The AI Consultancy Trap: Why Implementation Beats Strategy Decks

The difference between an AI consultancy and a hands-on implementation firm is simple.

One sells a long transformation. The other ships work, gets first numbers on the board, and iterates from real metrics.

That is the better path: implement, measure, adjust, move.

OpenAI could sell AI advice at any price it wanted. In May, it built the opposite. It stood up a Deployment Company that puts forward-deployed engineers inside client operations to ship working systems, and it bought a consultancy, Tomoro, along with its 150 engineers to staff it. The company with the most to gain from selling decks decided to build instead.

That is the tell for the entire consulting trade. The firms charging you to "transform with AI" are the ones AI is hollowing out from the inside. This piece is about not becoming their next line item.

For context, I'm Mark, founder of Limestone Digital (@LimestoneHQ).

We've been on the market ten years and we run over 170 client engagements. Here's the one number that frames everything that follows: every engagement we run has fewer people on it today than the day it started, and the firm is growing faster than it ever has. Hold onto that for a second.

The trap has two doors, and they catch different people.

Part 1: Door One, and Why Anyone Walks Through It

Door one sends in a partner whose time is billed at a rate that only makes sense if you never see the invoice broken down.

By published benchmarks, a McKinsey senior partner bills around $1,190 an hour and a first-year associate around $400, and a full engagement runs from roughly $1.7 million to $8 million. Bain bills a small team between $110,000 and $160,000 a week.

You're not paying for an answer. You're paying to keep an expensive person fed by the day, for as long as the work runs.

Most of the time that buys you a maturity curve, a junior's demo, and three months later the same broken workflow you started with. Nobody validated the outputs. Nobody built the thing that runs next Tuesday without the consultant in the room. Talk to the people one, two, three levels below the executive who signed the deal, and they can't stand it, because they watched the circus and still have to do the actual work when it leaves.

So why does anyone buy it? Because a consultant is a safety net for a manager who is scared to decide. Bring in a big name and the decision becomes "informed," and if it fails, the story writes itself: we hired the best, so it wasn't my call. Nobody ever got fired for hiring McKinsey. What they're buying is insurance against blame, and the company foots the bill.

I want to be fair, because there's a version of this worth every dollar. If you have one concrete, excruciating problem, a system that's down, a technical knot nobody in the building can untie, and a specialist walks in and fixes it, pay them. Pay them $150,000 for a week if they stop the thing that's been bleeding you for a year. That's not the trap.

The trap is generic knowledge work dressed up as insight. The deck about "unlocking your potential," the workshops, the quarter that ends with an upsell and never turns into anything you can point to. In 2026, generic advice is the cheapest thing in the world. You can ask a frontier model for the top strategy frameworks, have them in a minute, hand them to your team, and start learning from real implementation instead of paying to be told what you could have read. If you need a consultant billing four figures an hour to set your strategy, the honest move is to change the person setting your strategy.

Look at what the money buys, this year. A 2026 Teradata survey of 1,000 senior technology and data leaders found just 7% of enterprises had operationalized agentic AI, with 68% still stuck experimenting or developing. A consultant's incentive does nothing to move that number, because the meter runs whether or not the workflow ever ships.

Part 2: Door Two, and the Five Questions That Open It

Every firm you've heard of has an AI practice now. Accenture, IBM, Deloitte, McKinsey's QuantumBlack, BCG X, PwC, EY, KPMG. When a category grows this fast, three kinds of vendor show up at your door, and only one of them can deliver.

The first is the legacy firm in new shoes: big, established, not AI-native, running a strategy practice it rebranded last quarter. One question cuts through the pitch: how many of your own people has AI made redundant this year, and why are you hiring more? A firm selling efficiency it hasn't found inside its own delivery is selling a slide, not a system.

The second is the team that assembled to ride the wave, a group pulled together to catch the demand, with a case-study page built from the past lives of people they hired last spring. The logos are real, but the team's track record isn't. Ask when the company was founded and when those case studies happened, and watch the dates refuse to make sense.

The third is the medium firm that's actually adapting. Some are doing the genuine work of becoming AI-native, and those are worth your time. The rest adapt fast or they're gone inside a year, and you don't want to be the client mid-engagement when that happens.

The tell across all three is one thing: a real implementer's delivery team gets smaller as the technology gets better. A vendor whose headcount on your account only grows is selling you the old model with a new label.

Ours runs the other way, and I mean it literally. On a finance-operations build we've run since the spring, we started with two engineers and a fractional architect. Today it's one engineer a few days a week, because the agents carry the reconciliation volume and the exceptions are the only thing left for a person to touch. We invoice less now than we did at kickoff. When your CFO asks why the line item went down last quarter, that's the product working.

The buyers who place us see the same contrast in their own numbers. A project shop quoted one private-equity operating partner $120,000 to build a workflow for one of his portfolio companies, over six months. We shipped it in a third of the time, at less than half the price. Two months, not six.

Everything else he tells us is downstream of that. The other firms he talks to are project-based: six meetings, a scoping phase, then a fixed-price proposal for a nine-month build. By the time they finish scoping, our engineer has been shipping for two months. He buys us for "speed, certainty, and flexibility", not a deck.

When one of his companies keeps insisting it needs another engineer, the answer is a pod that costs about what that one hire would and ships far more. A consultant's model can't produce that, because a consultant's model needs the headcount to go up.

So here's the script. Run these five on any AI vendor, in a single call, before you sign anything. Each has a clean answer and a tell.

1. Does a successful engagement end with more of your people on my account, or fewer? If the honest answer is more, their P&L grows when yours doesn't.
2. When does billing start, and what am I paying for in month one? If the meter runs on discovery and slide-building, you're funding their learning curve at your rate. Ask them to eat discovery. The ones who believe in the work say yes without flinching.
3. Show me one workflow you operate every week without you in the room. Something that runs on a schedule and writes to a real system, not a demo you drive by hand. Deloitte's Tech Trends 2026 found only 11% of companies actually running AI agents in production. Vendors who don't operate their own work are how you stay out of that 11%.
4. Who owns this the day you leave, and what's the single number we're moving? No named owner and no baseline metric means you're buying a science project with a nice interface. The number has to be one you already track, so the before and after stay honest.
5. How much smaller has your own delivery team gotten as the models improved? A vendor selling AI efficiency should be finding it inside their own shop first. If they've only added bodies, either they don't believe their pitch or the pitch doesn't work.

Note where they dodge. The dodges tell you more than the answers.

What This Looks Like in Practice

A PE firm reached out to us last month to get them out of a consulting engagement. Six figures spent on AI consulting across seven portfolio companies, and all they had to show for it was a team chat subscription.

They needed real EBITDA impact by December to underwrite their 2027 exits. Instead they had a 150-hour discovery deck and a dev vendor billing extra "AI productivity" fees with zero boost in shipping speed.

Before you laugh, recognize you almost certainly have a version of this running right now.

The same four gaps show up every time, between what legacy consulting promises and what the engineering actually needs.

1. Verifiable proof, not slide decks. Consultancies spend weeks on interviews to sell a maturity assessment. We spend week one building a golden set: 20 to 60 real operational cases with pre-agreed correct outputs. Grant Thornton found only 9% of private-equity firms are confident they could pass an independent audit of their AI governance within 90 days. A golden set is what closes that gap. An assessment just promises to.
2. Production infrastructure, not abstract roadmaps. Instead of a phased journey, we stand up a live foundation inside the portco's own Azure tenant within three weeks. We lock down read-only data access, strip PII and PHI, and cap per-seat cost. The client owns everything from day one. If we leave, nothing breaks.
3. Lean automation pods, not billable headcount. Traditional firms add $350-an-hour bodies to grow the account. We deploy a $17K-a-month, month-to-month pod: one senior engineer, automated development harnesses, and a fractional architect. As the harness carries more of the load, our human footprint shrinks over time instead of growing.
4. Internal ownership, not vendor lock-in. Unscrupulous vendors keep the setup obscure to protect the billing. We embed the portco's product owner from day one, deliver a 90-day self-sufficiency plan, and leave telemetry running. Success means the portco ships its own updates without calling us.

Part 3: The Industry Already Knows

If this sounds like an outsider's cheap shot, it isn't. The firms selling AI transformation have told their own people, behind closed doors, exactly where this goes.

At a June town hall, Deloitte walked its own staff through a chart showing hourly-billed work shrinking to a thin sliver of the market by 2035, with AI agents taking the rest. One consultant's read afterward: "our model is toast, we're basically getting replaced by robots."

In May, KPMG cut about 400 US advisory roles, roughly 4% of that business, and trimmed close to 10% of its US audit partners. The stated reason was softening demand, which is almost worse, because the work is thinning before the automation even lands.

McKinsey's headcount is down from about 45,000 to around 40,000 over the last 18 months. The firm that will happily sell you an AI operating model is running that same cost math on itself first.

Meanwhile, the most-fought-over hire in the business is the person who ships. The forward-deployed engineer who sits inside your operation and builds it is the role OpenAI, Anthropic, and Google are all racing to hire. Everyone agrees now that the value is in deployment. The recruiting war is just the scoreboard.

By Fortune Business Insights' estimate, the AI consulting market reaches about $12 billion this year, on its way to $74 billion by 2034. A number growing that fast is a tell of its own. Most of the firms in it rebranded a strategy practice last quarter. The pitch got a new noun before the delivery got a new muscle.

What the Honest Version Looks Like

The model is changing under everyone, including the firms selling it. What replaces it comes down to two things: the order the work runs in, and commercial terms that match the technology. Demand both from any partner you let near your operation.

Start with the order. If you've been quoted a six-figure retainer to "assess your AI readiness," this is what that retainer skips. Five steps.

1. Frame. Pick one workflow with real users, a deadline the customer feels, and a number you already track. Sign the baseline before anyone builds. This step produces nothing to show a board, and it's the only thing that makes step five honest.
2. Prove. Build the smallest useful slice on real data in two to four weeks, with the assumptions, errors, and human review all visible. If the failure profile is ugly, you found out for the price of a month.
3. Harden. Add the security, the cost controls, the fallbacks, and a named owner. A working demo is where most programs stop. Production is a runbook and a person whose phone rings when it breaks.
4. Embed. Train the users, hand over the knowledge, watch adoption. Technical accuracy isn't business value. If nobody in the operation touches it on a Tuesday, it's a science project.
5. Scale. Adjacent workflows and reusable pieces, chosen from what step four measured, never from a platform pitch.

The consultancy version runs in the other order: a long diagnostic before anyone touches real data, a program justified by projected benefits, success measured in workshops held and slides delivered.

Then the commercial terms. This is what you should demand from any partner, and how we run it at Limestone.

1. Zero-cost discovery. We sign an NDA and learn how you actually work on our dime, not yours: how the deal team sources, how the controllers close the books, where your best people burn their days on version control instead of the job you hired them for.
2. Clear scope, no phantom billing. You see the SOW, the cost, and the exact business change before a single dollar moves.
3. Results-based billing. The meter starts only when the work reaches a baseline number you already track. Discovery and slides cost you nothing.
4. No lock-in. Month to month, a named owner on your side, telemetry left running, so you can walk whenever you want and nothing breaks.

A partner who builds for your independence prices it differently from one who builds for your dependence.

The way we staff it is an AI Velocity Pod at $17K/month. We're operational in three to five weeks, your people go back to their real jobs while we carry the technical risk, and if we don't deliver, you don't pay. We turn away about 30% of the companies that come to us, because the model only works when the fit is there.

When it is, it's the cleanest way I know to put AI inside a business without betting the quarter on a consultant's learning curve.

If you want to see this against one of your own workflows, book a discovery call, and come ready to talk about your least favorite one and what makes it so painful.

TLDR

• The firms charging you to transform with AI are the ones AI is hollowing out. OpenAI, the company best positioned to sell advice, built a Deployment Company and bought 150 engineers instead. That's the tell.
• The trap has two doors. Door one is the high-ticket strategy engagement, where a partner bills around $1,190 an hour and a full engagement runs $1.7M to $8M, and you're paying to keep an expensive person fed, not to get an answer. Nobody got fired for hiring McKinsey, which is exactly the problem.
• Door two is the vendor who changed shoes. Three types: the legacy firm in new shoes, the team assembled to ride the wave with borrowed case studies, and the medium firm actually adapting. The tell across all three is whether their team on your account gets smaller as the models improve, or bigger.
• The buyers see it in their own numbers. A project shop quoted one PE operating partner $120K over six months for a build; we shipped it in a third of the time at less than half the price. Two months, not six. He buys us for speed, certainty, and flexibility, not a deck.
• Run five questions on any vendor: more of your people or fewer, when billing starts, one workflow they run without them in the room, who owns it when they leave, and how much smaller their own team has gotten. Only 7% of enterprises have operationalized agentic AI and only 11% run agents in production, so the dodges are common.
• The industry already knows. Deloitte told staff "our model is toast," KPMG cut 400 advisory roles, McKinsey shrank from 45,000 to 40,000, and everyone is racing to hire the forward-deployed engineer who actually ships.
• The honest model is a standard you can demand from anyone: run the work in order (Frame, Prove, Harden, Embed, Scale), with terms that match it (free discovery, clear scope, results-based billing, no lock-in). Ours shrinks over time instead of growing.

P. S. Book a discovery call to skip the consulting trap: https://meetings-na2.hubspot.com/mark-ajzenstadt

— Mark Ajzenstadt, Founder @LimestoneHQ