Pi Mapper

Process Intelligence
for AI Agents

AI agents need to understand your processes to work reliably - the systems, the workflows, the policies, the exceptions and the approval needs. I help operations teams build the operating model, the process knowledge layer and evals for that.


MB

Malika Bhatia

Founder, Pi Mapper
London, UK

I've followed process transformation through every wave it's had. AI agents are the latest - and the most demanding.

I trained as an IT engineer and started my career on the technical side - understanding systems from the inside out. That foundation turned out to be more useful than I expected when I moved into internal consulting, working across group functions and getting close to how large organisations actually operate: the politics, the handoffs, the gap between what IT delivers and what the business needs.

From there I moved into Six Sigma and lean - learning to map how work actually flows, not how it's supposed to. Then outsourcing, where the discipline of handing off processes to third parties forced a rigour about documentation and process clarity that most organisations hadn't needed before. Then large ERP implementations, where the gap between what a system assumes and what an organisation actually does becomes very expensive, very fast.

After that came process automation and RPA - which taught me that automating a broken process just breaks it faster. The underlying process intelligence has to come first. That lesson turns out to apply even more forcefully to AI agents, which are considerably less forgiving than a well-configured RPA bot.

Each wave has raised the bar on how well you need to understand your processes before you embed technology into them. AI is the highest bar yet - and process intelligence, encoded as agent skills, is the foundation that makes it crossable. Get that right, and AI agents become a genuine operational capability. Pi Mapper is my attempt to help organisations build it.

20+ years in process transformation - Six Sigma, outsourcing, ERP, RPA, and now AI
Deep specialism in process transformation, change management, TOM design (including the new Human + Agents + Apps operating model) and process intelligence for AI agents
Sector experience in banking and financial services including tier 1 banks, finance and HR operations
Based in London, working with UK and European enterprise clients

I write for process owners, operations managers, and the people responsible for making AI investments actually land. No hype, no vendor positioning - just honest thinking on what works and why.

Your documented process is fiction. That's why your AI project is struggling.

Every large operation has two versions of itself. There's the official one - the flowchart in Confluence, the training manual, the SOPs signed off in 2019. And then there's the actual one.

April 2025
Read →

What process mining actually tells you - and what people get wrong about it

Process mining has become a fashionable word in enterprise transformation circles. Most of what gets called process mining is really just process mapping with better software. Here's the distinction that matters.

How to talk to your AI vendor without getting sold a beautiful slide deck

The questions you should be asking before any AI engagement - and the answers that should make you pause.

The rework problem: why AI can predict it but operations teams still don't use it

Rework is one of the most detectable patterns in operational data. It leaves clear traces in event logs. The models can see it coming. So why does it keep happening?


Pi Mapper turns the way your processes really run into agent-ready Plugins - across four phases and one tight loop. The output is process intelligence your agents can actually use, with quality and governance built in from the start.

Phase 01

Foundation

Discovery & Process Mining

See the process as it truly runs.

Every good skill starts with how the work really runs. Before writing anything, we map the real process end to end - including the variants, the exceptions, and the rework that never made it into the manual.

aCapture how work happens today, from system logs and screen recordings to interviews and existing playbooks.
bReconstruct the end-to-end flow, surfacing variants, exceptions and hidden rework loops.
cPut numbers on each step: how often it runs, how long it takes, what it costs, how often it goes wrong.
dRank candidate tasks by value, friction and feasibility.
Process map Bottleneck heatmap Task inventory

Phase 02

Design

Operating Model

Decide how the work runs.

Decide what an agent handles, what plain code handles, and where people stay in charge. Sort out how quality is governed across the whole flow before a single skill is written.

aDraw the line between what needs an agent's judgment and what's better as plain, predictable code.
bMap how people and agents work together: the handoffs, escalations and approvals.
cSet the governance: quality checks, output controls, an audit trail, and clear owners.
dSet the guardrails and the bar each step must clear.
Operating model / RACI Automation boundary Governance spec

Phase 03

Build ⇄ Test loop

Skill Writing

Translate model into a skill.

⇄ Loops with Phase 04

Turn the plan into something the agent can actually run. Write clear, versioned instructions, give it the right context and tools, then send it straight to testing.

aBreak the process into clear steps you can give instructions for.
bAuthor the instructions, context and the tools each step can reach.
cEncode the policies, edge cases and escalation paths from Phase 02.
dVersion every change so improvements stay traceable.
Skill definition Tool & context spec Versioned changelog

Phase 04

Build ⇄ Test loop

Evals & Scoring

Prove it against reality.

⇄ Loops with Phase 03

Find where it breaks, then feed that straight back into the skill. Phases 03 and 04 keep cycling until it clears the bar. Then it's ready to ship.

aBuild a test set from real cases, including the messy ones.
bRun the skill and score against ground truth and the quality bar.
cWork out why it failed: a gap in instructions, missing context, or the wrong split of work.
dGo back to Phase 03, refine, run again, repeat until it clears the bar.
Eval suite Scorecard Failure analysis

The packaged output: a Plugin

A Plugin brings it all together - the skill with the connectors, governance, and evals it relies on, shipped as one versioned unit you can trust. That's what makes your process intelligence ready for agents.

From Phase 03

Skills

The process logic: instructions, steps, policies, and edge cases.

Systems of record

Connectors

The system connections and tools the agent works through.

From Phase 02

Governance

Guardrails, approvals, audit trail and human-in-the-loop gates.

From Phase 04

Evals

The quality bar ships with it, watched live and not just at build time.

From Phases 01–02

Context

Domain knowledge, reference data and the operating model.

Metadata

Manifest

Version, owner, and dependencies - so it stays easy to trace and update.

In production, the Plugin becomes your next dataset. Live data, new variants, and new exceptions flow back into Phase 01 - and the whole method runs again.


I take on engagements where I think I can genuinely help - and try to be honest when I'm not the right fit. The work is always practical and grounded in what your teams actually need, not what looks good in a deck.

01

Process Mining

Using event log data to surface how your operations actually run - not how they're documented. The essential foundation before any AI can reason reliably about your processes.

02

Human + Agent Operating Model Design

Designing the division of labour between your people and AI agents: what agents handle, where humans stay in the loop, and how you avoid the failure modes that come from getting that boundary wrong.

03

AI Context Layer Design

Building the empirical grounding layer that AI systems need to reason correctly about your operations - rather than performing confidently against documented fiction.

04

AI Evals & Production Readiness

Structured evaluation of AI systems before and after go-live. Are your agents actually doing what you think? Where do they fail, and under what conditions? Honest answers before the stakes get high.

If you think there might be a fit, I'd love to hear what you're working on.

Whether you're in the middle of an AI rollout that's not landing, or you're earlier in the journey and want to think it through - feel free to get in touch. No pitch, no pressure.

malika@pimapper.com

"The problem with most AI in operations isn't the AI. It's that nobody bothered to find out what the operations actually look like before they started."

Thanks - I'll be in touch shortly. In the meantime, feel free to email directly at malika@pimapper.com.

Pi Mapper - Editor

Select a post to edit, or create a new one.