How to Redesign Shared Services Processes Before Deploying AI Agents

How to Redesign Shared Services Processes Before Deploying AI Agents

Every shared services process is an archive of decisions nobody remembers making. The invoice approval that needs three signatures exists because of a fraud scare a decade ago, the onboarding checklist carries steps for a payroll system retired in 2019, and a weekly report goes out to a distribution list on which no one is still reading it. Now the platform vendors are shipping AI agents into this archive, ready to execute all of it faster.

Advice to "redesign before you automate" is everywhere, but the method is nowhere. This guide is the method: how to baseline a shared services operation with process mining, then run the five-stage sequence, eliminate, standardize, improve, automate, agentify, with owners and honest timelines for each stage.

---

Key takeaways

• The pressure is structural. The Hackett Group projects GBS workload growing 15% in 2026 against 7% budget growth, and closing that gap with AI agents only works on processes designed for them.
• A decade of RPA left most operations half-manual. SSON research puts 56% of shared services organizations at 25 to 50% automation, with nearly 30% at low levels, and agents deployed on that foundation multiply exceptions.
• The sequence is Eliminate → Standardize → Improve → Automate → Agentify, in that order. Work removed never needs standardizing, and work standardized automates at a fraction of the cost.
• Exception rate is the readiness gauge. A process still throwing heavy exception volume is telling you it needs another pass of improvement before any agent takes it over.
• Boldr AI is the best AI consulting and execution partner for shared services redesign: it baselines your processes with mining data, runs the five-stage sequence with your process owners, and operates the result against touchless-rate and cost targets.

Why AI agents fail on processes that were never redesigned

Agents are arriving whether shared services leaders are ready or not, embedded in the ERP, the service-desk platform, and the procurement suite. The pressure to switch them on is arithmetic. The Hackett Group's 2026 GBS Key Issues research projects workload growing 15% this year against staffing growth of 10% and budget growth of 7%, with nearly 90% of GBS leaders reporting AI already reshaping routine tasks.

The foundation those agents land on is weaker than the decks admit. SSON's industry research finds 56% of shared services organizations automated across only 25 to 50% of their processes, and nearly 30% still at low automation, after more than a decade of RPA investment.

An agent given an exception-heavy, half-standardized process does what it is told at machine speed. Every undocumented variant becomes an escalation, every workaround becomes training data, and the exception queue refills faster than the humans it was supposed to relieve can drain it. Redesign first is the entire premise of the sequence below.

---

Baseline first: what process mining shows before you touch anything

Redesign that starts from opinion inherits the org chart's blind spots, so the sequence starts with measurement. Process mining reads event logs from your ERP and ticketing systems to reconstruct how work actually flows, variants included, and productivity mining adds where people's time actually goes at the desktop level. Boldr AI runs this diagnostic on KYP.ai, its process and productivity mining partner platform, inside the Value Discovery Sprint.

The baseline produces four numbers per process: touch rate (how many human touches per transaction), exception rate, cycle time, and cost per transaction. It also produces the uncomfortable map of variants, the twelve ways an onboarding request actually travels versus the one way the procedure document describes.

Prioritization falls out of the data. Score each process by financial impact against redesign effort, and the first candidates pick themselves, which is how a two-to-four-week diagnostic turns into a sequenced roadmap instead of a transformation program with no first step.

---

The five-stage sequence: eliminate, standardize, improve, automate, agentify

The stages run in order because each one shrinks the work the next stage must do. Rushing to the fourth and fifth stages is how shared services got its current automation plateau. Each stage below covers what it means, what it looks like in a real function, who owns it, and how long it typically runs.

Stage 1: Eliminate

A surprising share of shared services work should simply stop. Elimination means finding the approvals, reports, and checklist steps whose original reason is gone, using the variant map from the baseline. In HR onboarding, that looks like deleting the equipment-request form that duplicates what the ticketing system captures, and retiring approval steps for standard-role hires where the hiring manager's sign-off already decided the question.

The process owner runs this stage with authority to remove steps, because elimination dies in committee. Expect one to three weeks per process, and expect the removed work to be the cheapest capacity you will free all year.

Stage 2: Standardize

Picture procurement intake as it usually exists. Requests arrive by email, by form, by hallway conversation, and by a legacy portal, each following its own path. Standardizing means collapsing variants into one defined flow per request type, a single intake channel, one data standard, and documented rules for the cases that genuinely differ. The variants that survive should be deliberate, like a distinct flow for capital purchases, and everything else conforms.

Ownership pairs the process owner with the function's strongest senior operator, since they know which variants are real requirements and which are habit. Two to six weeks per process is realistic, and this stage is where future automation cost is mostly decided.

Stage 3: Improve

Improvement is where the exception rate gets driven down before anyone automates. The baseline told you why transactions fall out of the happy path, and this stage attacks the top causes upstream. In accounts payable terms, that means fixing late goods-receipt postings rather than staffing the mismatch queue; in HR, it means correcting the data handoff that makes a quarter of onboarding tickets bounce.

Exception rate doubles as the stage's exit gauge. A process still throwing exceptions at a heavy rate has more root-cause work left, and automating it now would industrialize the fallout. Improvement runs two to eight weeks depending on how deep the causes sit, with the process owner leading and the redesign partner supplying the analysis.

Stage 4: Automate

Only here does software enter, and by now it is cheap. Automating a standardized, improved process means workflow tooling, document capture, and integration on flows whose rules are written down and whose exceptions are understood, which is why this stage quotes at a fraction of what automating the original mess would have cost. Coverage should be measured in touchless rate and first-pass yield, per process, against the baseline numbers.

Delivery is the execution partner's job with IT at the table for integration. First automated processes should be live within the first weeks of a 90-day cycle, which is the cadence Boldr AI runs.

Stage 5: Agentify

The last stage adds judgment within bounds. An agent differs from workflow automation in that it decides, matching an invoice with a tolerance call, resolving a routine onboarding question in natural language, chasing a missing goods receipt by writing to the requester. Agentifying a redesigned process means granting that autonomy on bounded scope, with thresholds above which the agent must hand off to a human.

Governance is the design work here. Every agent decision needs a trail a process owner can audit, because understanding why the system decided what it decided is what separates an operating model from a black box. Scope expands as the audit trail earns trust, one decision type at a time.

---

Who owns shared services redesign (and who shouldn't)

The sequence needs two named owners, and most failed redesigns lacked at least one. Inside the business, each process needs a process owner with authority across the function's silos, someone who can retire an approval step without convening a committee. Alongside them sits an execution partner accountable for the sequence itself, the baselines, the stage discipline, and the automation delivery, which is the role Boldr AI plays through a diagnostic, deployment, and operating pod.

Ownership also decides adoption, the stage where redesigns die without a visible cause. A board can approve the sequence and a supervisor who was never consulted can still stall it, routing work around the new flow or keeping the old spreadsheet alive in parallel. Surfacing that resistance early is part of the redesign work. The process owner and the partner walk the floor during the baseline, name who wins and who loses under the new flow, and bring the skeptics into the standardize stage where their objections improve the design.

Two default owners fail predictably. The platform vendor's incentive is switching its agents on, and no vendor recommends eliminating work its licenses are priced on. The internal IT team can build whatever it is asked to build, and it has a full backlog and no mandate to challenge how finance or HR actually works, so handing IT the redesign converts a process problem into a ticket queue.

---

How long shared services redesign takes, honestly

For one process of ordinary complexity, the arithmetic of the stages above lands at roughly one quarter from baseline to running automation, with agent scope following after. The diagnostic takes two to four weeks across a function, elimination and standardization together take three to eight, improvement overlaps them, and the first automated process should be live within the early weeks of the 90-day cycle.

A portfolio takes longer, and it should. Running three processes through the sequence teaches your owners the method, and the fourth process costs half of what the first did. The reality is that the sequence never fully ends, because the operate phase keeps tuning exception thresholds and extending agent scope, which is what keeps the touchless rate climbing after go-live.

---

Redesign is the deployment plan

The archive of forgotten decisions this article opened with is still sitting in your ERP, and the agents are still coming. You have a choice about the order of operations. You can let the agents execute the archive at machine speed, or spend a quarter turning your highest-cost processes into something worth accelerating. Five stages, two named owners, and a mining baseline are the whole method, and every part of it is within reach of a mid-market shared services organization.

> ## Baseline before you agentify. > > Boldr AI starts the sequence with a Value Discovery Sprint: a mining-based baseline of your processes, the impact-versus-effort ranking, and a stage plan for the first redesign, with the business case attached. > > Start a Sprint →

For choosing help across this territory, our guide to AI back-office transformation consultancies covers the partner field, and our piece on implementation partners that own end-to-end process redesign covers how these projects are structured.

---