The Legacy Software Atlas

The map

Thirty-five of 46 candidate categories passed a strict legacy test.

Their legacy spend adds to about $178B a year on midpoints.

Scores run from 31 to 69.5 on a 0 to 100 scale.

Utility billing leads at 69.5, half a point ahead of mainframe.

The five factors

Each category is scored on size, pain, AI fit, lock-in and crowding.

Size and AI fit score whole numbers; the other three rank inputs into fifths.

The total is 35 plus 5 times the signed factor sum.

A one-step move in any factor moves the total by 5 points.

Where the money sits

Low and high estimates sum to $118B and $271B, each estimated independently.

Eighteen of 35 ranges cross a size-bin line, so boundary ranks are fragile.

Legacy spend is what buyers still pay for old systems, not sellable product.

Categories overlap: custom mainframe code runs inside banks, insurers and agencies.

The opening

The search asks for sizeable, painful, AI-fit categories with low lock-in and crowding.

One category passes all five tests, narrowly.

The fuller answer sits inside categories, at the level of single workflows.

Read the shortlist and workflow map next.

Buyers

Buyers are banks, insurers, hospitals, utilities, courts, agencies, distributors and manufacturers.

Enterprise buyers hold about $100.0B of legacy midpoint spend, government $42.1B.

Mid-market holds $20.6B and small business $15.3B.

Mid-market is usually the best first buyer: 27 of 35 categories broadly.

Build versus buy

About $42.5B of legacy spend sits with buyers who build or own the system.

That leaves $135.4B buyable across the 35 categories.

Twenty-two categories are critical; thirteen are important.

Critical categories hold $124.3B of legacy spend, of which $84.4B is buyable.

Two spend traps

Mainframe holds $14.5B of legacy spend but only $0.6B buyable.

Most of it goes to IBM platforms and integrators, not products.

Government benefits holds $12.5B of legacy spend and $0 buyable.

States still pay integrators under federal rules; there is no packaged core to sell.

A module at a time is the entry the study recommends.

Why buyers stay

Buyers stay because switching the system of record is regulated, slow and risky.

Lock-in averages contract length, non-competitive share, replacement time and regulator tests.

Seventeen categories carry a regulator yes; five calls are borderline.

Twenty-five of 35 categories rest on only two of the four inputs.

Regions

The US and EU hold the broadest evidence, at $81.6B of allocated spend.

India’s $28.8B is mostly a method artifact, driven by health facility counts.

Of 146 regional usage estimates, 20 are Grade B and 126 Grade C.

A second pass collected 369 named-organisation records not folded into the split.

Workflows

Each category splits into six workflows, 210 in all.

130 of 210 workflows do not touch the system of record.

That is where overlay products enter without a migration.

The pain

Pain scores run from 1.7 to 4.4 across the 35 categories.

Licensing and permits is most painful at 4.4; trade finance and mortgage servicing sit lowest.

Inputs count keep-alive awards, backend-failure reviews, app ratings, failures and workarounds.

Fifteen categories are graded on three inputs or fewer.

What AI can do

Twenty-four categories score AI fit 4 of 5, ten score 3, one rounds to 2.

AI fit is an ordinal task-level comparison, not an automation forecast.

The correlation between model fit and observed Claude usage is 0.46.

Fifty-one workflows record zero usage under that mapping, which is not missing data.

The attackers

The inventory holds 436 category-company challenger rows covering 423 names.

177 rows are coded AI-native, and 119 of those are overlays.

Overlay is the dominant observed strategy, followed by greenfield.

AI-native funding totals $11.0B disclosed across 298 recorded rounds.

73 of 177 AI-native rows are early, against 23 of 259 other rows.

AI-native traction itself counts 92 of those 177 rows.

Nineteen of 436 rows are stalled or failed; Olive and the networks sit outside.

Company age is unmeasured, so early traction is not a survival rate.

The workflow map

Workflow labour pools total about $6.3T across US occupations.

That counts whole occupations, so it ranks work but does not size markets.

Eighty workflows touch the core and carry a large ranking penalty.

The opportunity index runs from -644 to 498 with no pass mark.

The shortlist

The search demands size, pain, AI fit, low lock-in and low crowding at once.

Utility billing passes at 69.5; mainframe misses lock-in by half a point.

Agriculture misses only on size.

Construction accounting, telecom billing and student information miss on one test each.

The five tests

Utility billing has $2.5B of legacy spend, all of it buyable.

Its pain is 3.6, AI fit 4, lock-in exactly 2.0 and crowding 1.7.

The lock-in score sits on the boundary, where one award could tip it.

Manufacturing ERP and construction accounting follow at 67 and telecom and tax at 66.5.

Tiers are a different rule: nine categories sit at 62.5 or above.

A stricter founder filter needs 60-plus total and $1B of buyable spend.

It also allows at most two AI-native challengers with traction.

The five tests run simultaneously; the filter is a separate rule.

Five pass it: utility, tax, student systems, core banking and courts.

Eight more score 60-plus but fail it on traction, buyable spend or both.

Five miss on challengers alone: manufacturing, construction, telecom, payroll, licensing.

Agriculture and government benefits miss on buyable spend alone.

Mainframe misses on both; airline, lab systems and trade finance fell below 60.

How narrow the pass is

The AI-fit 3.61 could read near 3.12 on the blind audit scale, a fail.

Its lock-in 2.0 sits exactly on the bar, with zero non-competitive awards of twelve.

One non-competitive award would make the input 3 and lock-in 2.5.

All 50 sampled awards are European; no US award was recorded.

A regulator-rule change alone would push lock-in to 3.0 and fail it.

Kraken raised $1B at a reported $8.65B valuation in December 2025.

Sierra’s June 2026 utility agents contest the same opening.

Kraken and the unknowns

Kraken is utility billing’s selected challenger, with AI-native status recorded unknown.

An EDF-reported migration moved 5.8 million UK accounts in 15 months.

Low crowding means few AI-native challengers, not few challengers.

Missing mapped funding records research coverage, not the absence of competition.

How to use it

The full ranking sits on the map.

Check lock-in, crowding and the named events before pricing entry.

Every category page carries its evidence and limits.

The cases

Eight deep cases trace actual replacement attempts across five strategies.

Six cover companies with real traction; two cover failures.

None establishes a causal winner.

Five strategies

Overlays sell an AI layer on the incumbent core: 119 of 177 AI-native rows.

Greenfield sells a new core to buyers with nothing to lose.

Service sells the outcome with licensed software inside.

Migration-led sells the move itself, then runs the new system.

Acquire buys the legacy asset and modernizes it; nine of ten show traction.

AKASA measured gains beside the core; Adaptive entered below the incumbent price floor.

Rillet serves buyers forced off QuickBooks; bloop sold fixed-price conversion pilots.

Four stages

The first wedge lands a job the buyer already pays for, beside the core.

The first ten customers prove the hard part repeatability.

Expansion follows proof, usually modules or channels inside the buyer type.

Replacing the core comes last; most observed successes stop before it.

Kraken, Thought Machine and Candid show the pattern outside the eight cases.

What the eight show

Valon ran its own stack for six years before selling it.

Mechanical Orchard proved equivalence, not a public mainframe exit.

Adaptive and GovWell won beside and below incumbents.

Rillet sells a greenfield ledger; AKASA sells outcomes.

bloop and the trade finance networks failed.

Failure patterns

Distribution before product: selling before one buyer’s job is provably done.

Incumbent dependency: needing the incumbent’s cooperation to reach its customers.

Contracted-versus-live confusion: announced deals that never reach production.

Displacement stall: the overlay that never becomes the core.

How the failures worked

Incumbent absorption: at least 8 of 39 acquired rows went to same-category incumbents.

Cycle outran runway: bloop expected 18-to-24-month sales on $7.43M raised.

That pattern rests on one company; three AI-native migration rows exist to compare.

The migration-led strategy holds 33 rows across the whole table.

Consortium governance: no Contour bank owner led a round; 170 investors passed.

Parent-as-customer repeats it: Kaluza deadlocked and MassMutual shut Haven.

Point tools as features: platforms ship the step the specialist sells.

Epic, ICE, Intuit, NetSuite, AWS and HSBC each built or announced the same step.

Acquired rows read as about 26 strategic, 4 distressed, 6 absorptions, 3 unclear.

Pismo and Brace appear twice, so the 39 rows are 37 companies.

The GROW binding scheme is recorded pending; its exit quality is unmeasured.

Every test line is judgment, not a finding.

Diligence

The study narrows where to look; it cannot pick the company.

This chapter turns its findings into diligence steps.

Choosing a wedge

Pick a workflow that does not touch the system of record.

Prefer buyers who feel the pain and can sign in months.

Check whether the incumbent can block access to data or customers.

Treat contracted deals as promises until the system runs live.

The first buyer profiles

Utility billing: municipal and co-op utilities, 10K-100K meters, $60K-$600K a year.

Utility migrations plan at 6 to 18 months.

Government tax: 3,100 US counties buy, not build; collectors pay $50K-$3M yearly.

Student systems: US community and private colleges, 2,000-15,000 students, $150K-$1M yearly.

UK further-education colleges face a vendor-forced cloud move; many near Workday’s $300K minimum.

Core banking: banks and credit unions of $1B-$10B assets, $0.4M-$4M a year.

About a sixth of six-year contracts renew each year.

Courts: county and city courts serving 50K-1M people, $50K-$850K a year.

Best in states where each court picks its own system.

Utility risk: six non-AI challengers have traction; its awards are EU-only.

One non-competitive award would fail utility lock-in.

Tax risk: AI fit 3, a borderline regulator answer, and Tyler and FAST incumbency.

Student risk: lock-in 3.5 on 60-month contracts and 18% uncontested awards.

Banking risk: pain only 2.2, few switches, and Thought Machine at 68 banks.

Courts risk: lock-in 3.7 and a $1.04B market on the size line.

What to verify

Ask how long the last replacement took and what it cost.

Ask whether a regulator must approve or be told in advance.

Ask which workflows buyers already run outside the core.

Ask who else sells into the same workflow, AI-native or not.

Confirm the gap exists outside this study’s coverage.

Find the step the platform has not shipped; run it for one buyer first.

Valon raised over $290M running a servicer; Benefits Data Trust closed.

These are diligence candidates, not guaranteed markets or software TAM.

Where to look next

The workflow leads chart ranks the openings.

Each category page carries its providers, pain evidence and named events.

The limitations list what the study cannot answer.

Buyer interviews remain the missing evidence.