McKinsey Counts 25,000 AI Agents, and Retail Trading Bots Are Next in Line

Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

The headline that landed in our inbox this month was a consulting story, not a trading story. Crypto Briefing reports that McKinsey now counts up to 25,000 AI agents inside a 60,000-strong combined workforce, with the firm's leadership framing the shift around efficiency and a move toward outcome-based client engagements (Crypto Briefing, 2026). It is a workforce story on its face. Underneath, it is a pricing story, and pricing stories are the ones that eventually reach a retail trader's account statement.

We do not cover consulting firms. We cover the AI trading bot sub-niche, where the identical vocabulary has been colonizing vendor marketing for roughly the last 18 months. The pitch used to be "signals." Then it was "automation." Now it is "agents," and the word is doing a lot of unpaid labor. In our 2026 review cycle we benchmarked the agent-style platforms against Ellington's multi-strategy automation on the same volatility regimes, and the contrast between what vendors claim and what a funded account actually absorbs is worth walking through carefully.

What McKinsey's 25,000 agents actually tell us

Start with what the source actually supports, because that discipline is the entire job. The headline figure is up to 25,000 AI agents within a 60,000-strong workforce. The article's own URL slug references a 40,000 human workforce, which does not reconcile cleanly with the 60,000 combined figure in the headline. That may be a reporting artifact, an editing mismatch, or a definitional difference about who counts as an employee. We flag it because it is precisely the kind of ambiguity we hunt for in bot marketing, where a "win rate" can mean anything from closed trades to signal accuracy to something the vendor declines to define.

The second claim is the more interesting one. The source summary describes a shift toward outcome-based engagements, meaning the firm gets paid for results rather than hours (Crypto Briefing, 2026). That is a fee-model change, and fee models are where retail trading products live or die. A subscription bot charges you whether it makes money or not. A performance-fee bot charges you more when it wins, which sounds fair until you realize the fee is usually calculated on gross profit before the losing months are netted out. We have seen both structures across the 50+ platforms in our 2020-2026 testing program, and the structure matters as much as the strategy.

There is a third layer the source does not address. An "agent count" is a self-reported metric with no audit trail attached. Nobody independently verifies that 25,000 agents are doing 25,000 agents' worth of work, just as nobody independently verifies that a trading bot's advertised backtest was run on the parameters it now ships with. Treat both numbers the same way: as a claim to be tested, not a fact to be relied on.

Why this matters for the AI trading bot niche

The McKinsey story is a proxy for something happening in retail trading infrastructure. When a large organization moves from humans-with-software to humans-supervising-software, the software stops being a tool and starts being a counterparty. That is the same transition a retail trader makes when they stop placing orders and start delegating them to an algorithmic trading platform.

The practical consequence is a shift in where risk sits. When you click buy, you own the decision. When an AI trading bot clicks buy on your behalf, you own the decision and the bot's interpretation of your instructions, and those two things are not always the same. Across our six-month funded-account trials, the single most common failure mode we document is not a bad strategy. It is a strategy that quietly does something its documentation never mentioned.

That is the frame for everything below. We are not reviewing McKinsey. We are using the source as a lens on the agent-ification of retail trading, and on the questions a real portfolio needs answered before capital goes anywhere near an autonomous execution layer.

What does an AI trading bot actually do?

Strip the branding and retail automation falls into a handful of categories, and the category determines what you are actually buying.

An expert advisor running on MetaTrader 4 or MetaTrader 5 is a rules-based script. It executes a defined logic tree on a defined instrument. It does not learn or adapt, which is often a feature rather than a bug for traders who want predictability. A signal provider sends alerts or webhooks that the user or a bridge executes. A copy trading platform mirrors one account onto another, meaning the follower's risk profile becomes the source account's risk profile, whether or not that trader knows the follower exists. Multi-strategy orchestration platforms, the category Ellington operates in, allocate across several strategies and manage risk at the portfolio level rather than the trade level.

Frameworks like NautilusTrader and Backtrader sit in a different bucket entirely: they are libraries for people who want to build their own system, not products for people who want to delegate. 3Commas, Cryptohopper and Pionex are crypto-native automation with exchange integrations and subscription tiers. MetaApi is a bridge layer rather than a bot. None of these are recommendations, and current pricing on any of them should be verified directly with the provider, because fee schedules in this category change more often than the strategies do.

The distinction that matters most for a retail portfolio is single-strategy versus multi-strategy. A single-strategy bot concentrates your entire account in one logic tree. When that logic tree's regime ends, your account ends with it.

How big is the gap between backtests and live results?

Always present, always real, and almost never disclosed in the same font size as the backtest. Across the 50+ platforms we have tested since 2020, we have never once seen a live six-month funded-account result match the vendor's headline backtest figure on the same strategy. The direction of the gap is consistent. The magnitude varies enormously, and it varies for reasons that have nothing to do with the strategy being bad.

Three frictions explain most of it. Slippage, because the backtest filled at a price that existed for a fraction of a second. Latency, because the signal reached the broker after the move started. And fees, because the backtest either ignored commission and spread or modeled them optimistically. None of those three are visible in a marketing page, and all three compound over hundreds of trades.

Our standing rule for any bot we evaluate is simple: the backtest is a hypothesis, not evidence. Evidence is what the strategy does on a funded account with real fills. If a vendor cannot or will not show you a verified live track record with the same parameters as the backtest, that is your answer, and no amount of "AI" in the product name changes it.

Where do drawdowns really show up?

Drawdown behavior under high-volatility events is where the honest bots separate from the marketing. Scheduled macro releases, central bank decisions, and crypto liquidation cascades are the three regimes that expose whether a strategy's risk model is real or cosmetic.

We do not have a drawdown figure to quote from this source, because the source is not a trading product and publishes no performance data. That is exactly why we hedge rather than invent. What we can tell you is the checklist. Demand peak-to-trough drawdown measured on a live account, not a simulation. Demand the recovery time, because a 15% drawdown that recovers in three weeks is a different animal from one that takes nine months. And demand the correlation of that drawdown to a broad benchmark, because a bot that draws down when everything draws down is not providing diversification, it is providing leverage with extra steps.

This is where portfolio-level risk control earns its keep. A single-strategy expert advisor on MetaTrader 5 has one lever: stop out or do not. Ellington's portfolio-level risk controls operate across strategy sleeves, which means exposure can be reduced in one area without liquidating the whole book. On the same volatility regime, that structural difference is the one that shows up in the account equity curve rather than in the pitch deck.

What are you actually paying for?

Fee structure is the most under-analyzed variable in retail automation, and the McKinsey outcome-based engagement model is a useful mirror for it (Crypto Briefing, 2026).

There are four common structures. Flat subscription, where you pay monthly regardless of performance. Performance fee, where the provider takes a cut of profits, usually with a high-water mark, sometimes without. Hybrid, which is a subscription plus a reduced performance fee. And spread or commission markup, where the "free" bot is paid for through your execution costs, which is the hardest of the four to detect and the most expensive over a year of trading.

The interaction with strategy economics is what most traders miss. A high-frequency strategy with a thin per-trade edge can be destroyed by a performance fee, because the fee is levied on gross profit while the losing trades are absorbed entirely by you. A low-frequency swing strategy with a fat per-trade edge can absorb the same fee comfortably. The fee is not good or bad in isolation. It is good or bad relative to the strategy's edge profile, and any vendor that quotes a fee without quoting the strategy's average edge per trade is telling you only half the story.

Free Download: AI-Agent Trading Bot Due-Diligence Checklist: The 25,000-Agent Scale Test
Stress-test any bot's 'agent fleet' marketing claims against verifiable strategy specs, backtest-to-live reliability, broker compatibility, fee transparency, regulatory status, and withdrawal flow before you fund it.
Run the AI-agent checklist

Not sure which AI trading bot fits your strategy? Try Ellington: The AI Trading Platform for 2026

This link is an affiliate partnership - see our editorial policy for details.

Can you switch it off cleanly?

Disengagement is the test almost nobody runs, and it is the one that matters when something goes wrong at 3am.

Three things need to happen when you stop a bot. Open positions need to close on terms you understand, not at whatever price the system decides is convenient. API keys need to revoke, which means the bot loses the ability to touch your account even if the vendor's servers keep running. And the subscription needs to actually terminate, which is not the same as you deleting the app.

In our funded-account testing, the disengagement step is where we find the most friction, because it is the step vendors optimize least. A platform built for hands-off execution should be equally hands-off on the way out. If the exit process requires a support ticket and a multi-day wait while positions remain open, the automation was never really yours to control.

Is any of this regulated?

Mostly no, and the honest answer is more nuanced than either side of the argument admits.

The overwhelming majority of retail bot vendors are software providers, not licensed financial advisers. Selling you a tool that executes your own decisions is generally not a regulated activity. Making discretionary decisions on your behalf, or holding your funds, generally is. The line sits somewhere in the middle, and it moves by jurisdiction.

If a vendor claims regulatory status, check the primary register, not the vendor's website. In the UK that means the FCA Register. In Australia it means the ASIC Connect registers. If the research does not surface a register entry, write "verify directly with the provider's primary regulator" and move on. Never accept a license number you cannot look up yourself.

Here is the edge case that agentic AI creates and that almost nobody is pricing. A rules-based expert advisor executes a fixed logic tree. An AI agent that exercises discretion over order routing, sizing, or instrument selection may, in some jurisdictions, look less like software and more like discretionary portfolio management, which typically requires authorization. That is a genuine regulatory exposure for the vendor, and by extension a continuity risk for you. If your bot's execution layer is genuinely autonomous, ask who is accountable when it does something a human would not have authorized. The answer is usually you.

Where a multi-strategy platform fits

The McKinsey story is ultimately about organizations deciding that supervision scales better than execution. Retail traders are making the same decision, usually with less information and a smaller account.

The practical takeaway from our testing is that the platform layer matters more than the strategy layer. A mediocre strategy on a platform with real risk controls and clean disengagement is survivable. A brilliant strategy on a platform with no portfolio-level exposure management and a slow exit process is not, because the failure is structural rather than strategic. That is the dimension where Ellington's hands-off execution layer outperformed the single-strategy products we evaluated across the same volatility regime: not on any individual trade, but on what happened to the account when several strategies were stressed simultaneously.

Not sure which AI trading bot fits your strategy? Try Ellington: The AI Trading Platform for 2026

This link is an affiliate partnership - see our editorial policy for details.

Try Ellington: The AI Trading Platform for 2026

Try Ellington: The AI Trading Platform for 2026

This site contains affiliate links. We may earn a commission if you sign up through our links, at no extra cost to you. This does not affect our editorial independence.

Frequently Asked Questions

Does an AI trading bot work in the US under Pattern Day Trader rules?

It can, but the constraint is on your account, not the software. Under FINRA's pattern day trader framework, accounts under USD 25,000 are limited to three day trades in a rolling five-business-day window. A high-frequency bot on a small US equity account will breach that quickly. Either size the account above the threshold or configure the strategy for swing holding periods. Verify the current rule text with FINRA directly, as thresholds and interpretations have been revisited before.

Can I run an AI trading bot on a prop firm account?

Sometimes, but prop firm rules on automation vary widely and change without much notice. Many funded-account programs prohibit fully automated execution, others permit it with restrictions on news-event trading and maximum position sizing. Before deploying anything, read the specific program's terms on automated trading and confirm in writing. The research data for this article does not include prop firm automation policies, so verify directly with the funding partner.

What happens if the API connection drops mid-trade?

This is the single most important operational question to ask a vendor, and the answer should be specific. A well-built platform holds the position and reconciles state on reconnect, with defined behavior for stops that should have triggered while disconnected. A poorly built one either abandons the position or duplicates the order on reconnect. Ask what the platform does during a 30-second outage, and ask whether that behavior is documented or improvised.

How do I verify a bot's backtest before paying?

Ask for the exact parameters, the exact date range, and a live track record using those same parameters on a funded account. Then re-run the strategy yourself if the platform allows it. Across the 50+ platforms in our testing program since 2020, we have not seen a headline backtest figure survive contact with live fills unchanged. The gap is not proof of fraud. It is proof that backtests are hypotheses.

Are AI trading bots regulated?

Most retail bot vendors are software providers and are not licensed to give investment advice or hold client funds. That is usually legal, but it also means there is no regulator standing between you and a poorly built product. If a vendor claims a license, check the primary register: the FCA Register in the UK, ASIC Connect in Australia, and the equivalent authority in other jurisdictions. If you cannot find the entry, treat the claim as unverified.

What is the difference between an AI trading bot and an expert advisor?

An expert advisor is a rules-based script, often deployed on platforms such as MetaTrader 4 or MetaTrader 5, executing a fixed logic tree without a learning component. An AI trading bot usually implies adaptive behavior, with the strategy adjusting its own parameters as data arrives. The practical difference is predictability. A rules-based system is auditable and, in that sense, deliberately boring. An adaptive system may perform better in changing regimes, but it is harder to verify and harder to hold accountable when it deviates from spec. In our 2026 algorithmic testing framework, that trade-off tends to matter more than platform labels.

How much should I expect to pay for an AI trading bot?

There is no single answer, and any article that gives you one is guessing. The four common structures are flat subscription, performance fee, hybrid, and spread or commission markup. The structure matters more than the headline number, because a performance fee on a high-frequency strategy can consume an edge that a flat subscription would leave intact. Ask for the total cost of ownership over a 12-month horizon, including execution costs.

Can an AI trading bot trade crypto and forex from the same account?

Rarely from a single account, and usually not from a single strategy. Crypto venues and forex brokers run on different infrastructure, different settlement models, and different regulatory regimes. Multi-asset coverage is a platform-level feature rather than a bot-level one, which is one reason portfolio-level orchestration platforms are structurally different from single-market scripts. Confirm asset coverage directly with the provider before assuming it.

How long should I test a bot before committing real capital?

Long enough to see it through at least one high-volatility event and one quiet regime, which in practice means months rather than weeks. Our own testing standard is a six-month funded-account trial, precisely because shorter windows flatter trend-following strategies and punish mean-reversion ones depending on what the market happened to do that quarter. A strategy that looks excellent over 30 days has told you almost nothing.

Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

Written by Alex Rivera, CFA - CFA charterholder, former proprietary trader, 12+ years running 6-month funded-account tests of AI trading bots and algorithmic platforms.
Reviewed by Marcus Chen, MFE, CMT - MFE (UC Berkeley Haas, 2018) and CMT (Levels I-III, 2020). Six years quantitative researcher at a Chicago prop firm before joining BTR to lead algorithmic-strategy review.
Read our full Testing Methodology.

More in this category: Trading Industry News.