Questions we answer

Can we trust AI to act on its own?Governance and guardrailsWho allowed the AI to do that?Identity travels with every requestCan someone talk the AI into something risky?Risk-graded autonomyCan we see who did what, and why?One thread from request to resultWhere are the keys kept?One vault, no shared secretsWhy did the AI bill jump, and who spent it?One counted doorwayIs every task getting the right model?Right model, right taskWhat if our AI vendor goes down or raises its price?No single point of supplyAre we paying for tools that do the same job?One tool per jobDoes the AI forget everything between conversations?Governed shared memoryIs this producing real work, or just answers?The digital factoryWhich work should run itself?Managed, not scriptedCan we get more from the Copilot licences we already pay for?Copilot proposes, the platform verifiesWill people actually use it?Meet people where they workHow much time do people lose sorting inbound information?Intake to insight, untouchedCan the same platform run a real business?One governed foundation for every productWhat happens when it breaks at night?Detect, fix, escalateWhere is the list of what is open, and who owns it?One list, closed with evidenceCould we recover if we lost everything?Proven by restoringCan you prove it to an auditor?Challenge before acceptanceDoes it get better over time, or just older?Every lesson becomes a ruleWill this still work after the next upgrade?Clean Core for AIDo our documents match what is actually running?Documentation as a by-productCan we trust AI to act on its own?Governance and guardrailsWho allowed the AI to do that?Identity travels with every requestCan someone talk the AI into something risky?Risk-graded autonomyCan we see who did what, and why?One thread from request to resultWhere are the keys kept?One vault, no shared secretsWhy did the AI bill jump, and who spent it?One counted doorwayIs every task getting the right model?Right model, right taskWhat if our AI vendor goes down or raises its price?No single point of supplyAre we paying for tools that do the same job?One tool per jobDoes the AI forget everything between conversations?Governed shared memoryIs this producing real work, or just answers?The digital factoryWhich work should run itself?Managed, not scriptedCan we get more from the Copilot licences we already pay for?Copilot proposes, the platform verifiesWill people actually use it?Meet people where they workHow much time do people lose sorting inbound information?Intake to insight, untouchedCan the same platform run a real business?One governed foundation for every productWhat happens when it breaks at night?Detect, fix, escalateWhere is the list of what is open, and who owns it?One list, closed with evidenceCould we recover if we lost everything?Proven by restoringCan you prove it to an auditor?Challenge before acceptanceDoes it get better over time, or just older?Every lesson becomes a ruleWill this still work after the next upgrade?Clean Core for AIDo our documents match what is actually running?Documentation as a by-product

Jeffrey Walters · AI transformation leader

I take enterprise AI from pilot to production.

A career leading SAP and ERP programs taught me how large transformations succeed: standard first, a clean core, and decisions made on evidence. I now apply that to AI, and I built and run a governed AI platform to prove it.

See the concepts One-page brief

The overviewSix AI questions, and how we answered them

Jump to a questionAccess Cost Freedom Memory Trust Operations

Career

From developer to Director of SAP Architecture and Program Management, leading large ERP programs and international teams.

International delivery

An S/4HANA rollout in London, and programs delivered alongside major consulting organizations.

Now

Founder of ERP AI Integration, designing, building and running a governed AI platform as working proof.

Built on standard tools

Microsoft Teams

Microsoft 365 Copilot

SharePoint

Outlook

Planner

Microsoft Entra ID

Claude

Hermes

OpenClaw

n8n

Grafana

Langfuse

OneUptime

StackStorm

Docker

Cloudflare

GitHub

Case studies

Every hard question about AI, answered by a concept.

Each case study names the business question, the concept I used to solve it and the result. Pick the question you care about.

Secure by Design

AI that only does what each person is allowed to do.

Theme overview

“Who allowed the AI to do that?”Identity travels with every request “Can someone talk the AI into something risky?”Risk-graded autonomy “Can we see who did what, and why?”One thread from request to result “Where are the keys kept?”One vault, no shared secrets

Cost-Smart AI

Every request takes the cheapest path that still gives a top-quality answer.

Theme overview

“Why did the AI bill jump, and who spent it?”One counted doorway “Is every task getting the right model?”Right model, right task “What if our AI vendor goes down or raises its price?”No single point of supply “Are we paying for tools that do the same job?”One tool per job “Does the AI forget everything between conversations?”Governed shared memory

Work That Gets Done

Real work, delivered where people already work.

Theme overview

“Is this producing real work, or just answers?”The digital factory “Which work should run itself?”Managed, not scripted “Can we get more from the Copilot licences we already pay for?”Copilot proposes, the platform verifies “Will people actually use it?”Meet people where they work “How much time do people lose sorting inbound information?”Intake to insight, untouched “Can the same platform run a real business?”One governed foundation for every product

Run Like an Enterprise

Operated with the discipline of an enterprise program, sized for one owner.

Theme overview

“What happens when it breaks at night?”Detect, fix, escalate “Where is the list of what is open, and who owns it?”One list, closed with evidence “Could we recover if we lost everything?”Proven by restoring “Can you prove it to an auditor?”Challenge before acceptance “Does it get better over time, or just older?”Every lesson becomes a rule “Will this still work after the next upgrade?”Clean Core for AI “Do our documents match what is actually running?”Documentation as a by-product

In your organization

How I would apply this.

First

Listen and baseline

Find the AI already in use, map data, access and spend, and agree what good looks like with finance, security and audit.

Then

Prove one use case

One real use case with cost routing, an independent fallback, approvals and a way to undo, measured rather than assumed.

Then

Scale with one playbook

One change procedure with evidence, monitoring that fixes what is safe, and parts that swap without a rebuild.

Governance and guardrails

AI that knows its limits, and proves it.

Every agent works inside written rules: what it may do alone, what waits for a person, what it may never do, and how every decision is recorded.

Runs on its ownRoutine, pre-approved fixes

Known problems are fixed automatically from a short list of standard fixes, and every run is recorded.

Detect, fix, escalate →Waits for a yesRisky steps need the owner

A risky action posts an approval card in Teams. Only the owner can approve, and the decision is kept with who, what and why.

Risk-graded autonomy →Stops at a limitCaps and retry limits

Each agent has a spend cap, a quota guard moves work before a limit is reached, and retries stop and alert a person.

One counted doorway →Never allowedDestructive or protected

Destructive actions are blocked for automated fixers, critical services sit on a protected list, and a guest never sees owner tools.

Risk-graded autonomy →

See the guardrails and the governance behind them

Why it matters

Most AI pilots never reach production.

Gartner’s research on abandoned generative AI projects points to the same five causes. The concepts above answer each one.

Unclear business valuePilots answer questions but never change how work gets done.Work That Gets Done →Escalating costUsage spreads and nobody can say who spent what.Cost-Smart AI →Weak risk controlsNobody can say who asked the AI to act, or what it may touch.Secure by Design →Poor data qualityRecords drift away from what is really running.Documents that stay true →Poor change managementGood technology that nobody adopts, and no owner when things break.Adoption and change →

Source: Gartner research on why generative AI projects fail.

Lessons learned

What went wrong, and what it taught.

Problems every large company recognises, met on a real platform: the issue, the result and the lesson.

ReportingGreen on the report, red underneathAsk for evidence of outcomes, not status colours.Business continuityThe recovery plan nobody had testedDisaster recovery is proven only by recovering.Third-party riskOne supplier behind every safety netConcentration risk hides inside contingency plans.OperationsAlerts nobody sawDesign alerting around the person who must act, then prove that they receive it.GovernanceAn audit that agreed with itselfAgreement is not proof.Problem managementFixing the symptom, not the causeFund the fix only after the cause is proven.Cost controlRunaway spend from hidden retriesWatch failure rates as closely as spend.SecurityToo much access in one accountLeast privilege protects availability as much as it protects security.ArchitectureCustomization debtClean Core applies to AI: standard first, custom only by exception.

All lessons learned

Contact

Looking for someone to lead AI from pilot to production?

I would be glad to walk you through any part of this live and talk about the role you have in mind.

Email me LinkedIn One-page brief

j.walters@erpaiintegration.com

Go deeper: Lessons · The journey · How it’s built · Live health