Vol. III · Essays on work after AI
Saturday, October 3, 2026
evolving.work
Strategy · Essay № 051

The Next Big AI Market May Be Boring Services

The biggest AI opportunity of the next decade may not be building frontier models or flashy AI products.

H Héctor Reinoso Aug 8, 2026 12 minute read Share · Save

Every week, someone I know in tech announces they are building an AI product. A startup, a wrapper, a vertical SaaS, an agent platform. The decks all look the same. The market all looks crowded.

Meanwhile, I keep getting hired to do work that has almost nothing to do with building a model.

Workflow mapping. Permission policy. Escalation rules. A six-week training program for an AP team that does not know which Copilot prompt to trust. The work is boring. It pays well. It is steady. The consulting revenue gives us a way to see the demand for it.

The pick-and-shovel business of the AI wave is change management. The receipts are already in.

Everyone wants to build the model. Few want to implement it.

McKinsey’s 2024 survey found that 65% of organizations are regularly using generative AI in at least one business function. Only 13% have seen EBIT increases of 5% or more [1]. Adoption is wide. Value capture is thin. Those two numbers are not a contradiction — they are a market.

Look at where the dollars are moving:

These are the biggest professional services firms in the world. Their AI revenue is not coming from selling models. It is coming from implementation, governance, and operating-model redesign — built on top of a U.S. management consulting sector that did $392 billion in 2024 [9].

History rhymes here. The ERP implementation services market was larger than the ERP software market for decades [11]. SAP and Oracle made the licenses. SAP partners, regional integrators, and vertical specialists made the businesses that got the licenses live. The same pattern is starting to repeat — except the underlying technology changes faster, the use cases sprawl wider, and the gap between “we own it” and “we use it” is bigger.

The press wants a story about model labs and product launches. The P&Ls are telling a story about humans helping other humans rework what they do.

What “implementation” means

Looking inside one of these projects makes “implementation services” concrete. In these projects, it covers six unglamorous, indispensable steps, and most companies skip four of them.

  1. Workflow Discovery. Map what the organization does — not what the org chart says it does. The documented process and the real process are almost never the same. Surface the difference.

  2. Data Access Design. Define precisely what the AI can read, write, modify, or transmit, and what it must never touch. Least privilege as the default. This is a business decision, not an IT setting.

  3. Escalation and Guardrails. Specify when the system must stop and route to a human: transactions above a threshold, decisions involving regulated customer classes, outputs that need legal review. Test the rules against real data before launch.

  4. Training and Adoption. Build calibrated trust — teach people when to believe the system, when to override it, and how to flag a problem. Adoption is a behavior-change program, not a launch email.

  5. Governance and Ongoing Oversight. AI systems drift. Without monthly escalation reviews, quarterly output audits, and a named owner per system, accurate deployments become inaccurate ones silently [6].

  6. Measurement. Define the business metrics the system must move before it goes live. Without a baseline, “success” defaults to “we shipped it.”

This is the work that gets skipped when a team is excited and a license is sitting unused. The Hackett Group sells a four-phase version — strategy, data engineering, integration, post-deployment optimization — and McKinsey, Accenture, Deloitte, and Avanade all sell variations. The market is converging on a shared definition because the failure modes keep repeating.

A small example. A mid-market manufacturer had bought enterprise RPA + AI tooling to automate accounts payable, procurement approval, and inventory reporting. After eight months, nothing was live. The internal IT team understood the software. They did not have the muscle for workflow discovery, data access design, or exception rules. The license was about to be abandoned.

Automation Anywhere’s professional services team came in for six weeks. They interviewed the process owners, documented the real flows instead of the manual ones, designed read-only ERP access with a tightly scoped staging-table write path, wrote escalation rules for invoices above a dollar threshold, and ran a four-week change program for the AP and procurement teams.

Three AP workflows and two procurement approvals went live within 90 days. Processing time dropped 60% for covered workflows; error rates fell 80%. The client renewed and expanded the license. The engagement cost roughly $180,000; the client calculated annual savings of $340,000 in labor and rework.

The software worked the whole time. The boring work was missing.

Why most companies can’t do this themselves

Implementation is not a transitional pain point that disappears once tooling matures. It is structural.

The share of companies abandoning the majority of their AI initiatives before production has risen from 17% to 42% year over year [3]. A separate estimate puts it more starkly: 88% of companies use AI in at least one function, but only 26% have scaled beyond pilots [5]. That 62-point gap between “we are trying” and “it is working” is the market — the entire economic surplus of a technology left on the floor because the layer between purchase and use is missing.

The gap has known causes. A 2025 survey found that up to 87% of AI projects never reach production because of data issues, and 61% of firms have halted AI projects entirely due to skill shortages. Microsoft’s Work Trend Index reports that employees bring AI into work faster than organizations can govern it [4]. The governance layer is being built behind the usage curve, not ahead of it.

Internal teams can configure software. They cannot, by training or by mandate, redesign workflows, manage organizational politics, or build trust with frontline users. That is a different skill set living in a different part of the org chart, and most companies have no senior person whose job is “make the AI tools actually used safely.” Stanford HAI’s 2025 AI Index notes that deployment so far concentrates in a narrow set of tractable functions — customer support, marketing, software engineering — and the next wave of value requires translating AI into harder, messier operational workflows [10]. Harder workflows mean more external help, not less.

Gartner has already classified generative AI consulting and implementation services as a distinct market category whose purpose is to “bridge gaps in skills, experience and technology” [7]. That classification recognizes demand for help between buying the technology and using it.

Most companies are not held back by lack of model access. They are held back by lack of operational permission to use what they have already bought.

The objections worth taking seriously

There are three serious objections to this thesis. I have heard each of them in the room.

“AI will automate the implementation work itself”. If models can map workflows, generate governance docs, and write integration code, the services market collapses. It is the cleanest version of the objection — and it conflates technical configuration with organizational change. Configuration will get cheaper, faster, and probably nearly free. Organizational change will not. Workflow discovery requires sitting in a room with the AP supervisor who has never told anyone that the “approved” approval chain has been bypassed for two years. Data access design requires the CFO and the head of compliance to agree on what the system may touch. Training requires showing a skeptical team that the tool is right enough to trust and wrong enough to verify. None of that is a prompt away. SAP existed for thirty years and the partner ecosystem implementing it kept growing.

“Mid-sized businesses can’t afford consulting-style implementation”. Mostly true — which means the service has to fit the size of the business. A ten-person clinic does not need a six-figure roadmap. The opportunity has segments: large enterprises buy from the Big Four, the mid-market ($10M-$500M) buys from boutique vertical specialists, and the smallest firms buy productized service packages or vertical SaaS with embedded governance.

“The Big Four will capture everything”. They are capturing the enterprise tier. Their minimum project sizes, billing rates, and overhead mean they cannot economically serve companies below roughly $500M in revenue — the entire mid-market and most of the upper SMB. The boutiques and regional integrators that thrived in the ERP era did so for exactly this reason. The structural gap is real.

There is a subtler version of the optimistic case worth surfacing. BCG often frames AI value creation as 10% model, 20% data and technology, 70% business-process and operating-model transformation. That sounds like a self-serving consulting pitch until you take it literally: the firm with the biggest AI consulting practice in the world is telling clients that 70% of the work is the work most other people are not selling. With 79% of leaders calling AI adoption critical to competitiveness [12], the governance and process layer to support it is not optional.

What a durable AI services business looks like

Avanade — the Accenture-Microsoft joint venture — was retained by a $1.2B regional bank that had bought Microsoft 365 Copilot licenses for 2,000 employees at $30 per seat per month. After 90 days, fewer than 200 employees were using it meaningfully, and the CIO was about to face the budget committee. The software was not the problem. People did not know which tasks to use it for, did not trust the outputs, and had received no guidance on what data the tool could access for sensitive client information.

Avanade ran a three-phase engagement: use-case discovery per business unit; data classification and Copilot scope design (which SharePoint libraries it could see, which it could not, an escalation protocol for compliance-sensitive outputs); and a six-week role-specific training program with internal champions and a monthly effectiveness review.

Active use rose from 10% to 67% in six months. The bank documented 2-4 hours per week per active user in time savings. The engagement cost $280,000; the bank calculated $4.2 million in annualized productivity value. They expanded Avanade into a managed-services retainer covering ongoing governance and new feature rollouts.

The retainer is what turns implementation into an ongoing business. The implementation project creates the relationship. The governance retainer creates the economics.

The durable services firm in this market is built around four design choices:

The size of the prize is meaningful. The global AI consulting services market is projected to grow from $14 billion in 2024 to roughly $91 billion by 2035, at 26.2% annually [8] — faster than most technology product categories. And the broader services market AI can touch — engineering consulting, professional services, advisory work — is around $10 trillion, against a software market of roughly $650 billion. The picks-and-shovels market is, by definition, larger than the mine.

So where does this leave you?

The AI conversation will keep being about models, products, and platforms. That is what makes good copy. The money, for a while yet, will keep going somewhere else.

If you are an operator, run your existing deployment against the six-step list above before buying another license. The ROI you are missing is almost never in the model.

For founders building a services firm, the durable version is built around recurring governance, not deployment milestones. Pick a vertical. Resist the urge to define yourself by a vendor. The next Accenture may start as a small AI workflow integration shop.

Investors, the durable play is not a horizontal chatbot company — it is a vertical operating-capability firm where AI improves a specific process, and where switching costs, compliance requirements, and recurring governance create the moat.

As for the consultants already inside this practice — your next decade of work is not selling AI. It is helping organizations safely act on what they have already bought.

The miners get the headlines. The pick-and-shovel businesses get the decade.

References

AI strategyAI transformation
About the author
Héctor Reinoso writes about how organizations should be designed when intelligence becomes abundant. He works on enterprise AI adoption inside a large industrial organization, after 25 years as a strategy consultant and transformation executive across four continents. Harvard MBA, industrial engineer, and still writes the code he writes about. More →
Continue reading
Strategy · Oct 3, 2026

Twelve Agents, €200, and a Consultant With a Name

When agents make research cheaper, consulting must earn its fee through the structure of the analysis, the quality of judgment, and clear accountability.
Org Design · Oct 2, 2026

When Coding Gets Cheaper, the IT Handoff Gets Harder to Justify

If your supplier uses coding agents while you still specify and verify the work, it is time to reconsider what you are outsourcing.
Strategy · Sep 29, 2026

Ten Agents, 250x the Tokens

Forecasts of 24x token-consumption growth by 2030 are not aggressive, they are too low, because they model usage as if it grows linearly per user.
Org Design · Sep 26, 2026

The Myth of the 50-Agent Founder

The fantasy of a solo founder running a company with 50 agents misunderstands what agents do to cognitive load.
Strategy · Aug 11, 2026

Compute Is Not Free Lunch

Don't ask what AI costs — ask what your people can do with it.
Strategy · Aug 5, 2026

The Best AI Opportunities May Not Look Like AI Companies

Riding the AI wave does not require building an AI software company.