ACG Minnesota Private Equity Summit
October 20, 2026 at The Machine Shop

The First100 Days

The deal closes. The investment thesis is still unproven. Diligence leaves two lists: what it found, and what it could not establish. AI can keep both connected to what is happening inside the business after close, so the people who own the decisions see the change while they can still act.

Open the slides

02 Found, not lost

Two ways a thesis breaks: found and not acted on, or never established.

H.I.G. Capital and South Bay Mental Health Center, 2012 to 2021: a diligence report reached the sponsor before close, and the company rediscovered the same issue twice. Paragon Metals, 2018 to 2026: the buyer saw a red flag, accepted an explanation, and learned the truth the day after closing.

The chain, from the court record

D. Mass., Memorandum and Order, May 19, 2021 (summary judgment, not a verdict)

Early 2012, before close

A third-party clinical diligence report, sent to H.I.G., reads no serious survey compliance issues up top and flags poor quality of supervision further in.

2014

Internal "Tiger Teams" recommend more licensed supervisors.

July 2016

Staff tell an assessment: We were unwilling to spend the money after the Tiger Teams... It was a thorough process, but nothing was implemented.

May 2021

The court finds sufficient evidence that the sponsor, through its board seats, had the power to fix the regulatory violations... but failed to do so.

October 2021

$25 million settlement, no admission of liability. H.I.G. pays $19.95 million.

✓Observed: court record!Facts disputed; settled without admission

Why this is the right opening

The finding traveled under the wrong label. "Supervision quality" reads like a clinical comment. In a business billing a government payer, it bore on whether claims were payable. Nobody translated it into "if this is still true on date X, money consequence Y follows."

The failure was a decision about spending, made by people with authority. Searchable documents would not have changed it. An owner with a budget, a deadline and a test of completion might have.

More than 60% of about 100 PE leaders said diligence outputs reach value creation plans only "rarely" or "some of the time."West Monroe, October 2022. Survey; vendor-published; method thin.

No base rate exists. Public cases are a biased sample of disputes that reached a court. Among six cases one report classified end to end, discovery failed in only two, and in both access was blocked (Marriott) or the seller concealed (Yahoo). Paragon (2026) adds a third pattern: concealment plus a red flag the buyer saw and did not test.

The counter-case: not fully established

Paragon Metals Holdings v. Smith, Del. Supr. Ct., July 1, 2026

2018, during diligence

A private equity buyer reviews more than 10,000 documents and spends about $1 million on advisers for a $100 million auto-parts acquisition.

November 29, 2018

A struck-through line in a draft reads Mike says there's a letter from ZF saying [they're] going to decrease their purchases. The CEO says volume will net out. The buyer does not ask for the letter.

February 1, 2019, the day after closing

The buyer learns its largest customers are cutting orders. It defaults on its acquisition loan and later invests another $37 million to stabilize the company.

July 2026

The Delaware Supreme Court holds the buyer could rely on the CEO's warranties despite imperfect diligence, and remands for damages.

An assertion was never tested against the document it described. That is a different failure from South Bay, and it needs a different AI job: turn every unproven claim into a test, and keep what management said apart from what the record shows.

✓Observed: court record!Fraud claim; damages not yet set; the trial court also noted COVID and a GM strike

Some risks are found and dropped. Others are never established. AI has to carry both across the close.

03 Six failures

"Losing what diligence taught you" is six different failures.

They need different remedies, and AI plays a different role in each. Select one.

AI drafts and connects where findings fail to travel, and makes an unowned finding visible where governance fails. It never supplies the owner or the budget.

04 The clocks

The calendar outranks the checklist.

"Day 100" is a convention. The deadlines that forfeit money are set by contracts, insurers and regulators, and they fall on both sides of it.

The clock every deal team knows, and a third still miss

SRS Acquiom 2026 Working Capital PPA Study, 1,570+ private-target deals, 2020 to Q3 2025. 19% valued at $25M or less.

Delivered after deadline32%

More than 30 days late8%

Deemed-final language23%

Buyer-favorable initial claim57%

Read the chart carefully. 32% is the 2020 to Q3 2025 aggregate; each full year ran 34 to 39%. The aggregate sits below every annual bar and the study does not explain why. Safe wording: "about a third, in every year." Late is measured; forfeiture depends on the agreement, so 32% late never means 32% forfeited. The 90-day median deadline comes from a separate drafting sample of 100+ deals closed in 2024.

✓Data!Vendor book

Missing it has a price

Schillinger Genetics v. Benson Hill Seeds, Del. Ch.

$14M

asset deal, February 2019

~2 mo

late on the 90-day closing statement

adjustment recovered; ~$80K sought

The court held delivery was "mandatory and unconditional under the APA." The indemnity notice, received one business day late, survived only because the seller suffered no prejudice. The sum is small; the mechanism is the point.

Every rebate accrual, unissued credit or unbilled-work item the QoE found becomes cash only if someone puts it in that statement on time.

✓ObservedLaw-firm summary (Hirschler, Mar 2022)

Replace "Day 100" on the plan with the real dates, and give each one an owner on day 1.

05 What kind of money

An earnings finding is an enterprise-value finding. Once.

Diligence findings hit four different money lines. The talk loses credibility the moment it adds them together.

One-time cashA penalty, a refund, a purchase-price adjustment, an insurance recovery.

Run-rate EBITDAA recurring cost or revenue effect.

Enterprise valueRun-rate EBITDA times a multiple. The same effect expressed as value, not an additional one.

Capex and working capitalCash that does not pass through EBITDA.

Where paid RWI losses come from

Aon 2026 Global Claims Study, North American policies placed since 2019, 2,000+ notified claims

Financial statements38%

Material contracts21%

Compliance with laws15.1%

Intellectual property11%

68%

Of the amount paid to insureds in 2025 was calculated on a valuation multiple.

6.7797x

The deal multiple a Delaware court applied to lost customer earnings in Dura Medic, relying partly on the buyer's own IC memo.

✓Data!Broker book; 38% in one place, 40% elsewhereDura Medic award amount unverified

DistribCo: switch each finding on

An illustrative distributor. $60M revenue, $7.2M QoE EBITDA, 8.0x entry. Nothing here is a case.

one-time cash, through the closing statement

run-rate or one-year EBITDA at risk

EV memo line, 8.0x, only where recurring and priced

Toggle a finding. The three lines are never summed.

◇Illustrative

A delayed price increase is a timing gap, not EBITDA times seven.

06 Before close

Closing changes what your knowledge is worth.

RWI excludes what the deal team knew. Remedies are fixed by the purchase agreement. A known issue has to be priced, carved out or deliberately accepted before close.

Why insurers denied RWI claims

Lowenstein Sandler 2023 survey of 154 market participants

Deal-specific exclusion20%

Late notice20%

Deal team's actual knowledge13%

Knowledge, 2020 survey39%

Policies exclude matters "identified in the disclosure schedules or otherwise discovered by the knowledge group during diligence" and generally do not cover purchase-price adjustments or special indemnities. RWI policies generally exclude what the deal team knew, so a finding deliberately deferred past closing usually needs a price, escrow or special indemnity instead. Policy terms vary; counsel decides.

ƒSurvey; policyholder-side firmPractitioner: ACC, 2018

The remedy you choose before close is the one you have after

Chicago Bridge & Iron v. Westinghouse, Del. Supr. Ct., June 2017

Westinghouse bought a nuclear construction business for zero dollars, subject to a working-capital true-up against a $1.174 billion target, under an agreement that barred post-closing claims for breached representations. It believed before closing that the accounts misapplied GAAP, yet "chose to close anyway," then sought nearly $2 billion through the true-up.

The true-up was "an important, but narrow, subordinate, and cabined remedy."Delaware Supreme Court. Observed; large strategic deal, used for the mechanism.

The last-chance review: every red flag gets one treatment

Part of the 100-day plan has to happen before signing.

07 One signal

One signal is a question, not an answer.

Arwood v. AW Site Services is the clearest lower-middle-market case in the record. The margin gap alone proved nothing.

Broadtree's waste-brokerage acquisition

Del. Ch., March 9, 2022, amended March 24. ~$16M cash and equity, October 2018. The seller kept no formal financials; the buyer rebuilt them with access the court called "extraordinary."

54%

44%

Gross profit expectedObserved, August 2019

SignalA 10-point gross-margin miss, ten months after close. Ambiguous: the seller's expert blamed the portable-toilet line for "nearly the entire" gap.

Frontline tipTwo long-serving staff: be on the lookout for "overage weights that end in .88." A lead, not proof.

Transaction dataPre- versus post-close comparison of overage revenue and dump tickets in the operating system shows estimated weights billed as actual. Corroborated. The scheme "was real."

$3.9M

contractual cap recovered on the financial-statement reps

Failed

fraud claim: the buyer's own access defeated justifiable reliance

40%

founder rollover equity plus an executive role

✓Observed: judicial record!Award, not collection; hindsight-selected case

Acting too fast also destroys value

Channel Medsystems v. Boston Scientific, Del. Ch., December 2019

Boston Scientific terminated a signed deal over a real red flag, an employee fraud that reached FDA filings, 23 days after the FDA had accepted the target's remediation plan and without answering six requests to meet. The court found no material adverse effect and ordered it to close.

+25%

Rise in a firm's probability of default after late customer payment (about 40% beyond a month).

8 of 100

Failing firms that are exposed to that risk. Escalating on lateness alone mostly targets customers that never fail.

✓Research: Banque de France, 2020

Caterpillar found the Siwei accounting discrepancies in a physical inventory count during integration, not in the monthly results. When the problem sits inside the numbers, the reports being watched may be the thing that is wrong. The LMM equivalent costs little: an early count and a billing-to-contract reconciliation for the top customers, before the 90-day statement is due.

A signal triggers verification by a named owner within a set time. Corroboration is the bar for acting.

08 Keeping conclusions alive

Deferral is legitimate. Deferral without protection is not.

The regulator did not fault Marriott for keeping Starwood's systems separate. It fined the absence of cheap interim controls while the decommissioning slipped.

Marriott and Starwood, 2016 to 2020

ICO penalty notice, October 30, 2020

September 2016

Acquisition. Marriott "was only able to carry out limited due diligence" on Starwood's systems. An attacker's web shell had been in place since July 2014.

Plan

Keep the systems separate; integrate over 18 months; retire Starwood's system in Q1 2018.

2018

Decommissioning slips to year end. Only limited changes were made to the Starwood system because it was expected to be decommissioned sometime in the future.

October 2020

£18.4M penalty (proposed £99.2M). Monitoring, logging and alerting could have been implemented relatively quickly. Retirement does not disapply the obligation.

✓Observed: regulator!Post-May 2018 conduct only; ICO did not rule on pre-close diligence

The people who hold the reasoning leave

33.3%

Mean share of deals in which PE investors replace the CEO after investing (79 firms, $750B+ AUM).

ƒSurvey: Gompers, Kaplan, Mukharlyamov

72%

Of firms that backed management and later replaced the CEO "waited more than a year." Turnover was unplanned in 34% of investments.

ƒSurvey: AlixPartners 2018

Storing more is not the answer. NASA mandated a lessons-learned database; in 2001, 27% of program managers had never heard of it. When nine hospitals standardized the spoken-and-written handoff at shift change (I-PASS), medical errors fell 23% with no added time. The difference is a structured transfer at a fixed moment when control changes hands. For a deal, that moment is closing.

Every conclusion carried forward gets three fields: the fact that would reopen it, the person who watches for that fact, and the date it will be checked regardless.At each review, mark it consistent, changed or retired. Past its review date with no new test, it is unverified, not true.

The more surprising risk is remembering too well.

09 The weekly meeting

The weekly meeting is part of the problem.

Group discussion favors information everyone already holds. After close, the shared information is the investment thesis. The unshared information is what the billing clerk noticed this week.

The usual agenda

Status round on the red/amber/green tracker, each owner grading themselves in front of the new owner
Integration workstream updates
Risks and issues
Any other business

This is where bad news gets softened.

Changes first

"What do you know this week that nobody else here knows, and which diligence assumption does it touch?"
Missing inputs and approaching decision dates
Rows that changed, remain disputed, need authority, or missed their verification date
Decisions recorded; evidence expected next week

Routine status is read beforehand. "We checked and found no change" is distinguished from "we have not checked."

almost half

Of groups chose the option they started with even when all information was shared (Faulmüller et al., 2010).

✓Research: laboratory

7,137

Workers at 66 firms: Copilot cut email time but produced "no significant average changes" in meeting time. Meetings need colleagues to agree on new norms.

✓Research: NBER w33795, randomized

0 min

Cost of changing the opening question. AI does not change the meeting by itself.

ƒPractitioner

10 What AI contributes

The AI loop: read, connect, compare, flag, route, verify.

AI's job is to keep the investment thesis connected to new evidence. AI does the first four steps; people decide and certify the last two. Below the loop: what has been measured, and what reliably goes wrong.

1 ReadPull every claim, date and number from the deal documents, with the page.

2 ConnectLink each claim to its money line, its clock and its owner.

3 CompareCheck new, authorized data against each claim.

4 FlagSay what changed, what it touches and what it does not prove.

5 Route, 6 VerifyA short brief to the named owner with a deadline; closed only on evidence a third party could check.

Measured: legal AI versus a lawyer baseline

Vals Legal AI Report, February 27, 2025. Vendors opted in; LLM-graded.

Data extraction, AI75.1

Data extraction, lawyers71.1

Document Q&A, AI94.8

Document Q&A, lawyers70.1

Summarization, AI77.2

Summarization, lawyers50.3

Lawyers stayed ahead on redlining. The manual baseline is not perfect either. One extraction question needing three clauses was answered exhaustively by no tool: they "often stopped after one relevant clause when several were needed."

✓Research!Vendor participation; dated

Reliable failure modes

17 to 33%

hallucination rate of purpose-built legal research tools, including "misgrounded" answers that cite a real source for something it does not say (Magesh et al., preregistered)

8.7%

moderate-risk incompleteness in AI-drafted clinical handoff notes, against 0% for physicians (Hartman et al., JAMA Network Open 2024)

Omission, the handoff's own failure, is AI's typical failure too. Models given contradictory passages "struggle to generate answers that accurately reflect the conflicting nature of the context" (WikiContradict).

Flagged is not saved

$3.59M

$834K

$6,899

Duplicate payments flaggedAgencies agreedRecovered when reported

Texas statewide recovery audit, January 2021. HHS-OIG found adjusted savings from Medicare's Fraud Prevention System were about 10% of what CMS reported. Judge a pilot on confirmed, actioned and collected items, never on alerts raised.

✓Observed: public auditsNo public account of a PE firm reporting an AI handoff outcome

The gain is not that AI remembers more. It is that the right evidence reaches the right person while they can still act, after it passes tests on findings you already know.

11 The finding card

The minimum that must travel with an important finding.

One row per finding that could move cash, EBITDA or a deadline. Typically 15 to 40 rows for an LMM deal, not every observation. DistribCo's five, with owner, clock and completion evidence.

◇IllustrativeStatus rule: accepted risk only when likelihood, size, bearer, approver and warning indicators are signed; otherwise an open questionNothing closes on a summary

Completion evidence is an artifact a third party could check. Not a status word.

12 Three levels

Start with AI on the documents. Grow into monitoring.

Level one sets the operating discipline; each level up lets AI do more of the reading and watching. Move up when the tests pass. A custom memory platform is not a prerequisite.

Documents, a spreadsheet and a 30-minute weekly meeting

Finding cards in a shared spreadsheet; a curated read-only folder; changes-first review.

Incremental benefitOwnership, dates, evidence status and clock awareness become explicit

Burden8 to 15 hours to seed; 30 minutes a week plus owners' updates (estimate)

Main failureGoes stale when the meeting stops; slow retrieval

Level 1 plus an approved AI tool on a curated set

A notebook or project restricted to the 10 to 20 key deal documents, instructed to cite page and say "not in sources."

Incremental benefitFaster first draft of the register; faster "where did diligence say X?"; checking a management statement against the record

Burden2 to 4 hours to curate; 15 to 25 known-answer test questions; weekly spot checks

Main failureFluent answers that blur assertion into fact; omissions

Connected assistant with recurring follow-up

Authorized access to mailboxes, drives and new reports, running scheduled checks against open rows. A team assistant such as Local AI's Captain would sit here, held to the same tests.

Incremental benefitNotices new material without anyone uploading it; resurfaces old concerns; reminders

BurdenPermissions clean-up, privilege and confidentiality review, admin consent, ongoing testing

Main failureSurfaces what users were never meant to see; alert fatigue

Acceptance tests before relying on any AI-built register

Known answersOn 10 findings the team knows cold, at least 9 reproduced with the right page

Open the citations10 random citations opened by a person; one miss triggers a full check of that section

Not in the documentsThree questions with no answer in the record must return "not found"

Planted itemsThree known obligations buried mid-document must appear

Assertion vs fact"Management assertion" items attributed, not stated as fact

Level 1 to 2 only after the register has run two weekly cycles. Level 2 to 3 only after 30 days of passing tests and a permissions and privilege review. Connected assistants inherit existing permissions, and an acquired LMM company's shared drives are rarely clean.

13 Misconceptions

Five things the room believes. Flip each one.

Select a card to see the evidence and the reframe.

14 Two weeks

A two-week starting plan for a small team.

Ten business days, existing software, one company. Roles: deal lead, operating partner, portfolio CFO, counsel. A proposed design, not a validated benchmark.

The 30-day decision

Continue

Every red flag has an owner and a clock, and no deadline is at risk.

Change

Rows are stale or owners lack budget authority. Fix governance before adding tools.

Stop the AI layer, keep the register

Two consecutive weeks where correction time exceeds time saved, a failed "not in documents" or access test unfixed within a week, or an uncited answer reaching a board or lender decision.

15 Three questions

Stop asking

Did we finish the 100-day plan?

Ask every week

Does the thesis still hold? What has changed? Who has to act, by when, and how will we know it is done?

AI doesn't make a good acquisition. It keeps testing whether the thesis is still true, and tells the people who can act while there is still time.

Working on this? Local AI builds the findings register this talk describes with deal and operating teams, cited to the page and tested on findings you already know: live capabilities in 21 days, on the cloud you already use, with code you own from day one. Book 30 minutes with us.

Open the slides Read the full report

ACG Minnesota Private Equity Summit, October 20, 2026

The First
100 Days

The deal closes. The investment thesis is still unproven.

Mike Halagan, Founder and CEO, Local AI

Thirty seconds on where I'm coming from. I'm not a private equity operator; I build AI. Before Local AI, I built machine learning at scale inside C.H. Robinson. Local AI builds production AI systems on the cloud a company already uses, with code it owns from day one. Building them for real is why I'm skeptical of AI demos: a demo shows the answers the vendor picked, and production shows you every answer it gets wrong. Here is the idea for the next half hour. When a deal closes, the buyer owns a business whose most important assumptions are still unproven: that the big customer renews, that the margins hold, that the old system can wait. Those assumptions live in a PDF called the investment committee memo, and then nobody looks at them again. For the first time, we can take those assumptions, connect them to the information flowing through the business, and keep asking one question: what changed? That doesn't replace judgment. It gives the people making the judgments a much better chance of seeing the right problem while they can still do something about it. Every number today carries a label: observed, data, research, survey, practitioner, vendor or illustrative.

The AI loopReadConnectCompareFlagRouteVerify

Your thesis is a list of assumptions.

AI reads it as claims to test.

Illustrative DistribCo: $60M revenue, $7.2M QoE EBITDA, 8.0x entry. Nothing here is a case.

Here's the reframe that makes this an AI problem. Before closing, the buyer writes an investment committee memo: why we're buying, what we're paying, what has to be true. Read it as an AI person and it's a list of claims. This is DistribCo, a made-up $60 million distributor we'll use for the rest of the talk. The top customer is 18% of revenue and renews in month 14. A 3% price increase applies to $10 million of contracts, if someone sends a notice in time. Rebates owed to customers are fully booked; in fact they're $150,000 short. One account manager holds $9 million of relationships. The old ERP system can wait. Each of those is a claim that can be true or false next quarter. There are three ways a claim breaks. Nobody tests it. It gets tested but never handed to the people running the company. Or it's handed over and nobody acts. Reading the memo, pulling every claim with its page number, is something today's models do well and fast. That is step one of the loop you see at the top of the slide, and I'll keep coming back to it: read, connect, compare, flag, route, verify.

The AI loopReadConnectCompareFlagRouteVerify

The calendar outranks the checklist.

Day 100 is a convention. Contracts set the dates.

Illustrative DistribCo. Data: SRS Acquiom 2026 (about a third of working-capital statements delivered late); Aon 2026 (about 51% of RWI claims notified after 12 months).

Day 100 is a convention, not a deadline. The dates that cost money are written into contracts and insurance policies. Two quick definitions. The working-capital statement is the buyer's post-close true-up of the purchase price; the contract usually gives about 90 days. About a third of them are delivered late. Late is not the same as forfeited; that depends on the contract, but in one Delaware case, Schillinger v. Benson Hill, a buyer delivered two months late and lost its adjustment entirely. RWI is insurance on the seller's promises, and about half of its claims arrive after a year. And before close there's a last-chance review: every red flag gets priced, escrowed, specially indemnified, accepted by a named approver, or assigned to be fixed after close. Insurance generally excludes what the deal team already knew, so the remedy you pick before close is often the one you have after; counsel decides the specifics. Now the AI part, which is simple and valuable. In DistribCo, a 3% price increase needs 60 days' notice before an anniversary on day 75. That means the notice is due on day 15, buried on page 40 of a contract. A model reads every contract, finds the clause, computes the date, and proposes an owner. The owner confirms. Miss that date and it's $300,000 of earnings for the year. This is the most boring, most valuable thing AI does in the first two weeks.

The AI loopReadConnectCompareFlagRouteVerify

The AI loop.

AI watches. People decide.

AI does the work

read · connect · compare · flag

People decide

route · verify

The right evidence reaches the right person while they can still act.

Proposed operating design. Research measures pieces of this loop, such as extraction and citation accuracy, not PE outcomes.

This is the slide I'd want you to photograph. Six steps. Read: pull every claim, date and number out of the deal documents, with the page it came from. Connect: attach each claim to its money line, its deadline and the person who owns it. Compare: as new information arrives, sales reports, CRM notes, invoices, check it against those claims. Flag: say plainly what changed, which assumption it touches, and, just as important, what it does not prove. Route: send a short brief to the named owner with a deadline. Verify: close the item only on evidence someone else could check, like invoices showing the price increase actually landed. The first four steps are where models are genuinely strong today: reading at scale, linking, and spotting a difference. The last two are where people stay in charge: deciding, and certifying it's done. One more way to say it, for the engineers in the room: treat the thesis like a production system. Each assumption is a monitor with a threshold and an on-call owner, and the agents run on change events from the CRM and ERP, not on someone remembering to look. And any action that touches the outside world is a proposal a second person approves.

The AI loopReadConnectCompareFlagRouteVerify

Memory is the hard part.

Not a vector database. Five things a fact has to survive.

What kind of memory

When it stops being true

Who may see it

One record stays one record

Written means findable

Ridgeline Supply is looking at options

sales inbox · Oct 20

what happened

how we do it

what is true now

Looking at options

true as of Oct 20 · recorded Oct 20

superseded

Renewed through 2028

true as of Nov 3 · recorded Nov 3

v1 kept, never overwritten · two clocks on every fact

Looking at options

author: deal lead · audience: DistribCo team

Dana, DistribCo

badge: DistribCo

Another company no badge

Ridgeline Supply

Ridgeline Supply Inc

RSI

one customer · one record

updated by source, not appended

Ridgeline Supply, one record

written 10:41:07

found 10:41:07 · the same second

Illustrative DistribCo record. How our own team memory works; for a portfolio company, a proposed design.

This is where I get to be the AI person in the room. Everyone says "the AI will remember." Here's what that takes, because we've built it for our own team. Five problems. One: what kind of memory. What happened in a meeting, what is currently true, and how we do things are different kinds of memory, and if you pour them into one pile the answers get muddy. We keep them in separate stores. Two: when it stops being true. Every fact has to carry its source and date, so a newer fact can replace it, and so you can answer the question a dispute asks: what did we know, and when? Three: who may see it. Every record has an author and an audience, and that's checked on every single call, not once at login. We run isolation tests that deliberately try to read what an account shouldn't see. Four: one record stays one record. The same customer shows up under three billing names, and the same fact gets written again every time someone mentions it. Duplicates don't just waste space; they crowd the right answer out of the top results. The fix is resolving entities and updating by source ID instead of appending. Five: written means findable. A managed memory service we tested took 10 to 65 seconds before a new record could be found. For a handoff, a one-minute gap is a failure, so we moved to Postgres with pgvector, where what's saved is findable immediately. So memory isn't a vector database. It's these five problems, and you grade it like a search engine. That's what you're about to see.

The AI loopReadConnectCompareFlagRouteVerify

“What changed this week?”

The assumption · IC memo

Top customer renews in month 14.

18% of revenue

New evidence · CRM and orders

“Looking at options.”

orders, last four months

A conflict.
Not proof.

Routed to

Account owner

Confirms by Friday

$1.9–2.7M

run-rate EBITDA at risk. Exposure, not a loss.

Memory underneath

change event · week 6

the one it touches · IC memo p. 12

Stored: question, owner, date

renewal intent? · account owner · Friday

Customer is leavingnot stored

Illustrative DistribCo finding; not a real customer's records. The money range is from the report's model.

Here's what the output of that loop looks like for one finding. The question the system answers every week is: what changed that puts our original thesis at risk? The card shows the original assumption, with where it came from: the top customer renews in month 14, from the customer section of the investment memo. Then the new evidence: a CRM note that says the customer is looking at options, and three months of falling orders. Then the most important line, what AI can establish: these conflict with the assumption, and they do not prove the customer is leaving. A good system says what it doesn't know. Then the money at stake, labeled as exposure. Then a named owner and a deadline: the account owner confirms by Friday, and the CFO reruns the downside case. And what closes it: a written answer from the customer, or a named person accepting the risk. If we assumed a top customer would renew, I don't want that sentence sitting in a PDF for fourteen months. I want to know when the orders change and someone mentions competing suppliers. That's a very different use of AI than summarizing a diligence report. DEMO CUE: next slide is the live demo. Before advancing, have Teams and Claude Desktop open behind the deck, and the backup recording queued. Bridge line: "Let me show you the foundation this is built on."

The AI loopReadConnectCompareFlagRouteVerify

Live: memory across the close.

Recorded in Teams. Recalled in Claude Desktop.

Microsoft Teams · the deal lead

Remember for DistribCo: Ridgeline says it’s “looking at options” and orders are down four months. Dana verifies renewal intent by Friday.

Claude Desktop · the operating partner

What do we know about DistribCo’s top-customer renewal risk, who owns it, and by when?

source · owner · deadline

Under the hoodTeams or Claude Desktop→Gateway→Policy check on every call→Postgres + pgvector→Cited answer

Different person. Different app. Same knowledge.

Live demonstration on illustrative DistribCo data; no client information. A recording is the fallback. Shows shared team memory only; comparing and flagging new evidence are the steps built on top.

LIVE DEMO, 60 to 90 seconds. Setup is in the prep doc: three DistribCo records seeded the night before from the memory-test account, stale DistribCo records removed first, both apps signed in to Captain on the presenting laptop, notifications off, text size up, the backup recording open in another window. Nothing from a real client goes on the projector; DistribCo, Ridgeline Supply and every name are made up. Say: "Let me show you the foundation this is built on. The deal team learns something before close. Watch what happens when someone else needs it later." Step one, in Teams, as the deal lead (the memory-test account), paste: Remember for DistribCo: a sales email today says Ridgeline Supply is "looking at options" and its orders are down four months in a row. Dana Whitfield, the account owner, verifies renewal intent by Friday, October 24. Source: sales inbox, October 20. Step two, switch to Claude Desktop, as the operating partner, ask: What do we know about DistribCo's top-customer renewal risk, who owns it, and by when? The answer should combine the seeded assumption (Ridgeline, 18% of revenue, renews in month 14, IC memo page 12) with what you just recorded, and name Dana and Friday. Point at three things: the source, the owner, the deadline. Different person, different app, same knowledge. If there is time, one more question: Does DistribCo's memory say anything about a Ridgeline termination right? The right answer is "nothing recorded": a memory that says what it does not know. Point at the under-the-hood line: "Every request passes a policy check that knows who you are and which client you can see, and the answer cites the record it came from." Then frame it honestly: "That's the transfer step: knowledge crosses the close. Watching for change, comparing new evidence against what's stored here, is what you build on top, and it's held to the tests on the next slide." Fallback: if anything stalls for more than about fifteen seconds, say "Live demos. Here's the recording," and play it. Don't troubleshoot on stage.

The AI loopReadConnectCompareFlagRouteVerify

Trust it only after it passes your tests.

9 of 10 known answers

10 citations opened

3 “not found” answers

3 planted obligations

Claims stay claims

Permissions hold

Flagged is not saved · Texas recovery audit

flagged

$3.59M

agreed

$834K

recovered

$6,899

Ours: the right answer in the top five on 92% of 132 questions. Zero injection leaks.

Proposed acceptance tests, following NIST AI 600-1. Observed: Texas statewide recovery audit, January 2021. No public account of a PE firm reporting an AI handoff outcome.

Now the part I know best, because I build these systems. Models fail in predictable ways. They leave things out: in one legal benchmark, tools often stopped after the first relevant clause when several were needed. They cite a real source for something it doesn't say. And when two documents disagree, they tend to smooth it over. Published error rates for legal research tools vary widely by tool and task, so I won't put a single number on a slide. Here's what I'd do instead: test it on your own deal. Ten findings your team knows cold; it should get nine with the right page. Open ten citations at random. Ask three questions that have no answer and make sure it says not found. Plant three obligations in the middle of long documents and see if it catches them. Make sure management's claims stay labeled as claims. And make sure it only shows people what they're allowed to see, because an assistant connected to an acquired company's shared drives will happily surface payroll. We hold our own systems to this. Our team memory is graded on 132 known questions: the right answer is in the top five results 92% of the time, and we track exactly which questions it misses. Our cited-answer test includes a page that tries to hijack the assistant, and it leaked nothing. And the scariest bug we found was quiet: 68% of one data source was unreachable, and the answers still looked clean. A clean-looking partial answer is worse than an error. Last: measure the right thing. In a Texas audit, $3.59 million of duplicate payments was flagged, $834,000 was agreed, and $6,899 had been recovered when the report came out. Flagged is not saved. Score the pilot on confirmed, actioned and collected items. And start small: have AI draft the register from the deal documents this month, every row cited and checked by a person. Step back a level if correction time exceeds time saved two weeks running.

The AI loopReadConnectCompareFlagRouteVerify

Where this goes:

a thesis that checks itself.

Extractorevery assumption + page

Reconcilerlists conflicts, never resolves

Devil’s advocateargues against one at a time

Deal memorylessons with “applies when”

Alert budgeta capped number of flags

Everyone can rent the same model. Own the record of what happened on your deals.

Proposed design: what we would build with a design partner, not a shipped product. Illustrative DistribCo. The memory underneath is what the demo showed; the agents on top are the build.

So where does this go? Here's the system I'd build if a firm wanted a real edge in the first hundred days. Every assumption in the thesis becomes a live record, and the key AI idea is that each record carries two clocks: when the thing was true, and when we wrote it down. Most memory systems keep one. With two, you can answer what a dispute asks, what did we know and when, and you can tell a fact that changed from a fact we recorded late. Each record also has a reopen trigger: orders fall two months running, a renewal date moves. When the trigger fires, the record reopens, and the old version is kept, never overwritten. That matters: at NASA, Columbia's managers reused a stale written rationale, and long-term-memory benchmarks show AI loses about 30% of its accuracy exactly on updated facts. Around the record sits a small team of agents. An extractor per document. A reconciler that lists conflicts between documents with quotes, and never resolves them; that's a person's job, and no public benchmark even tests cross-document reconciliation yet. A devil's advocate that argues against one assumption at a time, because models are agreeable by default and research shows a dissenting AI improves how well people calibrate their trust. Deal memory: lessons from past deals, with gates. After fifteen HVAC deals, it raises refrigerant regulation on the sixteenth, but each lesson carries "applies when" and "does not apply when," and lessons from deals you lost travel only as questions, never as the seller's documents. And an alert budget: a fixed number of flags a week, scored on what got collected, because flagged is not saved. Sensitive material, like founder interviews, can run on a model inside your own environment. The output is a weekly thesis diff: what changed, against which assumption, who owns it. To be clear, this is what we'd build with a design partner. The memory underneath is what you just saw in the demo. And at eleven I'll show the other half: how you score a system like this against deals you've already closed. Everyone can rent the same model. What you can own is the record of what happened on your deals.

The AI loopReadConnectCompareFlagRouteVerify

Monday, with the tools you already have.

ChatGPT Business or Claude Team.

Do this

Where it stops

Build the assumption register

IC memo and QoE in a shared project: every claim, with its page

Pull every clock

SPA, contracts, insurance: the date, the clause, the page, an owner

Run the six tests

plant three items, ask three it cannot answer, open ten citations

Memory is per person

a shared project is a folder; the handoff across the close does not happen slides 5, 7

Nothing watches for change

it reads what you hand it, when you ask slide 6

No “not found”, no audit trail

who recorded what, when, and what it does not know slide 8

Great at reading what you hand it, once. Not built to remember, watch or prove.

Capabilities per OpenAI and Anthropic help pages, October 2026; limits are approximate and change often. Neither product trains on business-plan data by default.

Here's the part I owe you: what you can do Monday with the subscription most of you already have, ChatGPT Business or Claude Team, and exactly where it stops. Three things that work today. One: put the IC memo and the QoE into a shared project and ask for every claim with its page, owner, date and money line. You'll have the assumption register from slide 2 in an afternoon. Two: add the SPA, the key contracts and the insurance binder and ask for every date-driven obligation with the clause and the page. That's the clock calendar from slide 3; a person still confirms every date. Three: before you trust it, run the six tests from slide 8 on your own deal: plant three obligations, ask three questions it can't answer, open ten citations. These tools are extremely good at reading what one person hands them, once. Now the wall. Memory is per person: a shared project is a folder with project-only memory; it doesn't know who wrote what, when, or who may see it, and the handoff across the close is you re-uploading. Nothing watches for change: the connectors read your own mailbox and files, when you ask; there's no stored assumption for new evidence to be compared against. And there's no not-found discipline and no audit trail. That's the line between a subscription and a memory layer, and it's what you saw in the demo. One caution: plan details change monthly. ChatGPT's agent mode was pulled in August, and Claude's scheduled tasks are still beta on Team, so check the morning you set this up.

If you want help with the first two weeks

One company, one measured pilot, built on the loop you just saw.

0110 to 15 material findings, cited to the page.Each with its money line, its clock and a named owner.

02The clock calendar and the assumption list, built by AI.Tested on findings your team already knows before anyone relies on them.

03A weekly "what changed" check, scored on confirmed actions.Live capabilities in 21 days, in your cloud, with code you own from day one.

Local AI · localai.com · contact@localai.com

Stop asking

Did we finish the 100-day plan?

Ask every week

Does the thesis still hold? What has changed? Who has to act, by when, and how will we know it is done?

AI doesn't make a good acquisition. It keeps testing whether the thesis is still true, and tells the people who can act while there is still time.

1 / 13 Exit

Rotate your phone for presentation view. Swipe to move between slides.

Speaker notes

Research report prepared by Local AI. Current through October 17, 2026. A critical synthesis of 20 deep-research reports, with load-bearing claims re-checked against original sources.

The First 100 Days

Don't lose what diligence taught you

Evidence standard. Every number carries a label. Observed: a documented outcome in a court, regulator or filing record. Data: a measured dataset with a disclosed sample. Research: a study with disclosed methods. Survey: self-reported responses. Practitioner: a recommendation or opinion. Vendor: the source sells what it describes. Illustrative: a constructed example, never presented as a case. Confidence concerns the stated fact, not the effectiveness of the proposed method.

From Local AI. We build the findings register this report describes with deal and operating teams: cited to the page, carrying the real clocks and owners, and tested on findings you already know. Live capabilities in 21 days, in your cloud, with code you own. Book 30 minutes, or read .

Contents

Executive thesis

The buyer needs to transfer responsibility for testing and acting on the investment case, not merely transfer the documents supporting it. A finding has become operational only when someone understands its consequence, accepts authority for the next decision, knows the relevant deadline, and can demonstrate what changed.

In the public cases where a deal later lost value or exposure that diligence had touched, diligence usually found the problem. What failed was the conversion of that finding into an owned, dated, evidenced decision after close. Findings travel as conclusions without their reasons, without the type of money they put at risk, without the deadline that governs them, and without a person who has the authority and budget to act.

Meanwhile the deadlines that matter are set by contracts, insurers and regulators, not by "day 100," and several of them expire before or soon after closing. The working-capital statement is delivered late in about a third of deals. About half of RWI claims arrive after a year. RWI policies generally exclude what the deal team knew, so a finding deliberately deferred past closing usually needs a price, escrow or special indemnity instead. Policy terms vary; counsel decides.

Not every loss is a dropped finding. In Paragon Metals (Del. Supr. Ct., July 2026), the buyer reviewed more than 10,000 documents, saw a struck-through reference to a customer letter about falling purchases, accepted the CEO's explanation, and learned the day after closing that its largest customers were cutting orders. Diligence leaves two lists: what it found, and what it could not establish. Both have to cross the close.

That is where AI earns its place. It can read the deal documents and pull every claim with its page, connect each claim to its money line, clock and owner, compare new authorized evidence against those claims, and flag what changed and what it does not prove. People route, decide and certify. The opportunity is moving from a searchable diligence archive to an investment thesis that is tested every week. AI's characteristic failures, leaving things out, citing sources that do not say what it claims, and smoothing over disagreement, are the same failures a good handoff exists to prevent, so it is trusted only after it passes tests on findings the team already knows. Evidence that AI improves post-close outcomes in PE does not exist in the public record.

The recommended approach is deliberately small: a preserved evidence set, one findings register, and a weekly discussion of changed evidence and required decisions. A custom memory platform is not a prerequisite.

How to read this report

Seven labels. Confidence concerns the stated fact, not the effectiveness of the proposed operating method.

✓Observed: A documented outcome in a court, regulator or filing record.
✓Data: A measured dataset with a disclosed sample. Often a vendor's own book.
✓Research: A study with disclosed methods, including laboratory and clinical analogies.
ƒSurvey: Self-reported responses, including named practitioners' estimates.

ƒPractitioner: A recommendation or opinion from a law firm, adviser or standards body.
!Vendor: The source sells the service or product it describes.
◇Illustrative: A constructed example (DistribCo). Never presented as a case.
LMM transfer: Insured-deal and provider datasets skew larger. Each insight notes whether it transfers to $10M to $250M deals.

Six insights that carry the argument

Each with its evidence, how well it is established, the decision it changes and the operating response.

Six ways diligence knowledge fails

An editorial framework, not a measured distribution. It prevents buying a retrieval solution for a governance problem, and the opposite mistake of blaming management for an exposure the handoff never explained.

Failure	What it looks like	Documented example	What fixes it	Realistic AI role

Case dossiers

Six bounded cases. No AI role is documented in any of them; counterfactuals are labeled as such.

DistribCo: transparent calculations

Every number here is assumed. This is not a case. A hypothetical industrial distributor: $60M revenue, $7.2M QoE EBITDA, 8.0x entry, $57.6M enterprise value. Cash, EBITDA and EV of the same item are never added; EV is a memo line only where the effect is recurring and was priced.

Item	Money type	Arithmetic	Result	Sensitivity	Multiple applies?

A second illustration: reducing days sales outstanding by three days on $30M of credit sales releases about $30,000,000 ÷ 365 × 3 = $246,575 of cash tied up in receivables. That is modeled working-capital release, not revenue or EBITDA, and it must not be added to a pricing example as a single "value created" total.

The finding card

One row per finding that could move cash, EBITDA or a deadline. The acceptance criterion: an IC member unfamiliar with the review can verify a material finding without asking the analyst to reconstruct how the answer was produced.

Three levels of capability, and the tests

Level	What it is	Incremental benefit	Operating burden	Main failure mode

Acceptance tests before relying on any AI-built register. Known answers: on 10 findings the team knows cold, at least 9 reproduced with the right page. Open the citations: 10 random citations opened by a person; any that does not support its claim triggers a full check of that section. Not in the documents: three questions with no answer must return "not found." Planted items: three known obligations buried mid-document must appear. Assertion versus fact: "management assertion" items attributed, not stated as fact. Thresholds are team choices to calibrate, following NIST AI 600-1's guidance to assess generative output against known ground truth.

Day-30 measures, as counts: share of IC-memo red flags with an accepted owner (target 100%); share with a clock identified; rows updated in the last 14 days; open questions resolved; deadlines met; issues first raised in the weekly review that were not in the register; findings closed with artifact evidence; for AI, net minutes saved after correction time plus every caught error. Do not count summaries, alerts or generated tasks as success.

A two-week starting plan

Five misconceptions

Misconception	What the evidence says	Reframe

What to leave out

Claims that appear in the source material but should not reach a slide.

Claim or theme	Why to leave it out

Working with Local AI

This report argues for a small discipline first and AI only where it passes a test. That is how we work, inside the tools and permissions the deal team and the company already have.

Where we would start. One company, from the last-chance review through the first two weeks: a clock calendar traced to the agreement, a findings register seeded from the QoE, legal memo and IC memo with exact quotes and pages, and owners with budget authority. Where an AI layer helps, we test it on findings your team already knows before anyone relies on it.

How it runs. On the cloud you already use, inside existing permissions, with every row tied to the page it came from and a recorded disposition from your team. After close, the register stays with you as institutional memory, not in a vendor's product.

Live capabilities in 21 days

Discovery to live capabilities in 21 days, on the systems you already use. We start inside one real workflow and prove it against your own numbers.

Yours from day one

Production AI agents on the cloud you already use, with code, models, integrations and deployment patterns owned by you from day one. No vendor lock-in.

Built by operators

Our founder built machine learning at scale inside C.H. Robinson. A boutique senior team, not a generalist consultancy, not an SI, not a SaaS vendor.

Bring one real workflow. We will tell you, honestly, whether AI belongs in it yet, and if it does, how fast we can ship.

Book a 30-minute conversation Ask for a Workload-Fit Assessment

Sources

Sources actually used, grouped by evidence type. The underlying reports cite 48 distinct sources; this table carries the ones that support claims in this brief.

Source	Publisher	Date	Evidence type	Primary?