↓ Download PDF
Arjun Jaggi
Enterprise AI Research · The 2027 Master Guide
every workload owned core the frontier slice your boundary
The Enterprise Guide to Model Strategy · 2027
Owned
Intelligence
Rent the frontier for the slice that needs it. Own everything else.
Author
Arjun Jaggi
Framework
The Ownership Stack, Version 1.0
Edition
2027
Readership
Boards · C-suite · Every department
Owned Intelligence · The 2027 Master GuideContents
Contents
How to Use This Guide03
Executive Summary04
The Guide in One Page05
The Ownership Stack on a Page47
Part I · The Case06
01The Shift07
02The Lock-in Tax09
03The Frontier Slice11
Part II · The Stack14
04The Ownership Stack15
04.1The Harness17
04.2Memory18
04.3Models19
04.4The Learning Loop20
04.5Evaluation21
04.6Compute and Runtime22
04.7Security and Governance23
Part III · The Economics24
05The Proof25
05.1The Rent-to-Own Curve26
05.2The Business Case Worksheet27
05.3The Metrics That Matter28
Part IV · The Enterprise29
06The Department Map30
06.1R&D and Engineering31
06.2Product and Design32
06.3Marketing and Sales33
06.4Solution Engineering and Success34
06.5Finance and Legal35
06.6People, Operations, and IT36
Part V · Leading the Shift37
07The Board Brief38
07.1The Investor Lens39
07.2The 2027 Roadmap40
07.3The First 90 Days41
07.4The Operating Model42
07.5The Owners43
Tools and Reference
Myths and Reality44
Questions Leaders Ask45
Glossary46
The Ownership Stack on a Page47
Self-Assessment48
Notes and Sources50
Solid fill, cited data
Hatched fill, directional illustration
Red, exposure and risk

Every number is cited to a named source in the notes or labeled directional. The terms Owned Intelligence, the Frontier Slice, the Ownership Stack, the Lock-in Tax, and the Rent-to-Own Curve originate with this guide and are subject to the license on the back cover.

Arjun Jaggi · Enterprise AI Research02
Owned Intelligence · The 2027 Master GuideHow to Use This Guide
How to Use This Guide
One guide, written for every seat at the table

This is a field guide, not an essay. Each page carries one idea, one exhibit, and one takeaway, so a board member can read it in twenty minutes and an engineering lead can build from it for a year. Start with the Executive Summary, then follow the path for your role.

EXHIBIT 1
Find your path through the guide
Which parts each reader should read closely, skim, or skip
Part IPart IIPart IIIPart IVPart VToolsBoard and investorsCEOCFOCIO and CTOCISOR&D and engineeringProduct and designMarketing and salesSolution eng. and successFinance, legal, people, opsRead closelySkim the exhibitsOptional
Source. Reader map of this guide. Part IV holds one playbook page for each department.
If you have 5 minutes
Read one page

The Guide in One Page, then the self-assessment. You will know your rung.

If you have 30 minutes
Read the summary

The Executive Summary and the opening page of each Part.

If you are building
Use the tools

The stack chapters, the worksheet, the 90-day plan, and your department playbook.

Arjun Jaggi · Enterprise AI Research03
Owned Intelligence · The 2027 Master GuideExecutive Summary
Executive Summary
The argument in five steps, and ten findings
1Intelligence gotcheap and open2Renting carriesa Lock-in Tax3The frontier earnsonly a slice4Own the stack,rent the slice5Owning pays back,then compounds
01

Intelligence is no longer scarce. GPT-3.5-level output got over 280 times cheaper in two years, and open models came within points of closed ones on some benchmarks [1].

02

Renting carries a Lock-in Tax. Price, retirement, egress, silent change, rate limits, and learning that never becomes yours.

03

The frontier earns a slice, not the whole. Routing each task to the cheapest model that passes cut cost by up to 98% [2], and by over half in another study [3].

04

The harness is the first asset to own. Whoever controls every model call controls the model choice, the data, and the cost.

05

Memory beats a bigger context window. Models use the middle of a long context worst [12]. Own your memory and test it under a budget [10][13].

06

Small trained models win their jobs. A 1.3B tuned model was preferred over 175B [4]. A 770M model beat 540B on a benchmark task [5].

07

Tuning became cheap. LoRA cut trainable parameters 10,000 times [6]. QLoRA tuned 65B on one GPU [7].

08

Learning loops build skill renting cannot. Checked rewards train models that improve on real work [8][14], and real work contains the checks [15].

09

Every department owns part of the asset. Test sets and data live across the company, not only in IT.

10

Owning pays back, then compounds. More cost first, less every quarter after, and every task improves the next model.

EXHIBIT 2
Where the guide takes you
Share of AI work by where it runs, today and at the 2027 target, directional
TODAYTHE 2027 TARGETmost work on rented frontier modelsowned models carry the bulktuned owned modelsslice
Source. Directional illustration, original to this guide. Your own split comes from your routing logs.
Arjun Jaggi · Enterprise AI Research04
Owned Intelligence · The 2027 Master GuideThe Guide in One Page
The Guide in One Page
Rent the frontier for the slice that needs it. Own everything else.

A frontier model should be a tool you rent for a narrow slice of work, not a dependency your business runs on. Owned Intelligence is the capacity to run, tune, and train models on a stack you control.

280x
cheaper GPT-3.5-level inference, 2022 to 2024 [1]
1.7
points between best open and closed models on some benchmarks [1]
98%
cost cut while matching the best model, best case [2]
100x
smaller tuned model preferred over a 175B model [4]

The Three Moves

ROUTESend each task to thecheapest model that passesOWNOwn the harness, memory,models, and learning loopTRAINTurn every task intodata for your next model
The test
If your vendor doubled the price tomorrow, how much of your AI would still run? That share is your Owned Intelligence.
The risk
The Lock-in Tax

Every rented call pays a price you do not set and builds no model you own.

The rule
The Frontier Slice

Use a frontier model only for work no owned model can pass. Measure the slice. Shrink it.

The build
The Ownership Stack

Harness, memory, models, and a learning loop. Any open-weight model, swappable.

Arjun Jaggi · Enterprise AI Research05
Owned Intelligence · The 2027 Master Guide
The Case
PRICE OF INTELLIGENCEfalling every yearclosedopenOPEN VERSUS CLOSEDthe gap nearly gonegiant, generalSMALL AND TRAINEDwinning its own job
I
The Case
Why owning intelligence became the right strategy, and what renting really costs.
In this part
01The Shift07
02The Lock-in Tax09
03The Frontier Slice11
Read this if you are  a board member, an investor, a CEO, a CFO, or anyone deciding where AI spend should go.
Owned Intelligence · The 2027 Master GuideSection 01 · The Shift
01
Section 01 · The Shift
Capable intelligence is now cheap, open, and small

Three curves moved at once. The price of a given level of intelligence collapsed, open-weight models caught the closed frontier, and small trained models began to beat giant general ones on the tasks they were trained for. Together they turn model ownership from a research project into an operating choice.

EXHIBIT 3
The same intelligence got 280 times cheaper
Price per million tokens at GPT-3.5-level performance (64.8% MMLU)
NOV 2022$20.00OCT 2024$0.07$0.01$0.10$1$10$100per million tokens, log scale
Source. Stanford HAI, AI Index Report 2025 [1].
EXHIBIT 4
Open models closed most of the gap
Best closed model's lead over best open model, some benchmarks
05108.0 points1.7 pointsJAN 2024FEB 2025
Source. Stanford HAI, AI Index Report 2025 [1].
Curve 1
Price fell

The cost of GPT-3.5-level output dropped over 280 times in two years [1].

Curve 2
Open caught up

The best open model now sits within a couple of points of the best closed one on some benchmarks [1].

Curve 3
Small got good

A 770M model, trained step by step, beat a few-shot 540B model on a benchmark task [5].

The takeaway
The scarce thing is no longer access to a capable model. It is owning the system that makes a model yours.
Arjun Jaggi · Enterprise AI Research07
Owned Intelligence · The 2027 Master GuideSection 01 · The Shift
Section 01 · The Shift
Why 2027 is the year to start owning

Each of these changes alone is a trend. Together they move the sensible default from renting intelligence to owning it, and they do it now, not in some distant future.

EXHIBIT 5
Enterprise AI is entering its third era
The question leadership asks changes as AI moves from trial to rent to ownership
2023 to 2024Experimentpilots on rented modelsTHE BOARD ASKSCan AI do this?2025 to 2026Rent at scalespend and dependence grow togetherTHE BOARD ASKSWhat are we spending?2027 onwardOwnan owned stack, the frontier as a sliceTHE BOARD ASKSWhat do we own?
Source. Framing of this guide. Era boundaries are directional and vary by company.

Five Forces Behind the Shift

ForceThe evidenceWhat it means for you
Price collapsedGPT-3.5-level output over 280x cheaper, 2022 to 2024 [1]Raw capability is now a commodity input
Open caught upGap to the best closed model fell to 1.7 points on some benchmarks [1]You can own a model close to the frontier
Small got goodA 770M model beat a 540B model on a benchmark task [5]Training matters more than size
Tuning got cheapA 65B model tuned on one 48GB GPU [7]Making a model yours fits a team budget
Learning got checkableReward given only when an answer is verified correct [14]Your own work can train your own models
The takeaway
The window is open while most competitors still rent. The ones who build the stack in 2027 will own the cost curve in 2028.
Arjun Jaggi · Enterprise AI Research08
Owned Intelligence · The 2027 Master GuideSection 02 · The Lock-in Tax
02
Section 02 · The Lock-in Tax
What renting really costs

The Lock-in Tax is the premium an enterprise pays for depending on a model it does not own. The token price is the smallest part of it. The rest is paid in exposure, rework, and learning that never becomes yours.

EXHIBIT 6
Six ways a rented model taxes the enterprise
Exposures that come with renting the model your business runs on
Priceset and reset by the vendorRetirementwithdrawn on their scheduleData egressevery prompt leaves the buildingSilent changebehavior shifts under one nameRate limitscapacity rationed at your peakLost learningyour usage never trains your modelRENTEDMODEL
Source. The Lock-in Tax model, original to this guide.
EXHIBIT 7
Most of the Lock-in Tax never appears on the invoice
Visible and hidden costs of renting the model a workflow depends on, directional
Token priceON THE INVOICEOFF THE INVOICERe-testing when a model is retiredRewriting prompts after silent changeReviewing data that leaves the buildingThe training data you never kept
Source. Directional illustration, original to this guide. Band widths show the shape of hidden cost, not measured values.
Arjun Jaggi · Enterprise AI Research09
Owned Intelligence · The 2027 Master GuideSection 02 · The Lock-in Tax
Section 02 · The Lock-in Tax
Where the tax bites, and how to stop paying it
EXHIBIT 8
Agents and customer-facing work carry the heaviest Lock-in Tax
Exposure of common workloads to each part of the Lock-in Tax, directional judgment
PRICERETIREMENTEGRESSSILENT CHANGERATE LIMITSLOST LEARNINGCustomer-facing assistantsMEDIUMHIGHHIGHHIGHHIGHHIGHInternal knowledge searchMEDIUMMEDIUMHIGHMEDIUMMEDIUMHIGHDocument extractionHIGHMEDIUMHIGHMEDIUMMEDIUMMEDIUMCode assistanceMEDIUMMEDIUMHIGHMEDIUMMEDIUMHIGHAnalytics and forecastingLOWMEDIUMHIGHMEDIUMLOWMEDIUMAgents that take actionsHIGHHIGHHIGHHIGHHIGHHIGH
Source. Directional assessment of this guide. Score your own workloads with the same six columns.

The Lock-in Risk Register

ExposureEarly signalMitigationOwner
PriceRate-card notices, spend rising faster than usageRoute through the harness, keep an owned fallback that passes the test setCFO
RetirementA retirement notice for a model you depend onKeep test sets ready so a replacement is validated in daysCTO
Data egressSensitive fields appear in outbound promptsRedact at the harness, run sensitive work on owned modelsCISO
Silent changeTest pass rate drops with no change on your sidePin versions where offered, run the test set dailyEngineering
Rate limitsErrors and queues at peak hoursPut routine bulk on owned capacityCIO
Lost learningTasks and outcomes are not being recordedLog every call and its outcome into owned memoryEvery team
The takeaway
Every mitigation runs through the same two assets. A harness you own and test sets you trust.
Arjun Jaggi · Enterprise AI Research10
Owned Intelligence · The 2027 Master GuideSection 03 · The Frontier Slice
03
Section 03 · The Frontier Slice
Rent the frontier only where it earns it

The Frontier Slice is the share of your AI work that genuinely needs a frontier model, because no model you own can pass it yet. It is real, it is valuable, and it is small. Everything outside it belongs on models you control.

EXHIBIT 9
The frontier earns only a thin slice of enterprise work
Where each kind of task should run, directional
ROUTINE BULKsmall open-weight model, ownedSPECIALIZEDfine-tuned owned modelFRONTIER SLICErented, routed
Source. Directional illustration, original to this guide. Your own split comes from your routing logs, see Exhibit 11.
EXHIBIT 10
Routing each task to the cheapest model that can pass it cuts cost sharply
Share of model cost removed while holding answer quality, as reported
0%50%100%FrugalGPTa cascade of modelsup to 98%RouteLLMa learned routerover 50%
Source. FrugalGPT, best case across tested tasks [2]. RouteLLM, over two times cheaper on tested benchmarks [3]. Both are published research results.
The takeaway
Paying frontier prices for routine work is a common, and avoidable, AI overspend. Route first, then decide what to rent.
Arjun Jaggi · Enterprise AI Research11
Owned Intelligence · The 2027 Master GuideSection 03 · The Frontier Slice

Who decides what runs on the frontier? Not a preference and not a vendor default. A test set. Every task gets a set of examples with known good answers, and the cheapest model that passes it gets the work.

EXHIBIT 11
Does this task earn the frontier?
Three questions route any task to an owned model, a tuned model, or the Frontier Slice
NEW AI TASKOwned model passesyour test set?noyesOwned modelthe defaultTuning or memorycloses the gap?noyesTuned owned modelfine-tune, or add memoryWorth the priceand the exposure?noyesFrontier slicerouted, logged, redactedHoldbuild the data first
Source. Decision logic of this guide. The test set is the gate. A model you own gets the task whenever it passes.

Four Rules for the Slice You Do Rent

Rule 1
Route it

Every frontier call goes through your harness. Never straight from an app.

Rule 2
Log it

Every call becomes an example you keep, for testing and for training.

Rule 3
Redact it

Sensitive fields are stripped before anything leaves the building.

Rule 4
Re-test it

Each quarter, owned models retry the slice. What they pass, they take.

The measure
The slice is a number you read from routing logs, not a guess. A healthy program watches it shrink every quarter.
Arjun Jaggi · Enterprise AI Research12
Owned Intelligence · The 2027 Master GuideSection 03 · The Frontier Slice
Section 03 · The Frontier Slice
Sorting the work, and when the frontier is the right call
EXHIBIT 12
Two questions sort almost every task
How new the task is, and how sensitive the data is, decide where it runs
Owned model, alwaysroutine work on sensitive dataFOR EXAMPLEcontract extraction, HR questions, patient notesFrontier with redaction, or buildnovel work on sensitive dataFOR EXAMPLEdeal analysis, novel legal questionsCheapest owned modelroutine work on open dataFOR EXAMPLEsummaries, translation, tagging, first draftsThe Frontier Slicenovel work on open dataFOR EXAMPLEnew research synthesis, hard reasoningROUTINENOVELDATA SENSITIVITYHOW NEW THE TASK ISowned models take this work as they learn
Source. Sorting model of this guide. The arrow is the point. As owned models learn, the slice moves left.

When the Frontier Is the Right Call

Case 1
Truly new reasoning

Problems with no precedent in your data, where no owned model can be tested yet.

Case 2
Rare, high-value tasks

Work so infrequent that tuning a model for it will never pay back.

Case 3
As a teacher

Drafting training examples for owned models, where the vendor's terms allow it.

Case 4
As a second opinion

A check on high-stakes outputs from owned models, not the first answer.

The takeaway
Renting the frontier is not a failure of strategy. Renting it by default is.
Arjun Jaggi · Enterprise AI Research13
Owned Intelligence · The 2027 Master Guide
The Stack
HARNESSMEMORYMODELSLEARNING LOOPyoursthe frontier, rented
II
The Stack
The four layers that make any model yours, and how to build each one.
In this part
04The Ownership Stack15
04.1The Harness17
04.2Memory18
04.3Models19
04.4The Learning Loop20
04.5Evaluation21
04.6Compute and Runtime22
04.7Security and Governance23
Read this if you are  a CTO, a CIO, a CISO, or an engineering, data, or R&D leader.
Owned Intelligence · The 2027 Master GuideSection 04 · The Ownership Stack
04
Section 04 · The Ownership Stack
The four layers to own

The Ownership Stack is the system that makes any model yours. A harness that controls every call, memory that keeps what the organization knows, models you run and tune, and a learning loop that turns work into better models. The frontier sits outside, reached through one governed line.

EXHIBIT 13
The Ownership Stack, four layers you own and one you rent
Where each part of an enterprise AI system lives, and the one line that crosses the boundary
YOUR BOUNDARYYOUR PRODUCTS, WORKFLOWS AND AGENTSHARNESSrouting · test sets · guardrails · tools · logsMEMORYdocuments · history · decisions · retrievalMODELSany open-weight model, swappableLEARNING LOOPRL environments · reward checks · tune and distillCOMPUTEowned or leased hardware · local runtimeopen Aopen BtuneddistilledyoursTASKS BECOME DATAIMPROVESthe sliceFRONTIER APIrented, the slice onlyoutside your boundary
Source. The Ownership Stack, original to this guide. A local runtime such as Zorp runs the stack on hardware you own [11].
The design rule
Own the harness first. A model you can swap in a day can never hold you hostage.
Arjun Jaggi · Enterprise AI Research15
Owned Intelligence · The 2027 Master GuideSection 04 · The Ownership Stack
EXHIBIT 14
The learning flywheel is what renting can never give you
How the Ownership Stack turns daily work into a better model
Harness logs every taskMemory keeps what mattersEnvironments score the workTune and distill the modelDeploy the owned modelEVERY TASKTRAINS THENEXT MODEL
Source. Flywheel of this guide. Reinforcement learning on scored tasks follows the pattern shown for reasoning in DeepSeek-R1 [8].
Layer 1
Harness

Routing, test sets, guardrails, and logs. Model-agnostic by design.

Layer 2
Memory

Kept on your side, and tested under a fixed budget [10].

Layer 3
Models

A 3B open model tuned on 6,444 domain questions and answers [9].

Layer 4
Learning loop

Environments that score real tasks so models improve by reward [8].

Build, Buy, or Configure

LayerBuild, buy, or configureWhy
HarnessBuildIt is the control point, and it must outlive any model
MemoryBuild on bought storageWhat your organization knows should never sit in a vendor's product
ModelsConfigure open weights, then tuneStart from the best open model, make it yours with your data
Learning loopBuildYour test sets and environments are the asset nobody can copy
ComputeBuy or leaseCommodity, so own the choice, not necessarily the hardware
FrontierRentFor the slice only, through the harness
Arjun Jaggi · Enterprise AI Research16
Owned Intelligence · The 2027 Master GuideSection 04.1 · The Harness
Section 04.1 · The Harness
The harness is the first thing to own

The harness is the software every model call passes through, the place where routing, testing, logging, and safety rules live. It is small to build, and it decides everything else. A company that owns its harness can change models in a day. A company that does not is rebuilding every time the market moves.

EXHIBIT 15
Every model call passes through one control point you own
The parts of an enterprise harness and the path of a single request
APPSand agentsGATEWAYidentity, budgetPOLICY ANDREDACTIONstrip sensitive fieldsROUTERpicks the modelOWNED SMALL MODELroutine bulkOWNED TUNED MODELspecialized workFRONTIER APIthe slice, redactedSCORES AND COSTfrom the test runnerOUTPUT CHECKSformat, policy, groundingTRACE STORE · every request, model, cost, and outcome, kept as your training data
Source. Reference architecture of this guide.

The Eight Requirements of a Harness

Model-agnostic interface. Any model can be swapped behind it without touching the apps
Full logging. Every call recorded with model, version, cost, and outcome
Test runner. Every candidate model scored on your tasks before it gets traffic
Router. Each task class goes to the cheapest model that passes
Redaction. Sensitive fields stripped before anything leaves the boundary
Budgets. Spend limits per team and per workload, enforced, not reported
Fallbacks. An owned model ready when a vendor fails or retires a model
Releases. A model change ships like software, with a test run, never as a surprise
Arjun Jaggi · Enterprise AI Research17
Owned Intelligence · The 2027 Master GuideSection 04.2 · Memory
Section 04.2 · Memory
Own what the organization knows

Memory is everything a model can draw on that is not baked into its weights, from the current task to years of documents and decisions. It is where your knowledge lives, so it must live on your side. A bigger context window is not a substitute. Models use information in the middle of a long context worst [12].

EXHIBIT 16
Memory decides what the model knows when it answers
Four kinds of enterprise memory, managed into a fixed context budget
WORKINGthe task at handEPISODICpast interactions and outcomesSEMANTICdocuments and factsPROCEDURALplaybooks and how-tosMEMORY MANAGERselect, compress, citeCONTEXT WINDOWa fixed budgetmost relevant memoryinstructionstasksupporting memoryput what matters at the startor the end, not the middle
Source. Architecture of this guide. Managed memory tiers follow MemGPT [13]. Placement follows Lost in the Middle [12].
Rule 1
Own the store

Memory lives in your systems, never inside a vendor's product.

Rule 2
Budget it

Test memory strategies at the budgets you will actually run [10].

Rule 3
Place it

Put the most relevant memory at the start or the end of the context [12].

Rule 4
Cite it

Every retrieved item carries its source, so every answer can be checked.

The takeaway
Models are replaceable. Memory is the asset that makes any model yours on day one.
Arjun Jaggi · Enterprise AI Research18
Owned Intelligence · The 2027 Master GuideSection 04.3 · Models
Section 04.3 · Models
Any open-weight model, made yours

Start from the best open-weight model, a model whose weights you can download, run, and change. Then climb the ladder only as far as your test set demands. Much enterprise work can be won on the first three rungs. MEDFIT-LLM, a 3B model tuned with LoRA on domain data, sits on rung two [9].

EXHIBIT 17
The adaptation ladder. Climb only as high as your test set requires
Six ways to make a model yours, from lightest to deepest, directional
1Prompt and retrieval2LoRA tuning3QLoRA tuning4Distillation5Reinforcement6Own modelno training neededsmall adapters[6]larger models, one GPU[7]small learns from large[5]reward checked results[8][14]deep domain trainingLESS EFFORT, LESS CONTROLMORE EFFORT, MORE CONTROL
Source. Directional illustration, original to this guide. Step heights show relative effort and control, not measured cost.

Choosing an Open-Weight Model

CriterionWhat to check
LicenseCommercial use allowed, and what the terms say about outputs and tuning
FitSize after quantization fits the hardware you will serve it on
QualityPass rate on your own test set, not on a public leaderboard
CoverageYour languages, your domain vocabulary, your document types
ToolingMature serving runtimes and tuning support exist for it
ProvenanceWeights come from a verified source, with a known history
Arjun Jaggi · Enterprise AI Research19
Owned Intelligence · The 2027 Master GuideSection 04.4 · The Learning Loop
Section 04.4 · The Learning Loop
Where owned models get better on their own

A reinforcement learning environment is a safe practice space. It gives a model real tasks, lets it act, checks the result, and rewards only outcomes that are verified correct [14]. Reasoning skill itself has been built this way [8]. The enterprise advantage is that real work already contains the checks. A public benchmark of 2,294 real software issues uses nothing more than each project's own tests [15].

EXHIBIT 18
A learning environment turns real work into a better model
The anatomy of a reinforcement learning loop an enterprise can own
TASK BANKreal tasks, known checksMODEL ACTSwith tools, in a sandboxCHECKERverifies the outcomeREWARDonly when verifiedUPDATEtune the modelTEST-SET GATEmust beat the current modelDEPLOYa better owned modelnew work and outcomes refill the task bank
Source. Architecture of this guide. Verified rewards follow Tulu 3 [14]. Reasoning from reinforcement learning follows DeepSeek-R1 [8].

The Check Already Exists in Your Work

FunctionWhat the environment verifies
EngineeringThe tests pass on a real issue from your own backlog
FinanceThe reconciliation balances to the ledger
Customer supportThe customer accepts the resolution without escalation
LegalThe clause matches the approved playbook position
Sales and solution engineeringThe answer matches an approved, current source
Operations and ITThe system reaches the expected state after the runbook step
The takeaway
Renting buys you a model's skill today. A learning loop buys you a model that is better next quarter.
Arjun Jaggi · Enterprise AI Research20
Owned Intelligence · The 2027 Master GuideSection 04.5 · Evaluation
Section 04.5 · Evaluation
The test set is the most valuable thing you will build

A test set is a collection of real tasks with known good answers. It is how you know an owned model is good enough, how you catch a vendor's silent change, and how the router decides. It outlives every model you will ever run.

EXHIBIT 19
Four layers of testing decide which model gets the work
The enterprise test stack, from automatic checks to human review
HUMAN REVIEWsampled live workevery weekSCENARIOSmulti-step tasks end to endevery releaseGOLDEN SETtasks with known good answersevery model changeCHECKSformat, policy, safetyevery callHOW OFTEN IT RUNS
Source. Testing model of this guide.

What Makes a Good Test Set

Drawn from real work. Not invented examples, actual tasks with actual answers
Owned by the department. The people who know the right answer maintain it
Includes failures. The cases that went wrong, so they never go wrong twice
Versioned. Every change recorded, so scores stay comparable
Refreshed quarterly. New work added as the business changes
Kept private. Never shared with a vendor who could train on it

The Four Numbers to Track

MetricDefinition
Pass rateShare of test tasks a model gets right
Cost per passed taskSpend divided by tasks passed, the true unit cost
CoverageShare of live workloads that have a test set
RegressionsTasks that passed before and fail now
Arjun Jaggi · Enterprise AI Research21
Owned Intelligence · The 2027 Master GuideSection 04.6 · Compute and Runtime
Section 04.6 · Compute and Runtime
Own the choice, not necessarily the hardware

Owning intelligence does not mean building a data center. It means the stack can run on any tier you choose, and moving between tiers is a configuration change, not a rewrite. A runtime, the software that loads and serves a model, makes that possible.

EXHIBIT 20
Run each workload on the smallest tier that passes
Where owned models run, from a laptop to a cluster, and where the frontier sits
DEVICElaptops and workstationssmall quantized modelspersonal and offline tasksnothing leaves the deviceSERVERone to a few GPUsteam-level servingQLoRA tuningdepartment workloadsCLUSTERshared GPU pool, owned or leasedhigh-volume servingreinforcement learningdistillation runsFRONTIERrented, outside the boundarythe slice onlythrough the harnessredacted inputsYOUR BOUNDARY
Source. Deployment model of this guide. Tier heights show relative capacity, not measured figures.
Principle 1
Quantize to fit

4-bit weights let a far larger model fit on the same GPU [7].

Principle 2
Smallest that passes

Pick the smallest model and tier that clears the test set.

Principle 3
Run where data lives

A local runtime such as Zorp serves models on hardware you own [11].

Principle 4
Lease without lock-in

Leased capacity is fine when the stack can move off it.

Arjun Jaggi · Enterprise AI Research22
Owned Intelligence · The 2027 Master GuideSection 04.7 · Security and Governance
Section 04.7 · Security and Governance
Owning moves risk from a contract to your controls
EXHIBIT 21
Six controls you take on when you own the stack
Each new duty, attached to the part of the stack it protects
YOUR BOUNDARYHARNESSMEMORYMODELSWatch outputsharmful or off-policy outputGuard memorysame access as the sourceVerify provenanceweights from verified sourcesPlan for failureincident plan, fast rollbackCheck licensesevery model and datasetTest before releasesafety and misuse tests
Source. Control map of this guide.

Who Is Responsible, Rented Versus Owned

AreaRented modelOwned stack
Where data goesLeaves your boundary on every callStays inside unless routed to the slice
Who changes behaviorThe vendor, on its own scheduleYou, through tested releases
Access controlVendor account settingsYour identity and permission systems
LoggingWhat the vendor chooses to exposeEverything, in a store you own
Model supplyThe vendor's choiceYou verify weights and licenses
UpdatesApplied by the vendorApplied by your platform team
For the CISO
The sensitive work should move in-house first, because it is the work you can least afford to send out.
Arjun Jaggi · Enterprise AI Research23
Owned Intelligence · The 2027 Master Guide
The Economics
paybackrentingowning
III
The Economics
The evidence that small owned models win, and the arithmetic of payback.
In this part
05The Proof25
05.1The Rent-to-Own Curve26
05.2The Business Case Worksheet27
05.3The Metrics That Matter28
Read this if you are  a CFO, a board member, an investor, or anyone building the business case.
Owned Intelligence · The 2027 Master GuideSection 05 · The Proof and the Payback
05
Section 05 · The Proof and the Payback
Small owned models win, and pay back

The case for owning is not ideology. Published research shows the same pattern again and again. A model trained on the right data for a defined job matches or beats a far larger general model, and the tools to do that training now fit on a single GPU.

EXHIBIT 22
A small model trained for the job beats a giant one renting its general skill
How much smaller or lighter, at a matched or better result, log scale
1x10x100x1,000x10,000xInstructGPT [4]1.3B tuned model preferred over 175B100x smallerDistilling step-by-step [5]770M model beats few-shot 540Bover 500x smallerLoRA [6]trainable parameters vs full fine-tune10,000x fewerQLoRA [7]65B tuned on one 48GB GPU, not 780GB+about 16x less memory
Source. InstructGPT [4], Distilling step-by-step [5], LoRA [6], QLoRA [7]. Each result is from the cited paper, on the tasks and benchmarks it tested.
Field evidence · regulated domain
MEDFIT-LLM

A 3B open model fine-tuned with LoRA on 6,444 domain questions and answers for medicine, one of the most demanding domains there is. Owned, inspectable, and small enough to run in-house [9].

Field evidence · reasoning
DeepSeek-R1

Reasoning was learned through reinforcement learning, then passed into open models from 1.5B to 70B parameters. The learning loop built the skill, and distillation carried it into small models [8].

The takeaway
Size is what you rent when you have no data. Data, tests, and a learning loop are what you own.
Arjun Jaggi · Enterprise AI Research25
Owned Intelligence · The 2027 Master GuideSection 05 · The Proof and the Payback

The Rent-to-Own Curve is the shape of the decision. Owning starts with a build cost renting does not have. Renting carries a cost that rises with every success and a price set by someone else. Past the payback point, every quarter widens the gap.

EXHIBIT 23
The Rent-to-Own Curve. Owning costs more first, then less every quarter after
Cumulative cost of renting every task against owning the stack and renting only the slice, directional
Q1Q2Q3Q4Q5Q6Q7Q8paybackprice riskRENT EVERYTHINGOWN THE STACK,RENT THE SLICEthe savingcumulative cost
Source. Directional illustration, original to this guide. The payback quarter depends on volume, the size of the slice, and what you already own.

Where the Return Comes From

Lever 1
Unit cost falls

Routine work moves to models that cost a fraction per task [2].

Lever 2
Exposure shrinks

Fewer of the six Lock-in Tax exposures apply to each workload.

Lever 3
Learning compounds

Every task improves the next model, a return renting never pays.

Lever 4
Options stay open

When the market moves, you swap the model, not the business.

Arjun Jaggi · Enterprise AI Research26
Owned Intelligence · The 2027 Master GuideSection 05.2 · The Business Case Worksheet
Section 05.2 · The Business Case Worksheet
Fill in five numbers, get the answer

The business case for owning is not a forecast. It is arithmetic on numbers you can read from your own systems. Fill this in for one workload first, then add them up.

EXHIBIT 24
Four lines of arithmetic decide whether owning pays
The Rent-to-Own business case, using numbers your harness already records
RENT EVERYTHING= V × cfOWN THE STACK= B + V × [ (1 − s) × co + s × cf ]YEARLY SAVING= V × (1 − s) × (cf − co)PAYBACK, IN YEARS= B ÷ yearly saving
Source. Business case model of this guide. Lower cost is not the only return, see Levers 2 to 4 on page 26.

Your Five Inputs

InputWhere to get itYour number
VTasks per yearHarness logs, by workload
sFrontier Slice shareShare of tasks no owned model passes, from routing logs
cfFrontier cost per taskVendor invoice divided by tasks served
coOwned cost per taskServing and hardware cost divided by tasks served
BBuild costTeam time for the quarters it takes, plus compute and tuning
Arjun Jaggi · Enterprise AI Research27
Owned Intelligence · The 2027 Master GuideSection 05.3 · The Metrics That Matter
Section 05.3 · The Metrics That Matter
Measure ownership, not activity

Counting prompts or seats measures activity. These eight numbers measure whether the company is building an asset or renting one. Review them monthly at the operating level and quarterly at the board.

EXHIBIT 25
Eight numbers tell the board whether Owned Intelligence is growing
The Owned Intelligence dashboard, each metric with its good direction and owner
Owned shareof tasks served by owned modelsgood directionCIOFrontier Sliceshare of tasks sent to the frontiergood directionCFOSwap timedays to replace any modelgood directionCTOTest coveragelive workloads with a test setgood directionDeptsCost per passed taskthe true unit costgood directionCFOData in boundarysensitive tasks handled insidegood directionCISOLearning cadenceowned releases that beat the lastgood directionML leadDepartments livewith a test set and an ownergood directionCOO
Source. Metric set of this guide. Every metric comes from the harness, the test runner, or the routing logs.
The headline
Owned share

The single number for the board. What share of AI work runs on models you control.

The cost truth
Cost per passed task

Spend divided by tasks that passed the test set. Cheap answers that fail are not cheap.

The resilience test
Swap time

How many days to replace any model. The lower it is, the less the Lock-in Tax can bite.

The takeaway
If a metric cannot be read from your own logs, you do not own the system that produces it.
Arjun Jaggi · Enterprise AI Research28
Owned Intelligence · The 2027 Master Guide
The Enterprise
R&DEngProductDesignMarketingSalesSol. eng.SuccessFinanceLegalPeopleOps and ITONEPLATFORM
IV
The Enterprise
A playbook for every function, from R&D to sales to legal.
In this part
06The Department Map30
06.1R&D and Engineering31
06.2Product and Design32
06.3Marketing and Sales33
06.4Solution Engineering and Success34
06.5Finance and Legal35
06.6People, Operations, and IT36
Read this if you are  a leader or team member in any department that uses AI.
Owned Intelligence · The 2027 Master GuideSection 06 · The Enterprise
06
Section 06 · The Enterprise
Department by department
EXHIBIT 26
Every department has routine bulk, specialized work, and a thin slice
Where AI work sits in each function
OWNED MODELS CARRYTUNED FOR YOUTHE FRONTIER SLICER&DArchive search and summaryDomain model on your dataCross-field hypothesesEngineeringCode completion and testsModel on your codebaseLarge cross-system redesignsProductFeedback themes, release notesIn-product assistantNew product conceptsDesignCopy variants, accessibilityModel on your design systemNew visual directionsMarketingOn-brand variants, localizationBrand-voice modelConcepts for new marketsSalesAccount research, call notesModel on winning proposalsComplex deal strategySolution Eng.RFP and security answersProduct technical modelNovel integration designSuccess and SupportTicket triage, answer draftsModel on resolved ticketsNovel escalationsFinanceInvoice and contract extractionModel on chart of accountsUnusual events, new scenariosLegalClause extraction, playbooksModel on your playbooksNovel questions, redactedPeople and HRPolicy and benefits answersModel on your policiesRarely neededOperations and ITIncident summaries, runbooksModel on runbooks, incidentsNovel failure modes
Source. Department model of this guide. Dashed pills are rented, everything else runs on models you own.
The takeaway
Every function already holds the two things a model needs to get good at its work. The data, and the test of what good looks like.
Arjun Jaggi · Enterprise AI Research30
Owned Intelligence · The 2027 Master GuideSection 06.1 · R&D and Engineering
Section 06.1 · The Enterprise
R&D and Engineering
R&D
for research leaders and scientists
Your archive of experiments is the most defensible training data you have.
WHERE THE WORK RUNSOWNEDArchive search and summaryTUNEDDomain model on your dataSLICECross-field hypothesesHOW IT GETS BETTERDATA YOU OWNexperiment archiveOWNED MODELdomain modelTHE METRICanswers passing reviewTHE CHECKexpert reviewFIRST THREE MOVES1Archive into owned memory2Test set from experiments3Tune a domain model
Unpublished results in rented tools
Engineering
for engineering leaders and developers
Your test suite is already a learning environment.
WHERE THE WORK RUNSOWNEDCode completion and testsTUNEDModel on your codebaseSLICELarge cross-system redesignsHOW IT GETS BETTERDATA YOU OWNrepos and CI resultsOWNED MODELcode modelTHE METRICpass rate on own issuesTHE CHECKtests passFIRST THREE MOVES1Route assistants via harness2CI tests as environment3Distill a code model
Source code leaving the boundary
Arjun Jaggi · Enterprise AI Research31
Owned Intelligence · The 2027 Master GuideSection 06.2 · Product and Design
Section 06.2 · The Enterprise
Product and Design
Product
for product leaders and managers
The model inside your product sets your gross margin.
WHERE THE WORK RUNSOWNEDFeedback themes, release notesTUNEDIn-product assistantSLICENew product conceptsHOW IT GETS BETTERDATA YOU OWNusage and feedbackOWNED MODELin-product modelTHE METRICmargin per AI featureTHE CHECKuser outcomeFIRST THREE MOVES1Tier every AI feature2Eval set from outcomes3Keep the model swappable
Margins falling as usage grows
Design
for design and research teams
A design system is a test set waiting to be written.
WHERE THE WORK RUNSOWNEDCopy variants, accessibilityTUNEDModel on your design systemSLICENew visual directionsHOW IT GETS BETTERDATA YOU OWNthe design systemOWNED MODELdesign modelTHE METRICaccepted without reworkTHE CHECKsystem checksFIRST THREE MOVES1Design system as checks2Research in owned memory3Tune for voice
Participant data in rented tools
Arjun Jaggi · Enterprise AI Research32
Owned Intelligence · The 2027 Master GuideSection 06.3 · Marketing and Sales
Section 06.3 · The Enterprise
Marketing and Sales
Marketing
for marketing and brand teams
Your approved content is your brand model's training set.
WHERE THE WORK RUNSOWNEDOn-brand variants, localizationTUNEDBrand-voice modelSLICEConcepts for new marketsHOW IT GETS BETTERDATA YOU OWNapproved contentOWNED MODELbrand-voice modelTHE METRICapproved on first reviewTHE CHECKbrand approvalFIRST THREE MOVES1Approved-content library2Tune a brand voice3Customer data stays inside
Off-brand output at scale
Sales
for sales leaders and account teams
Your won and lost deals are the most valuable data in the company.
WHERE THE WORK RUNSOWNEDAccount research, call notesTUNEDModel on winning proposalsSLICEComplex deal strategyHOW IT GETS BETTERDATA YOU OWNCRM and won dealsOWNED MODELproposal modelTHE METRICtime to first draftTHE CHECKdeal outcomeFIRST THREE MOVES1CRM data in owned memory2Test set from won deals3Owned proposal drafts
Customer data in consumer tools
Arjun Jaggi · Enterprise AI Research33
Owned Intelligence · The 2027 Master GuideSection 06.4 · Solution Engineering and Success
Section 06.4 · The Enterprise
Solution Engineering and Customer Success
Solution Engineering
for solution engineers and architects
Every RFP you have answered is a test set.
WHERE THE WORK RUNSOWNEDRFP and security answersTUNEDProduct technical modelSLICENovel integration designHOW IT GETS BETTERDATA YOU OWNapproved RFP libraryOWNED MODELtechnical modelTHE METRICapproved-source answersTHE CHECKsource matchFIRST THREE MOVES1Answer library with sources2Score on past RFPs3A source for every answer
Stale answers resurfacing
Customer Success and Support
for support and success teams
Every resolved ticket is a labeled example.
WHERE THE WORK RUNSOWNEDTicket triage, answer draftsTUNEDModel on resolved ticketsSLICENovel escalationsHOW IT GETS BETTERDATA YOU OWNresolved ticketsOWNED MODELsupport modelTHE METRICaccepted without editTHE CHECKcustomer acceptsFIRST THREE MOVES1Tickets as environment2Conversations kept owned3Score on last quarter
Unsourced answers to customers
Arjun Jaggi · Enterprise AI Research34
Owned Intelligence · The 2027 Master GuideSection 06.5 · Finance and Legal
Section 06.5 · The Enterprise
Finance, Legal, and Compliance
Finance
for the CFO and finance teams
Finance owns the number that decides the program, the Frontier Slice.
WHERE THE WORK RUNSOWNEDInvoice and contract extractionTUNEDModel on chart of accountsSLICEUnusual events, new scenariosHOW IT GETS BETTERDATA YOU OWNledgers and invoicesOWNED MODELfinance modelTHE METRICcost per passed taskTHE CHECKbooks reconcileFIRST THREE MOVES1Slice as a budget line2Quarterly worksheet3Track cost per passed task
Spend outgrowing usage
Legal and Compliance
for general counsel and compliance teams
Legal holds the most sensitive text in the company, so it moves to owned models first.
WHERE THE WORK RUNSOWNEDClause extraction, playbooksTUNEDModel on your playbooksSLICENovel questions, redactedHOW IT GETS BETTERDATA YOU OWNplaybooks and contractsOWNED MODELlegal modelTHE METRICsensitive docs kept insideTHE CHECKplaybook matchFIRST THREE MOVES1Classify by sensitivity2Privileged stays inside3Playbook as checks
Privileged material leaving
Arjun Jaggi · Enterprise AI Research35
Owned Intelligence · The 2027 Master GuideSection 06.6 · People, Operations, and IT
Section 06.6 · The Enterprise
People, Operations, and IT
People and HR
for the CHRO and people teams
Employee data is the data you can least afford to send out.
WHERE THE WORK RUNSOWNEDPolicy and benefits answersTUNEDModel on your policiesSLICERarely neededHOW IT GETS BETTERDATA YOU OWNpoliciesOWNED MODELpolicy modelTHE METRICemployee data kept insideTHE CHECKapproved answerFIRST THREE MOVES1Employee data stays owned2Test set from policy answers3Teach the slice rule
Employee data outside the boundary
Operations and IT
for the COO, CIO, and platform teams
Runbooks and incidents are checkable work, ideal for a learning loop.
WHERE THE WORK RUNSOWNEDIncident summaries, runbooksTUNEDModel on runbooks, incidentsSLICENovel failure modesHOW IT GETS BETTERDATA YOU OWNrunbooks and incidentsOWNED MODELoperations modelTHE METRICchecked steps automatedTHE CHECKexpected stateFIRST THREE MOVES1One shared platform2Incidents as environment3Owned fallbacks
No fallback when a vendor fails
Arjun Jaggi · Enterprise AI Research36
Owned Intelligence · The 2027 Master Guide
Leading the Shift
RENTEDROUTEDREMEMBEREDTUNEDOWNEDone rung per quarter
V
Leading the Shift
How boards, investors, and executives lead the shift, quarter by quarter.
In this part
07The Board Brief38
07.1The Investor Lens39
07.2The 2027 Roadmap40
07.3The First 90 Days41
07.4The Operating Model42
07.5The Owners43
Read this if you are  a board member, an investor, a CEO, or the executive who will own the program.
Owned Intelligence · The 2027 Master GuideSection 07 · Leading the Shift
07
Section 07 · Leading the Shift
The board brief
The questionA good answerA red flag
What share of our AI runs on models we control?A number, and it rises every quarter“We use the best vendor.”
What happens if our main vendor doubles its price or retires a model?A tested fallback, and a swap time in daysA plan to renegotiate
What is our Frontier Slice, and is it shrinking?Read from routing logs, trending downNobody measures it
Do our critical AI workloads have test sets?Yes, owned by the departments that use themQuality is judged by demos
Where does sensitive data go when AI uses it?Inside our boundary, or redacted first“The vendor has a policy.”
What do we learn from every AI task?Outcomes are logged and feed memory and tuningNothing is kept
What is the payback on the stack?The worksheet, filled in with our numbersA vendor's return study
Who owns Owned Intelligence?One accountable executiveEveryone, which means no one
EXHIBIT 27
The board's year
What the board reviews each quarter of the first year
QUARTER 1Approve the planname the owner, fund the harnessQUARTER 2Check the foundationworkloads behind the harness, with testsQUARTER 3See owned share moveowned models carry real workQUARTER 4Review the paybackreal numbers, plan the next rung
Source. Board cadence of this guide.
Arjun Jaggi · Enterprise AI Research38
Owned Intelligence · The 2027 Master GuideSection 07.1 · The Investor Lens
Section 07.1 · The Investor Lens
How to tell durable AI advantage from a rented veneer

Two companies can ship the same AI feature and have opposite futures. One owns the system that makes its models better and cheaper with every customer. The other rents a capability any competitor can rent tomorrow.

EXHIBIT 28
Three numbers separate durable AI advantage from a rented veneer
What to ask any AI-heavy company in diligence
01Owned shareof AI work on models the company controls02Frontier Sliceshare of work still sent to rented frontier models03Swap timedays to replace the model behind the product
Source. Diligence lens of this guide.

Signals in Diligence

SignalDurable advantageRented veneer
Model dependenceRuns on owned or swappable modelsHard-wired to one vendor
Gross marginImproves with volume as owned models take workFlat or falling as usage grows
DataProprietary data feeds memory and tuningData passes through, nothing is kept
EvaluationPrivate test sets tied to customer outcomesPublic benchmarks and demos
LearningEach release gets better from real usageImproves only when the vendor ships
SwitchingCan change models in daysA rewrite to change models
The takeaway
A company whose margin falls as its AI usage grows is renting its future from its supplier.
Arjun Jaggi · Enterprise AI Research39
Owned Intelligence · The 2027 Master GuideSection 07.2 · The 2027 Roadmap
Section 07.2 · The 2027 Roadmap
Four quarters, five rungs

Build in order, because each layer feeds the next. The harness produces the logs, memory keeps them, tuning turns them into models, and the learning loop makes those models better on their own.

EXHIBIT 29
One year from rented to owned, one layer at a time
The 2027 roadmap, with the Frontier Slice re-tested at every quarter gate
Q1 2027Q2 2027Q3 2027Q4 2027Harnessroute, log, testone harnessMemoryown what you knowown the memoryTune and distillfirst owned modelsfirst owned modelsLearning loopenvironments, own modelsenvironments and own modelsSlice reviewowned models retry it
Source. Roadmap of this guide. Durations are planning guidance for one business unit, and phases overlap in practice.
EXHIBIT 30
Five rungs of Owned Intelligence
Each rung moves more work off the frontier and onto a stack you control
Rentedfrontier runs allRouteda harness decidesfrontier runs mostRememberedmemory you ownfrontier runs muchTunedowned models livefrontier runs someOwnedlearning loop livefrontier runs the slice
Source. Maturity model of this guide. The self-assessment on page 48 places you on a rung.
Arjun Jaggi · Enterprise AI Research40
Owned Intelligence · The 2027 Master GuideSection 07.3 · The First 90 Days
Section 07.3 · The First 90 Days
What to do this quarter
EXHIBIT 31
Thirteen weeks from rented to a first owned model in real use
The first quarter, with a gate at the end of each phase
WK 0WK 2WK 6WK 10WK 13BaselineHarnessFirst owned modelRead the numbers
Source. Plan of this guide, sized for one business unit.
Weeks 1 to 2
Baseline
  • List every AI workload and vendor
  • Pick three workloads, one of them sensitive
  • Name one accountable executive
  • Gate. The inventory exists
Weeks 3 to 6
Harness
  • Put the three workloads behind the harness
  • Log every call with its outcome
  • Write the first test sets with each department
  • Gate. Every call passes through it
Weeks 7 to 10
First owned model
  • Score open-weight models on the test sets
  • Route the share they pass
  • Stand up owned memory for one workload
  • Gate. An owned model carries real traffic
Weeks 11 to 13
Read the numbers
  • Fill in the business case worksheet
  • Report the eight metrics
  • Plan the next quarter's rung
  • Gate. The board sees owned share

By Day 90 You Will Have

A complete inventory. Every AI workload, vendor, and owner on one list
A working harness. Three workloads routed, logged, and redacted
Test sets. Real tasks with known answers for each of the three
A first owned model. Carrying the share of work it passes
Your real numbers. Owned share, Frontier Slice, and cost per passed task
A board-ready plan. The next rung, funded, with a named owner
Arjun Jaggi · Enterprise AI Research41
Owned Intelligence · The 2027 Master GuideSection 07.4 · The Operating Model
Section 07.4 · The Operating Model
Who builds, who owns, who decides
EXHIBIT 32
A central platform, department ownership, and one council
The operating model for Owned Intelligence
AI COUNCILslice budget, risk appetite, prioritiesAI PLATFORM TEAMowns the harness, memory, compute,and the learning loopR&DEngineeringProductMarketingSalesSolution eng.SuccessFinanceLegalPeopleOperationsITeach department names an AI leadwho owns its data and test setseach department names an AI leadwho owns its data and test sets
Source. Operating model of this guide.
ActivityCouncilPlatformDept. leadCISOCFO
Set the Frontier Slice budgetACICR
Run the harness and platformIA, RICI
Build and maintain test setsICA, RII
Release an owned modelIRACI
Set data sensitivity rulesIRCAI
Report the eight metricsIRCCA

R, does the work. A, accountable for the result. C, consulted. I, informed.

Arjun Jaggi · Enterprise AI Research42
Owned Intelligence · The 2027 Master GuideSection 07.5 · The Owners
Section 07.5 · The Owners
Everyone has one question and one first move
EXHIBIT 33
Six seats, one question each, one first move each
Who holds which lever in the program
CEOWould our AI survive a vendor's bad quarter?Name it a strategic assetwith one ownerCFOIs the Frontier Slice shrinking?Budget the slice asits own lineCIO and CTOCan we swap any model in a day?Put every call behindone harnessCISOWhat leaves the building today?Redact at the harness,keep sensitive work ownedEngineeringDoes every workload have a test set?Build the test sets first,they decide what runs whereEvery teamWhich tasks are routine, which are rare?Keep the record of your work,it trains your next modelOWNEDINTELLIGENCE
Source. Role model of this guide.

The Minimum Viable Team

Platform engineerowns the harnessML engineerowns tuning and environmentsData engineerowns memoryProduct ownerowns the test setsSecurity architectowns what may leavepart time

What Kills the Program

No harnessapps wired straight to a vendorNo test setsno proof an owned model worksTuning before memoryno clean data to tune onA fixed slicenever re-tested, never shrinks
Arjun Jaggi · Enterprise AI Research43
Owned Intelligence · The 2027 Master GuideMyths and Reality
Reference · Myths and Reality
Eight things leaders believe that are no longer true
1MYTH
The best model is the frontier model.
The best model for a task is the cheapest one that passes your test set [4][5].
2MYTH
Owning means training a model from scratch.
Owning starts with the harness and open weights. Training from scratch is the last rung.
3MYTH
Open-weight models are a security risk.
Weights run inside your boundary keep data inside. The risk moves to controls you can audit.
4MYTH
Tuning needs a giant GPU cluster.
QLoRA tuned a 65B model on one 48GB GPU [7].
5MYTH
Bigger context windows replace memory.
Models use the middle of a long context worst [12]. Managed memory wins.
6MYTH
We can worry about lock-in later.
Lock-in is cheapest to avoid on day one. The harness makes speed safe.
7MYTH
Owning is always more expensive.
More first, less after payback. The worksheet shows when, for your numbers.
8MYTH
This is a decision for engineering.
Every department holds data and test sets. It is built across the company.
Arjun Jaggi · Enterprise AI Research44
Owned Intelligence · The 2027 Master GuideQuestions Leaders Ask
Reference · Questions Leaders Ask
Straight answers to the questions that come up first
Q
Will we fall behind the frontier?

No. You still rent it for the slice, and the harness lets you adopt a new frontier model in days.

Q
Do we need to buy GPUs?

Not necessarily. Lease or own. Tuning a large model now fits on one GPU [7].

Q
Can we train on frontier outputs?

Only where the vendor's terms allow it. Your own data is the safer base.

Q
How big a team?

Five roles to start. See page 43.

Q
How long to pay back?

It depends on volume and slice size. Use the worksheet, page 27.

Q
Does this fit a smaller company?

Yes. The harness and test sets cost little, and owning compute is optional.

Q
Where exactly do we start?

The harness, three workloads, and test sets. See page 41.

Q
What if the owned model is worse?

Then it does not get the work. The test set decides, task by task.

EXHIBIT 34
Where to start, in five steps
The answer to the question leaders ask first, as a picture
STEP 1Build theharnessevery call, one control pointSTEP 2Pick threeworkloadsone of them sensitiveSTEP 3Write testsetsreal tasks, known answersSTEP 4Route toowned modelswhere they pass the testSTEP 5Read thenumbersowned share and the slice
Source. Sequence of this guide. The full plan is the first 90 days, page 41.
Arjun Jaggi · Enterprise AI Research45
Owned Intelligence · The 2027 Master GuideGlossary
Reference · Glossary
The words in this guide, in plain English
EXHIBIT 35
How the five terms of this guide fit together
Read left to right
LOCK-INTAXavoidedbyOWNERSHIPSTACKshrinksFRONTIERSLICEraisesOWNEDINTELLIGENCEpays backonRENT-TO-OWNCURVE
Source. Concept map of this guide.
DistillationTraining a small model to copy a larger one.
Fine-tuningFurther training a model on your own examples.
Frontier modelThe most capable models, usually rented through an API.
Frontier SliceWork no owned model passes yet, so it is rented.
GuardrailA rule the harness enforces on inputs or outputs.
HarnessThe software every model call passes through.
InferenceRunning a model to produce an answer.
Lock-in TaxThe full premium of depending on a model you do not own.
LoRA and QLoRACheap tuning that trains a small add-on, not all weights.
MemoryWhat a model draws on beyond its weights.
Open-weight modelA model you can download, run, and change.
Owned IntelligenceThe share of your AI that survives a vendor doubling its price.
Ownership StackHarness, memory, models, and a learning loop you own.
QuantizationFewer bits per weight, so models fit smaller hardware.
Reinforcement learningTraining by rewarding good outcomes.
Rent-to-Own CurveOwning costs more first, less after payback.
RetrievalFinding documents and placing them before the model.
RoutingSending each task to the model that should handle it.
Test setReal tasks with known good answers.
TokenA small piece of text, the unit models charge by.
Verifiable rewardA reward given only when an outcome is checked correct.
WeightsThe learned numbers inside a model.
Arjun Jaggi · Enterprise AI Research46
Owned Intelligence · The 2027 Master GuideThe Ownership Stack on a Page
The Framework on a Page
Pin this page
The Ownership Stack on a Page
YOUR BOUNDARYYOUR PRODUCTS, WORKFLOWS AND AGENTSHARNESSrouting · test sets · guardrails · tools · logsMEMORYdocuments · history · decisions · retrievalMODELSany open-weight model, swappableLEARNING LOOPRL environments · reward checks · tune and distillCOMPUTEowned or leased hardware · local runtimeopen Aopen BtuneddistilledyoursTASKS BECOME DATAIMPROVESthe sliceFRONTIER APIrented, the slice onlyoutside your boundary
Move 1
Route

Cheapest model that passes the test set gets the task.

Move 2
Own

Harness, memory, models, learning loop. Any model, swappable.

Move 3
Train

Every task becomes data, and every quarter the slice shrinks.

Owned Intelligence

The share of your AI that would keep running, improving, and paying back if every frontier vendor doubled its price tomorrow.

The Frontier Slice

The work that genuinely needs a frontier model, because no owned model passes its test set yet. Measured, and shrinking.

The Ownership Stack

Harness, memory, models, and a learning loop, owned, with the frontier reached through one governed line.

The Lock-in Tax

The full premium of dependence. Price, retirement, egress, silent change, rate limits, and learning that never becomes yours.

The Rent-to-Own Curve

Owning costs more first and less every quarter after payback. Renting rises with every success.

Arjun Jaggi · Enterprise AI Research47
Owned Intelligence · The 2027 Master GuideSelf-Assessment
Self-Assessment
Ten questions. Count the yeses.

Answer for one business unit. A yes needs evidence, a log, a test, or a number.

1Every model call passes through a harness you own.Harness
2You can swap the model behind any workload in a day.Harness
3You know what share of AI spend goes to frontier models.Slice
4Your documents, history, and decisions sit in memory you own.Memory
5Your memory strategy has been tested under a fixed budget.Memory
6Every live workload has a test set with known good answers.Learning
7At least one workload runs on a tuned open-weight model.Models
8You keep the data you would need to train your own model.Learning
9At least one environment scores real tasks for reinforcement learning.Learning
10The Frontier Slice is re-tested every quarter, and it is shrinking.Slice

Your Score, Your Rung

0-2Rented3-4Routed5-6Remembered7-8Tuned9-10Owned
Arjun Jaggi · Enterprise AI Research48
Owned Intelligence · The 2027 Master GuideAbout This Guide
About This Guide
How it was built, and how to use it

About the Research

A practitioner guide, built to be seen first and read second. Owned Intelligence, the Frontier Slice, the Ownership Stack, the Lock-in Tax, and the Rent-to-Own Curve are original constructs of this guide, along with its maturity ladder, operating model, and department playbooks.

They rest on published research on routing and cascades [2][3], tuning and distillation [4][5][6][7], reinforcement learning [8][14], domain fine-tuning [9], memory [10][12][13], real-task environments [15], and local runtimes [11], with market figures from [1].

Exhibits labeled directional show the shape of an argument, not a measurement. The only numbers that matter for your decision are the ones your own routing logs and test sets will produce.

Companion Reading

White Paper No. 03, The Future Is Local. White Paper No. 09, The Local Inflection. White Paper No. 11, The Metered Enterprise. All at arjunjaggi.com.

Citing This Work

Arjun Jaggi, Owned Intelligence. The 2027 Master Guide to Enterprise Model Strategy, Enterprise AI Research, arjunjaggi.com, November 2026.

Using This Guide

Share it freely inside your organization. Pin page 47 by your desk, and run the self-assessment on page 48 with your leadership team each quarter.

AJ
Arjun Jaggi

Arjun Jaggi writes and advises at the intersection of enterprise AI strategy, the economics of adoption, and the operating work of turning AI into results. His frameworks turn the hardest questions of running AI at scale into terms a board can decide on and a team can execute. Engage at arjunjaggi.com.

Arjun Jaggi · Enterprise AI Research49
Owned Intelligence · The 2027 Master GuideNotes and Sources
Notes and Sources
Every source cited in this guide
Arjun Jaggi · Enterprise AI Research50
Arjun Jaggi
Enterprise AI Research
The frontier is a tool you rent. Intelligence is an asset you own.
Owned IntelligenceThe Frontier SliceThe Ownership StackThe Lock-in TaxThe Rent-to-Own Curve
© 2026 Arjun Jaggi. Original framework. Academic citation permitted with attribution; commercial use requires written permission.
2027 Master Guide · arjunjaggi.com