Low switching costs and high AI risk in US stocks

By Harry Read (A5 Labs) and Christian Richter | Published 9 October 2026

Introduction

This is a worked example run across two MCP servers in the same Claude session. Baseline does the qualitative pass: it scores companies 1 to 5 on questions an analyst would ask, with a written rationale behind each score. Shibui Finance does the quantitative pass on prices, financial statements and valuation for US stocks. The output is one shortlist.

The screen looks for businesses whose customers can leave easily and whose product is exposed to capable, cheap AI. On Baseline that is low Barriers to Switching and a low AGI Scenario score. Both fields are scored so that 5 is the most favourable, so a low AGI Scenario score means high exposure. We use it as our standard example because it separates cleanly and the names it returns are ones most readers will have an opinion on.

Part one below is the Baseline side. Part two is written by Shibui.

Part one: the Baseline screen

The filter as run on the Baseline MCP on 6 October 2026, restricted to US listings because that is the universe Shibui covers:

Barriers to Switching <= 2
AGI Scenario <= 2
Country = US

This returns 26 companies out of 1,374 US names scored on both fields, about 1.9%. Across the full Baseline universe the same two conditions return 37, so the US accounts for most of the set. All but one score 2 on AGI Scenario, which in Baseline terms means a business that is not existentially substitutable but where most of the current profit pool is exposed to fee compression. Six score 1 on Barriers to Switching, Evercore and Marcus & Millichap among them.

The set leans heavily to Communication Services (12 of 26), mostly online marketplaces, review sites and ad-tech. Staffing and labour intermediaries make up a second cluster (Robert Half, ManpowerGroup, RXO), with advisory boutiques a third (Evercore, Perella Weinberg). The rationales for Robert Half and Evercore make much the same point, that the product is cognitive labour sold by the hour or by the mandate, and clients can take the work in-house once the tools are good enough.

The more useful split is on Baseline View, the top-level score for the direction of the business. Seven of the 26 still score 4, which means Baseline's read is that the business is improving despite both flags. Eleven score 3, six score 2 and two score 1. None score 5. That spread is what makes the screen worth handing to a quantitative tool, since the qualitative flags alone do not say which of these the market has already marked down.

The 26 names

Sorted by Baseline View, then Business Quality. Sub-scores and Baseline View are integers; Business Quality and Competitive Dynamics are group scores shown to one decimal place. Scores as pulled from the Baseline MCP on 6 October 2026.

The bottom of the list includes some very small companies, which may drop out once Shibui applies a size floor. Sector labels are Baseline's own and do not always match GICS: LegalZoom sits in Industrials, for instance.

Three of the names in more detail

Evercore (EVR). Barriers to Switching scores 1. Advisory mandates are won deal by deal, there are no systems or data to migrate, and clients rotate between Evercore, Goldman Sachs, Lazard and Centerview as a matter of course, sometimes following a departing managing director. AGI Scenario scores 2 because the analyst and associate layer (valuation, modelling, drafting, diligence) is the part of the delivery model most easily reproduced. The senior advisory layer is better protected by reputation, independence on special committees and the need for a named human to be accountable for an irreversible transaction. However Pricing Power still scores 4, because the fee is small next to the cost of bad advice on a large deal, and Baseline View is 4.

Duolingo (DUOL). Barriers to Switching scores 2. The rationale separates behavioural lock-in, which is strong (streaks, habit, social features), from structural lock-in, which is close to nil, since a user can leave with no downtime or data loss and can use a general AI tutor alongside the app at no cost. AGI Scenario scores 2 for a similar reason. Personalised instruction, practice generation and conversational tutoring are things a general model does directly, and subscriptions were around 84% of 2025 revenue. The Duolingo English Test is the more defensible piece because it is an accepted credential, but it is not large enough to offset the subscription exposure. Baseline View is 4.

Robert Half (RHI). Barriers to Switching scores 2. Clients can end engagements on short notice without penalty, they multi-home agencies through preferred supplier lists and vendor management systems, and each placement is a natural point to re-award the work. AGI Scenario scores 2 because both segments, Talent Solutions (64% of 2025 revenue) and Protiviti, sell white-collar cognitive labour. The rationale expects a smaller, more senior business rather than an obsolete one. Pricing Power is 3, largely a markup on wages passed through to clients, and Baseline View is 3.

Part two: the Shibui pass

Duolingo and Upstart have lost 54% and 55% of their share price over the past year, while their revenue grew 29% and 46%. Measured against their own history, both now trade close to the cheapest they have been relative to sales in five years. Seven of the 21 names are in that position, and all seven score 3 or 4 on Baseline View; the three names scoring 2 with a usable multiple trade higher in their range.

We took the 26 names through Shibui Finance to see which of them the market has already marked down, with prices as at the 6 October 2026 close, the same day as Baseline's scores, and financial statements through the quarter ended 30 June 2026. Three prompts in Claude with Shibui connected produce the numbers below:

Five names drop out before any valuation work, which leaves 21. Balincan International (ALTB) trades over the counter, outside Shibui's NYSE and NASDAQ coverage, and Fiverr and Onfolio fall below a $500 million market-cap floor. Paramount Skydance and Coursera have changed shape. Paramount Skydance became Skydance Corporation (SKYD) on 6 October alongside its takeover of Warner Bros. Discovery (8-K), and Coursera merged with Udemy in May (10-Q), so their five-year multiples compare different businesses.

Trump Media & Technology stays in, but with trailing revenue of about $4.5 million its price-to-sales ratio means little, so the table marks it n/m.

The 21 names

Sorted by Baseline View, then by where today's price-to-sales ratio sits in its own five-year range, from 0 at the five-year low to 100 at the high. Returns are share price only, without dividends, from the 6 October 2025 and 6 October 2023 closes. Revenue growth compares the last four reported quarters with the four before, and operating margin covers the last four. Data as pulled from Shibui Finance for the 6 October 2026 close.

MNTN listed in May 2025 and RXO in October 2022, so their ranges start at listing, and MNTN's growth compares the June quarter with a year earlier. CarGurus restated 2024 and 2025 for a discontinued business, so its row uses the restated filings and its percentile is approximate. People Inc's range spans its 2025 spin-off of Angi, and most of RXO's revenue growth comes from a full year of Coyote Logistics, which it bought in September 2024.

Where the market and Baseline View line up

As a group, the 21 names fell 13% over the past year, equal-weighted, against a 16% gain for SPY, the S&P 500 ETF, and only six of them beat it. Over three years the average is +36%, but Innodata accounts for almost all of that; without it the group is up 2%, against 81% for SPY.

The seven names in the bottom fifth of their range split into two groups. Duolingo and Upstart lost 54% and 55% over the year while revenue grew 29% and 46% and operating margins rose. So far the risk shows up in the share price but not in the reported numbers. Yelp, TripAdvisor, LegalZoom and Playtika had revenue growth of 12% or less, and Robert Half's revenue fell 5% as its margin dropped to near zero. For those five, the price and the results point the same way.

Evercore, the first of part one's worked cases, sits near the middle of its range at the 43rd percentile, after an 18% fall over the year and with net revenue up 45%.

At the other end, PC Connection trades at its highest multiple in five years and RXO at the 84th percentile. Innodata scores 4 on Baseline View and sits at the 80th. Opendoor scores 2 and still sits at the 70th percentile after a 75% fall in its shares, because its revenue has shrunk from $15.6 billion in 2022 to $3.3 billion over the last four quarters.

The joint shortlist

The shortlist takes two names from each of the two groups above and Opendoor from the other end. None of this is a view on whether a share is cheap. It compares each price with the company's own history and its latest results.

Duolingo (DUOL). At 6.0 times sales, the shares trade at less than half their five-year median of 13.9, on revenue of $1.1 billion over the last four quarters.

Upstart (UPST). The shares trade at 1.8 times sales against a five-year median of 4.4, and operating income over the last four quarters was positive after a loss in the four before.

Yelp (YELP). At 0.68 times sales, the shares sit near their five-year low of 0.66 and well below the median of 1.80, while the operating margin has held near 12%.

Robert Half (RHI). The shares are down 55% over three years, and the June 2026 quarter brought an operating loss of $62 million.

Opendoor (OPEN). Revenue over the last four quarters, $3.3 billion, is about a fifth of its 2022 level, which keeps the multiple above its five-year median of 0.28 even after the fall in the shares.

Sources

Prices, valuations and financial statements from Shibui Finance, with revenue and most operating income figures checked against each company's 10-Q and 10-K filings on SEC EDGAR; Evercore's revenue is net revenue. Corporate actions from the Skydance Corporation 8-K on the SKYD listing and the Coursera 10-Q for the quarter ended 30 June 2026.

Frequently asked questions

Can Claude use two MCP servers in the same conversation?

Yes. With both servers connected, Claude calls each one as the question needs it. In this example Baseline supplied the qualitative scores and Shibui Finance the prices, valuations and financial statements, all in the same Claude session.

Which US stocks have low switching costs and high AI risk?

On Baseline's scores as at 6 October 2026, 26 US companies score 2 or lower on both Barriers to Switching and AGI Scenario, among them Evercore, Duolingo, CarGurus, Robert Half, Yelp and Opendoor. Both fields are scored so that 5 is the most favourable, so low scores mean customers can leave easily and the product is exposed to capable, cheap AI. The full list with scores is in part one.

Which of these stocks has the market already marked down?

Of the 21 names left after Shibui's checks, seven trade in the bottom fifth of their own five-year price-to-sales range: Yelp, TripAdvisor, LegalZoom, Playtika, Upstart, Duolingo and Robert Half. Duolingo and Upstart grew revenue 29% and 46% while their shares fell. For the other five, revenue grew 12% or less, or fell.

What do Baseline's scores measure?

Baseline scores companies from 1 to 5 on questions an analyst would ask, covering business quality, competitive dynamics, pricing power, management and risks, each with a written rationale. Baseline View is the top-level score for the direction of the business. The scores describe the business and are not a view on the shares.

What does Shibui Finance add to the screen?

Shibui Finance is a free MCP server with daily prices, financial statements, valuations, technical indicators and SEC filings for 10,000+ NYSE and NASDAQ securities, queried in plain English. Here it supplied each company's market cap, 1- and 3-year price returns, price-to-sales ratio against its own five-year range, revenue growth and operating margin.

How can I run this screen myself?

Connect both servers to Claude. Run the Baseline filter from part one (Barriers to Switching 2 or lower, AGI Scenario 2 or lower, US listings), then paste the three prompts from part two. Shibui is free and Baseline has a free tier. Scores and prices change over time, so a new run will not match these tables exactly.

Is the shortlist a recommendation?

No. The shortlist compares each price with the company's own history and its latest results, and Baseline's scores describe the business, not the shares. None of it is a view on whether a share is cheap.

Run it yourself

Connect both servers in Claude, then start from the Baseline filter in part one and the three prompts in part two.