A few months ago I was in a working session with a client’s procurement team, and one of them asked a question I haven’t been able to shake since. We were reviewing a managed services renewal, and she said, almost casually: “If your delivery teams are using AI now, why is the price the same as last year?”
Nobody at the table had a good answer. That question, in one form or another, is about to be asked of every IT services and consulting firm on the planet. And the firms that can’t answer it are going to find out the hard way that agentic AI is not just another technology cycle.
I’ve spent the last couple of years building these systems for Fortune 100 clients. A multi-agent loan processing platform for an auto lender. A RAG-based generative AI platform for a power utility, which is about as regulated an environment as you’ll find. So what follows isn’t analyst speculation. It’s what I’m seeing in actual production work, and what I think it means for the industry I’ve spent more than twenty years in.
This is not the chatbot wave
Let me start by drawing a line, because the word “AI” has been stretched to cover everything from autocomplete to autonomy.
The generative AI most enterprises adopted in 2023 and 2024 was assistive. It drafted emails, summarized documents, helped developers write code a bit faster. Helpful, sure. But a human still pressed every button. The AI advised; people acted.
Agentic AI flips that. An agent is given a goal, not a prompt. It plans the steps, calls the tools and systems it needs, handles intermediate results, and finishes the job. With something like MCP (Model Context Protocol) sitting underneath, agents can talk to enterprise systems and to each other in a standardized way, which is the piece that was missing until fairly recently.
Take the loan processing system I worked on. The old process looked like every loan process you’ve ever seen: an officer collects documents, someone validates them, an analyst checks income and pulls credit, a reviewer applies policy, exceptions bounce around in email for days. In the agentic version, those steps are agents. One reads and validates documents. Another verifies income through tool calls. A third applies underwriting rules. A supervisor agent watches the whole thing and pulls a human in only when something genuinely needs judgment. Days became minutes. And the cost per application? Let’s just say the business case wrote itself.
Now picture that same pattern applied to claims, onboarding, invoice reconciliation, service desk tickets, IT incident response. That’s the next three years of enterprise IT.
What actually changes inside the enterprise
The thing people miss is that a huge share of the enterprise application portfolio exists for one reason: to give human beings screens to click through. When agents execute processes directly against APIs, a lot of those screens stop earning their license fees. Systems of record stay. The thick layer of workflow and case management apps sitting on top of them starts to thin out. I’m already watching CIOs redirect budget from application licenses toward agent platforms, and we’re early.
IT operations gets hit first and hardest, in my view. L1 and L2 support is structured, well documented, and tool-rich, which is exactly the environment agents thrive in. That tier also happens to be the volume base of most managed services contracts. Do the math.
Integration work changes shape too. I came up through the middleware era, connecting systems to systems, and a lot of that muscle memory now needs retraining. The new problems are agent registries, tool governance, context management, orchestration patterns. Different discipline, same architects, if they’re willing to learn.
And then there’s governance, which I’d argue becomes the most important conversation in the room. When an autonomous system is touching customer transactions, you get questions traditional IT governance never had to answer. How do you bound what an agent is allowed to do? How do you audit a decision a reasoning model made at 2 a.m.? How do you even test something non-deterministic? My utility client’s risk team asked all of these before we wrote a line of code, and they were right to. Banks and insurers will be tougher still. The frameworks barely exist yet, which, if you’re a consultant, should sound like opportunity.
The uncomfortable part for services firms
Here’s the structural problem, stated plainly: the IT services industry sells human effort. Time and materials, FTE-based managed services, headcount-driven support. Every one of those models prices labor. Agentic AI deflates the labor content of precisely the work those contracts cover, quarter after quarter, from here on out.
Clients have figured this out. That procurement question I opened with is already turning into a demand: 20, 30, sometimes 40 percent off managed services pricing, on the theory that AI productivity should flow to the buyer. You can resist, and lose the renewal to someone who doesn’t. The deflation is coming either way. The only decision a firm gets to make is whether it leads the repricing or gets dragged through it.
That’s the threat. The opportunity is just as large, and I don’t say that to soften the blow. Every big enterprise is going to re-architect its core processes around agents over the next several years. We’ve seen this movie twice before, with ERP in the nineties and cloud in the 2010s, and both times the transformation wave was the richest revenue pool the industry had ever seen. This one will be too. It just won’t be won with the old playbook.
So what should firms actually do?
I’ll resist the urge to hand you a neat framework, because I don’t think this reduces to one. But after two years in the trenches, a few things seem clear to me.
Stop billing hours for work agents perform. If your revenue model punishes you for becoming more efficient, the model is the problem. The firms moving fastest are pricing per transaction (per loan processed, per claim settled, per ticket resolved) or contracting on outcomes, or running agentic platforms as a subscription. Is the transition painful? Brutally. You cannibalize annuity revenue before the new revenue matures, and your CFO will hate the J-curve. Delay doesn’t make the curve go away, though. It just hands it to a competitor.
Build the real engineering capability, not the slideware version. I’ve sat through plenty of “agentic AI practice” decks lately, and the gap between a strategy deck and a production multi-agent system running inside a regulated bank is enormous. Evaluation harnesses, guardrails, observability, human-in-the-loop design, MCP tool integration. These skills are scarce and will stay scarce for the next two or three years. Your best enterprise architects and integration engineers can learn them. Start retraining them now, not after the first big RFP loss.
Productize. The blank-page, build-everything-bespoke era is ending. Pre-built, industry-specific agent libraries (underwriting agents, claims agents, KYC agents, service ops agents) that work with Claude, Bedrock, and Vertex AI, deployable in weeks instead of quarters. IP becomes the margin story. Headcount stops being one.
Make governance your wedge, especially in regulated industries. Honestly, the fastest differentiation available in the next 24 months is being the firm that can walk agentic AI past a bank’s model risk team or a utility’s compliance function and come out the other side with an approved architecture. Technical capability gets you into the room. Trust closes the deal.
And rethink the pyramid. The traditional delivery model needs a wide base of junior people doing routine work, and that’s exactly the work agents take. What’s left looks more like a diamond: fewer juniors, more senior architects and domain experts, plus what amounts to a fleet of digital workers in the middle. Hiring models, career paths, the whole talent machine has to be redesigned around that shape. This is the change firms will procrastinate on the longest, and it’s the one with the longest lead time.
Why I keep saying two to three years
Past transitions gave the industry a decade to adjust. Cloud certainly did. This one won’t, for three reasons I can see clearly from inside client work: the models are improving on a six-to-twelve-month cadence, the protocol and tooling layer matured faster than anything I’ve watched in thirty years of enterprise technology, and boards are pushing CIOs to show AI economics now, not in 2030.
By 2028 or so, the agentic operating model will simply be how enterprise IT works, and the revenue mix of every major services firm will look different than it does today. Some firms will have chosen that new mix. The rest will have had it chosen for them.
I know which side of that I’d rather be on. The work starts now.
Dr. Harish Kotadia, Ph.D. is an Enterprise AI Architect with 20+ years of Fortune 100 consulting experience, specializing in agentic AI systems built on Anthropic Claude, AWS Bedrock and Google Vertex AI.
© Dr. Harish Kotadia, 2026. All Rights Reserved.