, must be tied to outcomes and tracked to realisation. The PMO CoE that holds automation to the same value discipline as other investments produces automation that delivers; one that lets automation claim value by activity produces a portfolio of unverified bots. The PMO disciplines are what keep automation honest.