One company proves it. The portfolio compounds it.
An implementation-ready roadmap to the best path to AI ROI, then the people who build it.
Portfolio AI fails when it runs as N independent projects. We work at operator level: with the operating partner to put the tools and one standard in place, and inside each portfolio company to find the work, size it, budget it with the team, fit it to how the company already runs, and build systems that serve several use cases at once. That is what justifies the ROI, and it is why the first workflow is live in 100 days.
The shape of the engagement is the same for every fund; the plan is fitted to each one.
- days to the first production workflow
- 100
- example companies below, and what their first use cases return
- 3
- levers every use case is tagged by: revenue and cost
- 2
Azure
Follow one order through it.
Three layers: the systems the company already runs, a context layer that gives AI the company's own facts, and the people and agents who do the work. Switch who does the steps and watch the route change: the same order, the same company, a person at every step or one review at the end. The building blocks are the same at every company; the use cases below are what they serve.
Software runs the repeatable steps; a person approves the one that does not match.
- Your systems1. The order arrivesA PDF in the shared inbox, from a retailer portal.
- People and agents2. A person reads it and asksThis customer's prices, terms and stock, in one question.
- The context layer3. AI answers from the company's factsThe price list, the contract, the inventory, with the source shown.The dataThe documentsThe rules
- People and agents4. The person keys it in and confirmsInto the ERP, then the confirmation email.
- Your systems1. The order arrivesThe same PDF, picked up on arrival.Permissions: reads only what it is allowed to
- The context layer2. The agent reads it against the context layerPart numbers, prices, terms and stock, checked line by line.The dataThe documentsThe rulesConfidence: below the threshold, a person
- Your systems3. The agent keys it into the ERPStructured, checked, with the confirmation drafted.Limits: nothing above the approval limit
- People and agentsReview gate: A person approves the exceptionThe one order that does not match, with the diff in front of them.Review: the action shown before it is final
Foundation
Cloud
among othersOperator level, inside the company.
We do not run a survey from the fund. A pod sits with the operators, finds the work, sizes it in the company's own P&L, budgets it with the team and builds it on what the company already runs. Five steps, the same at every company; the plan is the company's own.
Identify
Every workflow across the company goes in: interviews with the people who own the work, the systems of record, the data that already exists. We work from the operator's seat, not from a survey sent from the fund.
Output: The inventory of work, by function.
Size
Each workflow gets an impact band, a time to value and a confidence, in the terms of this company's P&L rather than an industry average. The bands below are where that starts; the diagnostic narrows them to the company.
Output: The ranked inventory.
Budget with the team
A written business case per use case, an owner inside the company, milestone billing, and a plan the CFO signs before anything is built. The budget is the company's, not a line in the fund's.
Output: The budget and the owners.
Fit
The plan fits how the company already runs: the existing process, the systems of record, what already works. Technologies and best practice are chosen per use case, on the best tools for the job, not imposed from a template.
Output: The architecture and the stack.
Build once, serve many
One data foundation and one agent layer under several use cases: the second and the third cost a fraction of the first, which is what justifies the ROI. The first workflow is live in 100 days, and what ships at one company starts the next.
Output: The roadmap, and the first workflow in production.
Packaged workflows
A proven agent workflow, packaged so the next company can run it: the prompts, the connections to its systems, the tests and the runbook, kept in source control. The variance-close agent built at the first company is the starting point at the second, not a blank page.
Shared governance
One standard across the cohort: an audit trail, the cost of every agent's work attributed, and tests built into every agent, in each company's own cloud account. The same practice at every company, nothing new to buy.
Playbooks
How we ran it, written down: the implementation steps for each category of work, the sign-off criteria, the test baselines, and the measured economics from every engagement feeding back into the inventory.
Reuse used to demand standardized data and common systems before anything could move between companies. Agents do not. With the right context and a proven workflow, a workflow moves between portfolio companies even when data is messy and processes differ; the agent adapts to the company, not the company to a system. Transfer is hardest where the process is unusual and data access is poor, exactly what the scan detects.
What the operating partner gets
- The portfolio map: every company placed by AI maturity and use-case clarity
- One standard across the cohort: governance, review gates, the audit trail
- The ranked inventory of work per company, sized in its own P&L
- A wave plan and a budget the fund can underwrite
What the portfolio company gets
- The first workflow live in 100 days, on the systems it already runs
- A business case per use case, with an owner inside the company
- A team that owns the workflow after the pod leaves
- The data foundation the next use case starts from
Inside three portfolio companies. What it pays, and when.
A $200M precision components manufacturer
- Year 1
- $2.1M
- By year 2
- $5.3M
- By year 3
- $8.6M
Spend classification and supplier quotes
Impact$1Ma year · up to $2.5MAddressable spend runs about 25% of revenue. We count 2% saved; the band runs to 5%.
Time to value8 to 16weeksStarts with the spend data in the ERP. Savings arrive as supplier contracts come up for renewal.
Quote margin guidance and discount guardrails
Impact$1.2Ma year · up to $2.4MQuoted work runs about 60% of revenue. We count one point of margin recovered on it.
Time to value8 to 14weeksUses the quote and order history in the ERP and CRM. The first weeks go to the margin rules.
Order intake from emailed purchase orders
Impact$1Ma year · up to $1.4MOrder entry runs about 1.2% of revenue. We count half that effort removed, then 15% off for overlap.
Time to value6 to 10weeksReads the inbox the team already works. No new data foundation needed.
Illustrative companies, sized from the bands in measured engagements. Low end of every band at the company’s revenue; the third use case counted at 85% because it shares data and people with the first two; each one paying from the late end of its window and at full rate a quarter later. Value before the Build investment. Ranges, not promises.
Segment the fund before you build anything.
Where a company sits on the five AI maturity levels we defined is one axis: Level 1 (L1), people ask a chatbot; Level 3 (L3), software runs a whole workflow and a person reviews the exceptions; Level 5 (L5), an operation runs itself under human oversight. The other axis is whether the company knows what it wants AI to do or needs discovery first; Assess places every company on this map through a read-only service account in days, and the quadrant sets which 100-day wave a company joins.
Capable, searching
Data in place, work unranked. Inventory first.
Ready to build
Clear use case, data in place. Pilot now.
Needs discovery
Low maturity, unclear work. Scan and sequence.
Eager, unready
Named use case, data not ready. Fix the data first.
Parallel pods. Production inside the window.
Two named pods, two portfolio companies, one timeline: a pod is our business analysts, forward-deployed engineers and AI specialists assigned to one company for the wave, so the implementation fits the company where it is. Prototyping starts the moment a use case is concrete, production lands inside the first 100 days, and pods scale up or down at the fund's discretion.
Scan sorts the cohort first: ready-to-build companies skip to design; discovery-track companies run the same spine with one extra stage. Inside each build we still check our own work at every step.
What shipped, and what it changed.
A deal team and a lender: the first a Level 3 workflow in production (software runs it, people review the exceptions), the second the clean, connected data a portfolio motion is built on.
Client result · Deal team · Level 3 of 5Private equity firm4x faster CIM reviewManual processing down 75%; analyst queue from 70 to 10.Built withClaude on Databricks
Client result · Real estate · Level 2 of 540% faster underwritingAutomated valuation pipeline across the lending book.Built withDatabricks on Azure
Common questions.
What does the engagement look like for a PE firm?
The portfolio view is how a PE firm puts AI and data to work across the whole book. Each company goes on one map using our five AI maturity levels (L1 to L5), the work is ranked by what it does to revenue, cost and the pace of improvement, and parallel pods rebuild the best workflows. What ships in one company transfers to the next.
What does AI actually change inside a portfolio company?
How the work gets done. Our five levels show what kind of system a company needs: Level 1, people ask a chatbot; Level 2, it knows the company's data; Level 3, software runs a whole workflow and a person reviews the exceptions; Level 4, agents coordinate across functions; Level 5, an operation runs itself under human oversight. We move each company up where it pays.
How does Blue Orange decide which portfolio company moves first?
With the portfolio map: where the company sits on our five AI maturity levels on one axis, how clear its use case is on the other. A clear use case with the data in place pilots now; strong data with unranked work gets inventoried first; low maturity and unclear work is scanned last. The scan takes days per company, read-only.
Where does AI pay inside a portfolio company?
In three places. Revenue: pricing and margin guidance, sales effectiveness, retention and aftermarket, AI in the product itself. Cost: procurement, manufacturing and quality, supply chain, and back-office work that agents run instead of added headcount. Engineering: AI-assisted software delivery, so the company keeps shipping its own improvements after the pod leaves. Every candidate is banded by impact, time to value and confidence first.
How many companies does a wave cover, and how fast does production land?
Two named pods, two portfolio companies, one timeline. A pod is our business analysts, forward-deployed engineers and AI specialists assigned to one company for the wave. Prototyping starts the moment a use case is concrete, and the first production workflow lands inside 100 days: work that ran by hand now runs on an agent, with a person signing off on exceptions.
Why does the portfolio motion get cheaper with each company?
Because three things compound: the workflows (proven agent workflows, packaged so the next company can run them), the governance (one standard across the cohort) and the engineering practice (each company left able to ship its own improvements). Every workflow shipped makes the next one cheaper, instead of scaling linearly with headcount.
Map your portfolio.
Run the scan across your top ten. You get the portfolio map, each company's level, the rated inventory of work, and an implementation-ready roadmap to the best path to AI ROI: the first wave's plan, before a dollar is committed to building.